Which AI agents for construction actually act on their own in 2026?

By their own published descriptions, almost none. Procore's AI plans page tells buyers that field and office teams review and approve every output. Autodesk states that the professional still gives instructions and controls the final product. Most of the rest of the market is computer vision (Buildots, Doxel, DroneDeploy, Togal.AI, EarthCam), constraint optimisation (ALICE), predictive modelling (nPlan) or classic workflow automation with document extraction (Kojo, Agave) — capabilities that predate the agent framing and were renamed into it. The honest 2026 answer is that construction AI drafts, detects and forecasts; a named human still adopts the output.

Is Document Crunch still an independent company?

No. Trimble Inc. completed the acquisition of Document Crunch, Inc. on 4 April 2026 for $246.4 million, disclosed in Trimble's Form 10-Q for the quarter ended 3 July 2026. Trimble announced the agreement on 2 April 2026 and closed two days later. The product still ships, and the site currently serves an interim co-branded Document Crunch and Trimble logo. Any 2026 listicle that describes Document Crunch as an independent contract-review startup is working from a stale roster.

Has Procore acquired DroneDeploy?

Not yet. Procore entered a merger agreement to acquire DroneDeploy for approximately $845.0 million in cash on 27 July 2026, announced publicly on 29 July 2026. Procore's Form 10-Q for the quarter ended 30 June 2026 states the completion is anticipated later this year and is subject to closing conditions, and lists the risk that the merger is not completed. DroneDeploy's own homepage carries a banner saying it has signed a definitive agreement to be acquired. Signed is not closed. A contractor signing a multi-year DroneDeploy agreement in late 2026 should ask in writing what happens to pricing, roadmap and integration neutrality at close.

What happened to Autodesk Construction Cloud?

It was renamed. Autodesk's own construction site states that Autodesk Construction Cloud is now part of Autodesk Forma and that ACC product names have been updated. The rename happened in the first half of 2026; trade sources give conflicting exact dates and Autodesk's announcement page could not be retrieved directly, so no specific day should be asserted. Corroboration that the new brand has propagated: competitor integration lists now name Autodesk Forma rather than ACC. Existing licences, project data and seat counts reportedly roll over unchanged.

Was the 2023 Davis-Bacon rule struck down?

No. Three provisions of it were vacated. In AGC of America v. DOL, No. 5:23-cv-00272 (N.D. Tex.), final judgment on 24 June 2026 vacated the truck-driver item at 29 CFR 5.2(2)(iv)(D), the material-supplier carve-back in 29 CFR 5.2(2), and 29 CFR 5.5(e) — the incorporation-by-operation-of-law provision — in its entirety. The vacatur is nationwide. The Fifth Circuit appeal was voluntarily dismissed on 26 June 2026, so nothing is pending. The 30 percent rule in 29 CFR 1.2 was never challenged, and the site-of-the-work expansion survives. Two traps: the eCFR still prints the vacated text with no editorial note, and DOL's own page still describes the 2024 preliminary injunction rather than the vacatur.

Can an AI agent seal a drawing or take responsible charge of engineered work?

No, and this needs no AI-specific rule to establish. NCEES Model Law section 110.20(E) defines responsible charge as exercising full professional knowledge of and control over work, requiring the licensee to be personally aware of the project's scope, to be capable of answering questions relevant to the engineering decisions made, and to accept full responsibility for the work. Model Rules section 240.20(D) permits the seal only on work under the licensee's responsible charge and gives the board final authority over that determination. NCARB's Model Law and Regulations (July 2024) states that by sealing a technical submission the architect represents that the architect was in responsible control. Every branch of the closed list of what may be sealed terminates in a licensed human. There is no category for work prepared by a tool.

Has any licensing board actually ruled on AI in engineering?

One has, and it declined to write a new rule. Texas Board Policy Advisory Opinion No. 71, adopted 14 November 2024, states that neither the Texas Engineering Practice Act nor the Board rules directly ban the use of AI software, that it is acceptable for licensees to use it, and that licensees are ultimately responsible for any work product they sign and seal. It concludes that no new Policy Advisory Opinion will be developed because the Act and Board rules adequately address the use of artificial intelligence software at this time. That is a single opinion, not a trend. NCEES has issued no AI position statement, and its August 2025 Model Law and Model Rules contain zero occurrences of artificial intelligence, machine learning or software.

Is beyond-visual-line-of-sight drone flying legal for jobsite capture in 2026?

Not by rule. 14 CFR Parts 108 and 109 are marked [Reserved] in the current CFR — there is no Part 108. The FAA published its BVLOS NPRM on 7 August 2025 at 90 FR 38212 and no final rule has been issued as of 23 August 2026. The statutory deadline is section 930 of the FAA Reauthorization Act of 2024, Pub. L. 118-63, codified at 49 U.S.C. 44811; the FAA missed the NPRM deadline by roughly eleven months, but the final-rule date of around 7 December 2026 is not yet due. BVLOS today runs on individualised authorisations: a Part 107 waiver under 107.205(c), or a 49 U.S.C. 44807 exemption, whose sunset was extended to 30 September 2033. Separately, Remote ID under 14 CFR Part 89 is fully enforceable, and flying over the trades below means dealing with the operations-over-people rules in 107.39.

Is it true that 98 percent of megaprojects run over budget?

That figure should not travel, and it is not Bent Flyvbjerg's. It comes from McKinsey's June 2015 report The construction productivity imperative, which says its own research estimates that 98 percent of megaprojects suffer cost overruns of more than 30 percent — a stronger claim than the version in circulation, with no sample, dataset or published method. Flyvbjerg's peer-reviewed work is what to cite instead: across 258 transportation infrastructure projects worth $90 billion, Flyvbjerg, Holm and Buhl (2002) found costs underestimated in almost nine out of ten projects, with actual costs averaging 28 percent above estimates.

What does an AI agent build cost for a construction firm, and how long does it take?

Frenchy Digital's published bands: discovery and workflow audit $9k–$22k over 2–4 weeks; a single-workflow agent $28k–$70k over 4–9 weeks; a multi-workflow platform with system integration $70k–$180k over 9–16 weeks; an enterprise, multi-site or regulated build $180k–$420k+ over 14–24 weeks. Senior-led work is $150–$225/hr, retainers run $2,500–$9,500/month, there is a 30-day post-launch warranty, and full source-code and IP ownership transfers to the client. A fixed-price phased proposal follows within 5 business days of a discovery call.

Top 10 AI Agents for Construction Firms in 2026

The claim under test

Every ranking of construction AI agents you can find asserts that these products are agents and that they are accurate. Neither claim survives contact with the vendors' own documentation. By their own published descriptions, the large majority of the products marketed as construction AI agents in 2026 are computer vision, constraint optimisation, predictive modelling or classic workflow automation, with agent language applied afterwards. And not one of them has been evaluated by anyone other than the company selling it.

That is not a rhetorical position. It is what the pages say. Procore's AI plans page carries the headline "Let agents do the work, you make the call," and then states plainly that your field and office teams review and approve every output. Autodesk's construction AI page states that AI will serve as a digital assistant but the professional will still be the one giving instructions to the system, engineering prompts, making selections from various design alternatives that the system generates, and controlling the final product. Two of the largest vendors in the category have already conceded the autonomy question in their own marketing copy. Everything downstream of that concession — including how much of your risk an agent can absorb, which is none — follows from it.

The second claim fails on a different axis. No independent benchmarkexists for construction AI agents. We checked, vendor by vendor, on 2026-08-23: no academic benchmark, no standards-body test suite, no government evaluation, and no third-party comparative test of output quality for any product in this roster. What exists instead is a set of numbers the vendors publish about themselves — "reduce delays by up to 50%," "11% ahead of schedule," "1.5x bid volume per estimator," "a degree of accuracy unattainable by human operators" — none with a disclosed methodology, sample, control group or denominator. One vendor publishes an ROI calculator, which is a form the buyer fills in to be told what the vendor already assumed.

The one time a US regulator actually audited a vendor's published AI performance metrics, the numbers did not hold. In In the Matter of Presto Automation Inc., Securities Act Release No. 11352 (14 January 2025), the SEC addressed a company that had told investors its drive-thru voice AI delivered over 94% accuracy even in noisy environments and 95%–99% automated order completion. The order found the product "lacked the capability to take orders on their own and required substantial human involvement," with human order takers located abroad processing the vast majority of orders. Presto consented without admitting or denying the findings; the remedy was a cease-and-desist order with no civil penalty. The findings are against Presto only — the order's "Supplier A" is a different company against whom the SEC made no findings whatsoever, and Presto Automation Inc. is not Presto Phoenix Inc. That is a different industry and a different product. It is also the reason this ranking scores only what you can check yourself.

There is a third claim, quieter and more damaging, buried in most of these articles: that the roster is current. It is usually not. Between January and August 2026 this market consolidated hard. Four of the biggest names on any 2025 list are no longer independent companies, or are no longer called what they were called. Ranking an acquired product under its dead independent name is the error that gets an article dismissed by exactly the person it was written for — and it is the error that a general contractor, who has been on the wrong end of a stale vendor list before, will spot in about fifteen seconds.

This piece is one of ten vertical rankings built on the same methodology; the parent 2026 guide to AI agents across industries sets out the shared standard and links the other nine. What follows is the construction version, and everything in it was checked on 2026-08-23.

How we ranked, and what we refused to rank on

We scored eight attributes, all of which a contractor's own team can re-check without talking to a salesperson. We scored nothing that only the vendor can see. Every cell in the table below is either a fetched fact with a public source, or the literal words "not publicly disclosed" or "not verified." We never estimated a price.

Methodology — what was scored, what was excluded, and how to re-check it

Scored:documented public integrations (a developer docs or API reference URL, not a logo on a marketing page); whether the vendor publicly offers a DPA; SOC 2 Type II, ISO 27001 or equivalent status readable on a public trust page; pricing transparency (published figure versus "contact sales"); locked to a suite versus standalone; ownership and funding from public record — SEC filings, Companies House, dated press releases; whether any independent evaluation exists; and published data residency or retention commitments.

Excluded, deliberately:accuracy, detection rate, deflection, containment, resolution rate, ROI, time saved, hours returned, delay reduction, bid-volume lift. Every published figure of that kind in construction AI is vendor marketing about the vendor's own product. Also excluded: software-directory star ratings, which measure user sentiment rather than output correctness, and several of which surfaced during this research as AI-generated affiliate pages of unknown provenance.

Date checked:2026-08-23. Every corporate status was confirmed against a regulated filing or the vendor's own live site on that date.

How to re-check it in about ninety minutes: open the vendor's pricing page and note whether a number appears; open /security and /trustand note the HTTP status; search EDGAR full-text for the vendor name to catch acquisitions disclosed in an acquirer's 10-Q; look for a developer documentation subdomain; and ask, in writing, for the SOC 2 Type II report and the DPA. What comes back — and how fast — is itself the score.

Two of the eight attributes deserve a note, because they are where the market is least honest with itself. The first is integrations. "Integrates with Procore" appears on almost every vendor site in this category. Three of the sixteen products we examined publish developer documentation that a buyer's IT team could actually read before signing: Procore, Autodesk Platform Services, and Agave. Everyone else asserts integrations on a marketing page. That is a real difference, and it predicts how a mid-implementation integration conversation will go.

The second is certification. A trust page that exists is not the same as a trust page you can read. Two vendors in this roster run client-rendered trust centres that returned nothing but a page title to an automated fetch — the certification list is behind a request-access flow. That is not evidence of absence, and we say so in the limitations. But it is a verifiable fact about the buying experience: you cannot see the attestation without a sales conversation, and that is precisely the attribute being scored.

The finding that inverts the usual narrative: the incumbent platforms are more transparent about security than most of the AI-native challengers. Procore, Trimble and OpenSpace publish complete, itemised trust centres including a DPA. Two well-funded pure-play AI startups in this market return HTTP 404 on both /security and /trust and publish no certification claim we could locate anywhere on their sites. If your procurement process starts with a security questionnaire, that ordering matters more than any feature grid.

The comparison table

Here is the ranking. It is ordered by verifiable transparency and corporate stability, not by capability, because capability in this market cannot be compared without a benchmark that does not exist. Read the "what it actually is" column first — it is the column that will change how you brief your team.

#	Product	Corporate status (23 Aug 2026)	Published price?	Public trust page / certifications	What it actually is
1	OpenSpace	Independent; acquired Disperse (announced 28 Oct 2025)	No — pricing link resolves to a quote request	SOC 2 Type 2, SOC 3, CSA STAR Level 1, Cyber Essentials Plus; DPA published as a downloadable document; subprocessors listed	360-degree capture plus computer vision, with an agent and API layer added recently
2	Procore AI	Independent, NYSE: PCOR; acquired Datagrid (closed 16 Jan 2026)	Tier structure published; every tier says contact sales	ISO/IEC 27001:2022, SOC 1 Type 2, SOC 2 Type 2, FedRAMP Moderate, SOX, EU-US DPF; DPA listed; dedicated AI governance section	Drafting copilot inside Procore — the vendor states teams review and approve every output
3	Autodesk Forma	Autodesk, NASDAQ: ADSK; renamed from Autodesk Construction Cloud in H1 2026	Not publicly disclosed for the AI features	Not verified in this pass	Classic ML and computer-vision features plus a conversational assistant
4	Document Crunch	Acquired by Trimble Inc., closed 4 Apr 2026, $246.4M	Not publicly disclosed	Own trust centre unreadable; parent Trimble publishes ISO 27001:2022, ISO 27701, SOC 1/2 Type 1 and 2, SOC 3, CMMC — product scope unconfirmed	Contract and specification review copilot with an agentic layer
5	Togal.AI	Independent, Miami FL	Yes — Growth at $299/mo per user, billed yearly; Business is custom	SOC 2 Type II, announced on the company news page; no trust page (/security returns 404)	Computer vision on drawings for automated takeoff — not an agent
6	Agave	Independent; marketing domain moved to useagave.com; $15M Series A announced July 2026	Not publicly disclosed	SOC 2 indicated by an AICPA badge; report type not stated; no public DPA located	Integration plumbing and document automation; ships an MCP server other agents call
7	Kojo	Independent; Wesco International is a strategic investor and an integrated supplier	Not publicly disclosed	Public security page states SOC 2, report type not stated, report on request; ISO 27001 not claimed	Materials procurement workflow automation plus invoice extraction
8	nPlan	NPLAN LIMITED, company no. 11043916, Active on UK Companies House	No — there is no pricing page (/pricing returns 404)	ISO 27001, SOC 2 and Cyber Essentials stated on the vendor site; certificates not independently retrieved	Predictive ML forecasting on schedule data — a forecasting engine, not an agent
9	Trunk Tools	Independent, venture-backed; $40M Series B reported July 2025	Not publicly disclosed	Vanta-hosted trust centre exists; certification list is client-rendered and not publicly readable	Retrieval and drafting copilot over the project document set
10	DroneDeploy	Signed to be acquired by Procore for approx. $845.0M — NOT CLOSED as of 23 Aug 2026	Not publicly disclosed	None found (/security returns 404)	Aerial and ground capture plus computer vision with an alerting layer

Four things fall out of that table that no feature comparison will tell you. Two of sixteen products publish a price. Three publish a complete, itemised trust centre with a DPA, and they are the two big platforms plus OpenSpace. Four of the best-known names are not independent companies any more, or are about to stop being one. And by the vendors' own descriptions, the majority of the category is computer vision or workflow automation marketed as agents.

Six more products were examined and did not make the ten. Each exclusion is for a reason drawn from the same scoring sheet, not from a judgement about quality:

Product	What it is	Why it is not in the ten
Buildots	Site-capture imagery compared against the BIM model and schedule to compute completion state; acquired Genda, sold as Buildots Field	No public trust or security page — /security and /trust both return 404 and no certification claim is publicly readable. Publishes a privacy policy referencing EU and UK GDPR
Doxel	Reality capture compared against schedule and budget for automatic progress tracking; active and independent	No public security page (/security returns 404); latest disclosed funding round is 2021, which is a fair, checkable observation about runway
ALICE Technologies	Construction optioneering — generates large numbers of alternative build sequences under resource and crew constraints	Compliance not verified, and the pricing page is a multi-step qualification questionnaire that asks company type, project size and geography before showing any figure
Handoff	Generates estimates from a voice description or plans, plus CRM, e-signature, invoicing and payments; the most transparent pricing in the roster	Scoped to residential remodelling contractors and handymen, with AI takeoff limited to residential plans up to 5,000 sq ft. Wrong tool for a commercial GC or a sub on a $50M job
Bild AI	Extracts door schedules and hardware sets from plans and specs, diffs drawing sets, flags plan/schedule/spec inconsistencies	Pivoted to CSI Division 08 only — doors, frames and hardware. bild.ai now redirects to a Division 8 product. Correct for that niche, not a general drawing reader
EarthCam	Jobsite cameras plus Control Center 9 software and vision analytics the vendor says identify over 110 types of materials, vehicles and safety issues	Camera hardware with computer-vision analytics, explicitly not an agent; current ownership structure is not publicly disclosed

Five names that should not appear in any 2026 ranking. Datagrid is dead as a standalone name — Procore completed the acquisition of Toric Labs, Inc. (d/b/a Datagrid) on 16 January 2026 for $168.0 million in cash. Disperse was absorbed into OpenSpace and Genda into Buildots. SafeAI cannot be confirmed to exist in 2026: safeai.ai and www.safeai.ai both 301-redirect to pronto.ai, which sells off-road autonomous haulage for mining and quarries and makes no mention of SafeAI anywhere. Whether that was an acquisition, an asset sale or a wind-down could not be determined. Beam AI is not a construction vendor at all — it is a horizontal enterprise agent platform with no construction, takeoff or estimating product, and it advertises 98% accuracy with no methodology, which is a textbook example of the figure this ranking refuses. Do not confuse it with Bluebeam.

One more table, because it is the single most useful procurement filter in this market. These are the products with real, public developer documentation as of 2026-08-23 — the kind your IT lead can read before you sign, rather than after:

Vendor	Public developer documentation?	Note
Procore	Yes	Docs, API Reference and Changelog at developers.procore.com
Autodesk	Yes	Autodesk Platform Services at aps.autodesk.com
Agave	Yes	Full developer centre with API reference, auth flow and source-system guides — note the docs still sit on the old agaveapi.com domain while marketing moved to useagave.com
OpenSpace	Partial	Integration partners page; no public developer or API reference located
Trunk Tools, Buildots, Doxel, Togal.AI, Bild AI, EarthCam, ALICE, nPlan, Kojo	No	No public API reference located; integrations asserted on marketing pages only

The ten: platforms and document agents (1-5)

Each entry below states what the product does, what is verifiable about it, what is not disclosed, who it fits, who it does not, and who owns it. Where a vendor publishes a performance number, we name it and refuse it rather than quietly omitting it — refusing in public is the whole point.

1. OpenSpace — Open Space Labs, Inc.

What it does:360-degree cameras, drones and smartphones turn a site walk into a navigable visual record pinned to the drawings, with progress tracking, field notes, issue tracking and BIM coordination layered on top. OpenSpace Field reached general availability on 3 February 2026 — a smartphone, image-based field-execution tool for punch list, observations and issues, with bi-directional Procore and Autodesk Construction Cloud integration. The platform navigation names an "AI Agents and APIs" capability under an "Act" pillar.

What is verifiable: the only fully public, itemised trust centre in this roster — SOC 2 Type 2 and SOC 3, CSA STAR Level 1, Cyber Essentials Plus, GDPR/CCPA/CPRA alignment, EU-US and Swiss-US Data Privacy Framework participation, a downloadable DPA, a named subprocessor list (AWS, Salesforce, Zendesk, Atlassian, Auth0 by Okta), and a published risk profile of 24-hour RTO and RPO. Regional site selectors are published for the US, Australia, Canada, the EU, Japan, Saudi Arabia, Singapore and the UK. Corporate: independent, and an acquirer rather than a target — it acquired Disperse, announced 28 October 2025, terms undisclosed.

What is not disclosed:price. The pricing link resolves to a "Request a Quote" form. Specific data-residency guarantees live in the DPA rather than the overview page. The vendor's "trusted on 110K+ projects" is a vendor-published figure and is not evidence of anything about output quality.

Fits / does not fit: fits a GC that wants a defensible visual record for delay, defect and payment disputes, and whose procurement process starts with a security questionnaire. Does not fit a firm expecting the tool to act — the documented, mature capability is documentation and verification, not autonomous action.

2. Procore AI / Procore Agents — Procore Technologies, Inc. (NYSE: PCOR)

What it does: a library of construction-specific agents that run inside Procore against the project record — Contract Review, Daily Log, Deep Search, RFI, Submittal Review, Bid Analyzer, Bid Leveling, Change Analysis, Drawing Analysis, Drawings-and-Specs Specialist, Fast Search, Field Observations, Financial Analyst.

What is verifiable, and it is the most important fact in this article: Procore describes its own agents as drafting assistants, not autonomous actors. From procore.com/ai/agents: the agents "instantly cross-reference project files, help flag risks, and generate submission-ready drafts directly inside your daily workflows—leaving your team to review, approve, and build."

Let agents do the work, you make the call. AI agents draft your RFIs, review submittals, and fill out daily logs. Your field and office teams review and approve every output, so you stay in full control.
— Procore, AI plans page (procore.com/ai/plans), retrieved 2026-08-23

Also verifiable: a public trust centre listing ISO/IEC 27001:2022, SOC 1 Type 2, SOC 2 Type 2, SOX, FedRAMP Moderate, VPAT and the EU-US, UK Extension and Swiss-US Data Privacy Frameworks, with a DPA listed under Legal, policies stated to be based on the NIST Cybersecurity and Risk Management Frameworks, and — unusually — a dedicated AI section covering AI Overview, AI Security and AI Monitoring. Procore is one of very few vendors here documenting AI governance in its trust centre at all. Pricing tiers are published (Digital Coworker Starter Pack, Pro and Enterprise) with agent counts, action limits and term lengths, and the underlying model providers are disclosed as Anthropic and Google on Starter and Anthropic, Google and OpenAI on Pro and Enterprise. That is real transparency and worth crediting.

What is not disclosed:every tier says "contact sales." No dollar figure is published, though the page notes Procore does not charge by the seat. And note the asymmetry: strong security attestations, zero output-quality evaluation. SOC 2 tells you the data is handled properly. It tells you nothing about whether the submittal review is right.

Ownership, and a checkable illustration of why filings beat press releases: Procore's own release of 20 January 2026 announcing the Datagrid acquisition said financial terms had not been disclosed. Its Form 10-Q for the quarter ended 31 March 2026discloses the number: "On January 16, 2026, the Company completed the acquisition of all outstanding equity of Toric Labs, Inc. (d/b/a Datagrid)... The purchase consideration was $168.0 million in cash." Allocation: developed technology $50.5M on a seven-year life, customer relationships $4.0M on ten years, goodwill $114.926M.

Refused: "multiplying team capacity and protecting your margins" is vendor marketing, not a measurement. Fits / does not fit: fits a firm already standardised on Procore, because the agents operate on data inside Procore — a hard suite lock. Does not fit a firm running its project record elsewhere.

3. Autodesk Forma (formerly Autodesk Construction Cloud) — Autodesk, Inc. (NASDAQ: ADSK)

The roster trap first:Autodesk Construction Cloud is now Autodesk Forma. Autodesk's own construction site states that ACC is now part of Autodesk Forma and that ACC product names have been updated. The rename landed in the first half of 2026; trade sources disagree on the exact dates and Autodesk's announcement page returned HTTP 403 to automated retrieval, so no specific day is asserted here. Reported component renames: Autodesk Docs to Forma Data Management, BIM Collaborate Pro to Forma Design Collaboration, and Autodesk Build, Takeoff and Estimate to Forma Build, Takeoff and Estimate, with existing licences, project data, settings and seat counts rolling over. That the rename has propagated is corroborated by two independent competitors now naming "Autodesk Forma" and "Autodesk (Forma) Build" in their own integration lists.

What it does: named AI features are Autodesk Assistant (conversational), Construction IQ (risk analysis), AutoSpecs (submittal automation), Automated Symbol Detection, Automated Drawing Extraction, Bid Forwarding, Financial Data Extraction and Autotags on Photos.

What it actually is: mostly classic ML and computer vision plus a copilot. Construction IQ, AutoSpecs and symbol detection are pattern recognition features that predate the agent framing. Autodesk is explicit about who is in control:

AI will serve as a digital assistant, but the professional will still be the one giving instructions to the system, engineering prompts, making selections from various design alternatives that the system generates, and controlling the final product.
— Autodesk, AI for construction (construction.autodesk.com), retrieved 2026-08-23

What is not disclosed:pricing for the AI features specifically, and we did not verify a trust page in this pass — that cell says "not verified," not "none." Autodesk is reported to be publishing MCP servers so third-party agents can call its data, which would be a meaningful interoperability move; that report comes from a trade summary rather than Autodesk documentation, so treat it as unconfirmed. Fits / does not fit: fits design-led and BIM-centric organisations already inside the Autodesk stack. Does not fit anyone hoping to buy the AI features independently of the suite.

4. Document Crunch — now owned by Trimble Inc. (NASDAQ: TRMB)

Corporate status first, because most rankings still get it wrong: Document Crunch is not an independent company. Trimble's Form 10-Q for the quarter ended 3 July 2026states: "On April 4, 2026, we completed the acquisition of Document Crunch, Inc. ('Document Crunch') for consideration of $246.4 million for 100% of the equity interests in Document Crunch. We financed the acquisition by borrowing from our credit facilities." Net identifiable assets were $39.4M including $32.4M of intangibles, meaning roughly $207M of the price was goodwill. Trimble announced the agreement on 2 April 2026 and closed two days later, on the first day of its fiscal Q2. The site currently serves an image asset named for an interim Document Crunch and Trimble co-branded logo — the integration is visibly mid-flight.

What it does: construction-specific contract and specification review — reads the contract, flags risk provisions, notice deadlines, payment terms and spec compliance obligations, and routes them to the responsible person. Products named on the site are CrunchAI (risk identification across documents) and Project Assist (described as an agentic layer for document review and compliance automation). It tells a human where the risk is and where the clause lives. It does not negotiate and it does not serve notice.

The compliance question changed with the acquisition, and this is the question to put in writing:the Document Crunch site states "SOC 2 by default" and links a trust centre that is live but JavaScript-rendered and returned no readable certification list, with no report type stated on any page retrieved. So do not write that Document Crunch is SOC 2 Type II. Its parent, however, publishes a very complete trust portal listing ISO/IEC 27001:2022 with a Statement of Applicability, ISO/IEC 27701, ISO 9001:2015, SOC 1 Type 1 and 2, SOC 2 Type 1 and 2, SOC 3, NIST 800-171 Rev. 2, Cyber Essentials and Cyber Essentials Plus, GovRAMP, TX-RAMP, CMMC and WCAG. Those are Trimble corporate attestations. It does not automatically follow that the Document Crunch product is in scope of any particular report. Ask Trimble, in writing, whether the Document Crunch product is named in the SOC 2 Type 2 scope. That single question is worth more than a feature demo.

Attributed vendor figures: "more than 10,000 projects" and "500-plus general contractors and construction managers" come from Trimble's acquisition release and are vendor-published counts, not evaluations. Fits / does not fit: fits a contractor whose real exposure is missed notice deadlines and flow-down provisions. Does not fit a firm that wants to stay outside the Trimble ecosystem, because that is the direction of travel.

5. Togal.AI — Togal LLC

What it does: automated takeoff — detects, measures and classifies spaces and areas from architectural drawings so an estimator gets quantities without manual tracing. Aimed at precon and estimating. Independent, headquartered at 5959 Waterford District Drive, Miami, Florida; founder and CEO Patrick Murphy. The name is from the Irish tógálaí, builder.

What is verifiable, and it is rare here:a published price. The Growth tier is $299 per month per user, billed yearly, with unlimited automated takeoffs, chat prompts, image and symbol search and collaboration. The Business tier is custom-priced for teams of four or more and adds onboarding, dedicated support, a classification library template, quantity discounts, SSO — and "security compliance." That security compliance is a paid upgrade is itself a finding worth one line in your evaluation notes.

A correction we owe the reader: Togal.AI does hold SOC 2 Type II. There is no /securitypage — it returns 404 — but the company's news page announces the certification. A conventional desk check of one URL would have produced the wrong answer. That cuts both ways in this article and we say so again in the limitations.

What it actually is: computer vision on drawings. Not an agent. Not disclosed: whether a DPA is publicly offered, data residency, retention. Integrations with Procore, Bluebeam, PlanSwift, Destini, Ediphi and eTakeoff are reported; only the Ediphi native integration is confirmed by a Togal.AI announcement on its own news page, and no public API reference exists. Fits / does not fit: fits an estimating department drowning in manual tracing on repetitive building types. Does not fit anyone expecting it to produce a priced estimate — it produces quantities, and an estimator still owns the number.

The ten: integration, forecasting and capture (6-10)

6. Agave — San Francisco

What it does:started as a unified API across construction systems and moved up the stack into financial operations — ERP sync, AP invoice automation with AI matching, expense capture and coding from the jobsite, analytics, and vendor compliance marked "coming soon." It also ships an MCP server, letting a user interact with ERP and project-management data inside an assistant such as Claude or ChatGPT. That is a genuinely different position: Agave supplies the context layer other agents call, rather than competing to be the agent.

What is verifiable: the integration list is specific and documented — accounting connectors for Acumatica, CMiC, Deltek ComputerEase, Foundation, QuickBooks Online and Desktop, Sage 100 Contractor, Sage 300 CRE, Sage Intacct, Viewpoint Spectrum and Viewpoint Vista; field connectors for Autodesk (Forma) Build and Procore. It publishes a full developer centre with an API reference, auth flow and source-system guides. Note the seam: marketing moved to useagave.com while the docs remain on the old agaveapi.com domain. Funding: a $15M Series A announced July 2026 led by Accel with Y Combinator participation, reported total above $20M since late 2021, founded 2021.

Not disclosed: pricing — the site references "simple and flexible pricing" and routes to a demo booking. SOC 2 is indicated by an AICPA badge with no report type stated, and no public DPA or residency statement was located. Fits / does not fit: fits a contractor with a legacy ERP that everything else has to talk to. Does not fit a firm looking for a finished, field-facing product — this is plumbing, and it is honest about that.

7. Kojo — d/b/a Kojo, San Francisco

What it does: materials procurement and back office for specialty trades — requisition to purchase order to delivery to invoice to payment, plus warehouse, tool tracking and prefab. The AI component is invoice processing and AP automation. Workflow automation with document extraction; useful, and not an autonomous agent.

What is verifiable: the best-documented integration list in the roster, named on the vendor site — ERP and accounting connectors for Acumatica, CMiC, Deltek ComputerEase, QuickBooks Desktop and Online, Sage 100, Sage 300 CRE, Sage Intacct, Viewpoint Spectrum and Viewpoint Vista; project management for Autodesk, Access Coins Evo, eCMS, FOUNDATION and Procore; and supplier connections including Wesco, Mayer, Rexel, Hajoca, Graybar and Border States. A public security page states SOC 2, without the report type, with access to the report available to customers and partners on request. ISO 27001 is not claimed, and no DPA, residency or retention commitment is published.

One honest disclosure:Wesco International appears both in Kojo's supplier integration list and on its cap table, via a reported $10M Series C extension in October 2025. That is not a disqualification, and it is not hidden — but if you are relying on the platform to compare supplier pricing, it is a question worth asking directly. Vendor-published counts of "more than 600 contractors" and "$5 billion in materials annually" are exactly that.

8. nPlan — NPLAN LIMITED (UK)

What it does: ingests the project schedule from P6 or Microsoft Project and forecasts activity-level duration uncertainty against a large archive of historical schedules, producing risk-adjusted completion forecasts and flagging problematic schedule segments; it increasingly markets generative schedule creation and validation. This is predictive ML on schedule data — a forecasting engine, not an agent.

What is verifiable: corporate existence, from a statutory register rather than a vendor site. UK Companies House shows NPLAN LIMITED, company number 11043916, incorporated 2 November 2017, registered office 68 Hanbury Street, London E1 5JL, status Active. The vendor site states ISO 27001, SOC 2 and Cyber Essentials certification, which is the second-strongest compliance posture in this roster on vendor-page evidence — the underlying certificates were not independently retrieved.

Refused:the site's claim that it forecasts "with a degree of accuracy unattainable by human operators" is marketing, and it is unbenchmarked. Named clients and totals — 750,000+ historical schedules, $500bn of projects under active management — are vendor-published. Not disclosed: pricing. There is no pricing page; /pricing returns 404. Fits / does not fit: fits a large programme with a properly coded CPM schedule and a scheduler who can act on a probabilistic forecast. Does not fit a firm whose schedule is a bar chart maintained monthly — garbage in, confident forecast out.

9. Trunk Tools — Trunk Tools, Inc.

What it does:agents over the project document set, branded Cortex — TrunkText for question-answering across project documents, TrunkSubmittal for automated submittal review, an auto-generated Submittal Register, TrunkReview for drawing revision analysis, TrunkRFI for RFI drafting and TrunkBid for trade package bid analysis. In practice this is retrieval plus drafting: answering "what does the contract say about X" against a corpus, and producing a draft review. Nothing on the public site claims unattended execution.

What is verifiable: independent and venture-backed, with a $40M Series B led by Insight Partners reported in July 2025 and a reported total around $70M — treat the round as reported, not audited. Live syncs are asserted with Procore, Autodesk, SharePoint, Box, Dropbox and Egnyte, and it publishes submittal registers to Autodesk Forma and Procore. Named customer logos on the vendor homepage include Gilbane, Suffolk, Cleveland Construction, Haskell, AMLI Residential, Granite and Consigli — those are vendor-published logos, not references you have called.

What is not readable: a trust centre exists at trust.trunktools.com, hosted on Vanta, but the certification list is rendered client-side and returned only a page title to automated retrieval. No SOC 2 or ISO 27001 claim could be read from any public page, so we do not state that Trunk Tools is SOC 2 certified. Given the Togal.AI lesson above, a certification may well exist. The verifiable fact is that a buyer cannot see it without contacting sales. No API reference is public, and pricing is not published.

10. DroneDeploy — DroneDeploy, Inc. (pending acquisition, handle with care)

Status, precisely: independent as of 2026-08-23, but under a definitive merger agreement with Procore that has not closed. Procore's Form 10-Q for the quarter ended 30 June 2026states: "On July 27, 2026, we entered into the DroneDeploy Merger Agreement to acquire DroneDeploy for a purchase price of approximately $845.0 million in cash... The completion of this transaction is anticipated to occur later this year and is subject to the satisfaction of certain closing conditions." The same filing discloses a bridge loan entered the same day in connection with the financing, and lists among its risk factors the risk that the company is unable to complete the merger. DroneDeploy's own homepage carries a site-wide banner: it has signed a definitive agreement to be acquired by Procore. Note the two dates — merger agreement 27 July 2026, press release 29 July 2026.

Once the acquisition closes, we plan to immediately cross-sell DroneDeploy's solutions into our broader customer base and vice versa.
— Ajei Gopal, President and CEO of Procore, acquisition announcement, 29 July 2026

Read the tense. "Once the acquisition closes" is the cleanest possible proof that it has not. This is the single most defensible procurement observation available in this market right now: a buyer evaluating DroneDeploy in Q4 2026 is buying into an unresolved transaction, and the acquirer says so in a regulated filing.

What it does:from the vendor's own navigation on 2026-08-23 — DroneDeploy Aerial and DroneDeploy Ground, plus Robotics, Dock Automation and Industrial Inspection, with a visual-intelligence line of Safety AI, Progress AI and Earthworks AI. Capabilities as stated: earthworks analysis, automated schedule tracking and quantification, visual defect detection ranked by severity, and automated safety risk identification "from missing toe boards to tripping hazards." Trade coverage listing "Inspection AI" and "Embodied AI" as product names is not corroborated by the vendor site, so those two names do not appear here.

What it actually is: capture plus computer vision with an alerting layer. The vendor does use agent language — AI agents proactively surfacing schedule or safety risks — but surfacing a risk to a human is detection and alerting, not autonomous action. Refused: the vendor's ROI calculator. A vendor-built estimator is not evidence of ROI; it is a lead form with arithmetic. What to ask before signing: what happens to pricing, roadmap and integration neutrality at close, in writing, with a term that survives the transaction.

Read the ten together and a pattern emerges that no vendor will put on a slide. The products that are genuinely closest to agentic — Procore's agent library, Trunk Tools' Cortex, Document Crunch's Project Assist — all sit on top of the project document set and all stop at a draft. The products that are most technically impressive — ALICE's optioneering, nPlan's forecasting, Buildots' and Doxel's progress computation — are optimisation and prediction systems that mostly predate the agent framing and were renamed into it. And the products that are most commercially transparent — Togal.AI, Handoff — are the narrowest. Transparency, capability and autonomy are three different axes in this market, and no product is top of all three.

Four numbers that should not travel

Four statistics carry nearly every construction AI pitch, and three of them are weaker than they look. We chased each to its actual publication rather than to a blog citing it. If your vendor opens with one of these, you now know more about it than they do.

98% of megaprojects run over budget — refuse it, and the source is not Flyvbjerg

The number is usually attributed to Bent Flyvbjerg. It is not his. It comes from McKinsey's June 2015 report The construction productivity imperative, page 1, which states: "Our research estimates that 98 percent of megaprojects suffer cost overruns of more than 30 percent; 77 percent are at least 40 percent late." No sample size. No dataset. No method. No external citation. And note the qualifier almost every repetition drops — it is not 98% over budget, it is 98% over budget by more than 30 percent, a far stronger and far less plausible claim than the version in circulation.

What Flyvbjerg actually published is better and citable. Flyvbjerg, Holm and Buhl (2002), in the Journal of the American Planning Association, examined 258 transportation infrastructure projects worth $90 billion and found that costs are underestimated in almost nine out of ten projects; for a randomly selected project the likelihood of actual costs exceeding estimates is 86%, and actual costs average 28% higher than estimated. His 2014 overview in Project Management Journal gives the iron law of megaprojects: over budget, over time, over and over again. Note also that the widely quoted 99.5% and 8.5% figures come from a 2023 trade book, not a refereed paper. Print the 86% and the citation.

"Construction productivity has been flat for sixty years" — genuinely contested

This one is not a fabrication, but the slogan is a self-citation and the underlying reality is disputed by the statistical agency that would have to certify it. McKinsey's 2015 report says productivity has been flat for decades "according to McKinsey research." The 2017 MGI report says productivity in construction has barely increased at all since 1945. Neither publishes a dataset or method.

Evidence for the pessimistic view is real: BEA value added per worker in construction was roughly 40% lower in 2020 than in 1970 (Goolsbee and Syverson, NBER Working Paper 30845 — a working paper, not peer-reviewed), and BLS has described construction as the only major industry with negative average productivity growth since 1987. But BLS's own economists documented the measurement problem in the Monthly Labor Review in January 2018: the deflators typically used to measure construction output have not matched the output they are being used to deflate, because there is no good price index for a unique built asset.

And here is the correction almost nobody makes. BLS does not publish a productivity measure for the construction sector as a whole.It publishes labour productivity for four industries covering roughly 12.5% of the sector's employment. For 2024: single-family residential +6.1%, multiple-family residential −12.8%, industrial building +16.0%, and highway, street and bridge declining in every year from 2021 to 2024. Independent tests — including García and Molloy (2025), peer-reviewed in Regional Science and Urban Economics— find an upward bias in the deflators "but the magnitude is not large enough to alter the view that construction-sector productivity growth has been weak." The honest summary is that the pessimistic story survives, but not in the form the slide deck uses, and the agency's own subsector series show positive growth in three of the four industries it covers.

"The industry wastes $1.6 trillion" — safe only if stated precisely

MGI's Reinventing Construction (February 2017) says: "If construction productivity were to catch up with the total economy, the industry's value added could rise by $1.6 trillion a year." That is annual, global, and a conditional counterfactual projection. It is not a measured loss and not money currently on the floor. MGI's own summary compresses it to "there is a $1.6 trillion opportunity," which is how it circulates as a waste figure — a reading the report does not support. Use it only in the conditional form, and never as an addressable market.

"Rework costs 20% of project value" — no source; the real number is smaller and narrower

We could not locate any source for 20% in the Construction Industry Institute's research or the peer-reviewed literature. Hwang, Thomas, Haas and Caldas (2009)measured 359 projects from CII's benchmarking database and reported a mean Total Field Rework Factor of 5.4% with a standard deviation of 6.2% — by group, buildings 4.6%, heavy industrial 4.4%, infrastructure 5.7%, light industrial 9.3%. Three restrictions travel with it: field rework only, as a share of construction-phase cost rather than total project value, and a CII-member sample that is large, sophisticated and industrial-weighted. The distribution is heavily right-skewed, so the mean is not typical, and the data is now 15 to 25 years old. Also worth knowing: the famous "5%" line in that paper's abstract is the authors citing CII, not their own finding — a literature-review sentence that acquired a peer-reviewed veneer by sitting in an ASCE abstract.

A fifth family of claims we refuse outright, because no methodology survives contact with any of them. "95% of GenAI pilots fail" rests on 52 interviews and 153 conference surveys of custom-built tools. "85% of AI projects fail" misquotes a 2018 prediction about output quality. "87% never reach production" began life as a rhetorical question in a sponsored post. Gartner's "40% of agentic projects cancelled by 2027" measured investment posture, not outcomes. All four circulate heavily in construction-tech marketing, on both the optimistic and the pessimistic side. We do not print them, and none belongs in a procurement decision.

One number we will print, with its provenance attached: the Associated Builders and Contractors estimated on 15 January 2026 that the industry needs to attract about 349,000 net new workers in 2026 and 456,000 in 2027. ABC publishes its methodology, which converts forecast construction spending into labour demand at roughly 3,450 new jobs per billion dollars of additional spending, plus projected retirements. Two things must travel with it: it is a modelled demand gap rather than a measured shortage, and ABC itself says a majority of 2026 demand is attributable to retirement rather than to growth in demand for construction services. ABC is a trade association with an advocacy interest in workforce policy. Attribute it; do not launder it.

The binding constraint: contract, seal, risk

On a construction project the contract documents are the source of truth — not the model, not the field data, and not the agent. An AI agent cannot be a party to the contract, cannot bear the risk of a schedule or an estimate it produced, and cannot hold the professional licence required to seal engineered work. Every agent output is therefore a draft that a named, accountable human adopts. That single sentence caps how much of a construction workflow any agent can actually own, and it is not a technology limitation that will be engineered away next year.

Start with the contract. AIA A201, General Conditions of the Contract for Construction, current edition A201-2017, is the backbone of most AIA-family contracts, paired with the owner-contractor agreements and the change order and payment application forms. ConsensusDocsis the competing family — 116 documents across general contracting, design-build, subcontracting, program management, CM at-risk, collaborative delivery and P3, endorsed by a coalition of roughly 45 industry associations. Here is the finding: neither family's public materials contain an AI-specific provision that this research could locate. The standard-form contracts that allocate risk on nearly every US project have not been updated to say who owns the consequences of an AI-generated submittal review, RFI response or schedule. That is a gap we found, not a claim that no such clause exists anywhere. It means the allocation defaults to the existing terms — which is to say, to you.

Then the doctrine, which is 107 years old and still governs. In United States v. Spearin, 248 U.S. 132 (1918), decided 9 December 1918, opinion by Justice Brandeis, the Court held — in the opinion, not the syllabus — at 248 U.S. 136:

But if the contractor is bound to build according to plans and specifications prepared by the owner, the contractor will not be responsible for the consequences of defects in the plans and specifications.
— United States v. Spearin, 248 U.S. 132, 136 (1918), Opinion of the Court

And at 137, the Court added that the implied warranty "is not overcome by the general clauses requiring the contractor, to examine the site, to check up the plans, and to assume responsibility for the work until completion and acceptance," and that "the duty to check plans did not impose the obligation to pass upon their adequacy to accomplish the purpose in view." The countervailing rule sits on the same page: "Where one agrees to do, for a fixed sum, a thing possible to be performed, he will not be excused or become entitled to additional compensation, because unforeseen difficulties are encountered."

Spearin allocates design risk to whoever furnished the design. An AI agent furnishes nothing. It has no legal personality, no capital and no insurance. When an agent drafts a submittal review that misses a spec conflict, or produces a schedule that proves undeliverable, the risk lands exactly where the contract already put it: on the party whose employee adopted the output. Using an agent does not move a single unit of risk. It changes how fast the paperwork is produced.

Then licensure, which is the strongest version of this argument because it needs no AI-specific rule at all.NCEES Model Law § 110.20(E) defines responsible charge as exercising full professional knowledge of and control over work. A licensee in responsible charge must have and exercise the authority to review and to change, reject or approve both the work in progress and the final work product; must be personally aware of the project's scope, needs, parameters, limitations and special requirements; must be capable of answering questions relevant to the engineering or surveying decisions made; and must accept full responsibility for the work. Model Rules § 240.20(D) — not § 240.30, which is continuing professional competency — provides that the seal and signature shall be placed on work only when it was under the licensee's responsible charge, and that the board has final authority regarding the determination of whether work was actually prepared under a licensee's responsible charge.

State law says the same thing in enforceable terms. Texas Occupations Code § 1001.402 provides that a public official may accept a plan only if it was prepared by an engineer, as evidenced by the engineer's seal; 22 TAC § 137.33(b) states that upon sealing, engineers take full professional responsibility for that work; and 22 TAC § 131.2(11) defines direct supervision as the engineer personally making decisions or personally reviewing and approving proposed decisions prior to their implementation. California Business and Professions Code § 6703 requires the independent control and direction, by the use of initiative, skill and independent judgment. On the architecture side, NCARB's "Model Law and Regulations," July 2024 — the current title; "Legislative Guidelines and Model Law/Model Regulations" is superseded — provides at § 401(2) that by sealing a Technical Submission the architect represents that the architect was in Responsible Control over the content, and Model Regulation R401.1(1) is a closed list of what an architect may seal. Every branch of that list terminates in a licensed human. There is no category for "prepared by a tool."

One board has actually addressed AI, and what it said is more useful than a ban. Texas Board Policy Advisory Opinion No. 71, adopted 14 November 2024, states that neither the Texas Engineering Practice Act nor the associated Board rules directly ban the use of AI software, that it is acceptable for licensees to use it, and that licensees are ultimately responsible for any work product they sign and seal. It then concludes that no new Policy Advisory Opinion will be developed, because the Act and Board rules adequately address the use of artificial intelligence software at this time. Texas considered whether AI needed a new rule and decided the existing sealing, competence and confidentiality rules already cover it. Florida reached the same place thirty-three years earlier by writing technology-neutrally: Fla. Admin. Code R. 61G15-30.008, in force since 1993, provides that the engineer shall be responsible for the results generated by any computer software and hardware that he or she uses in providing engineering services.

Be careful not to read a trend into that. PAO 71 is the only clean AI-specific board opinion we located. NCEES has issued no AI position statement, and its August 2025 Model Law and Model Rules contain zero occurrences of "artificial intelligence," "machine learning" or "software." NCARB's July 2024 document mentions none of them either. The absence is not an oversight — the licensure regime already terminates responsibility in a natural person, so it does not need to name the tool.

Finally, the practical constraint underneath the legal one.Every computer-vision progress tool in this roster computes progress by comparing captured reality against a model or a schedule, and that comparison inherits the error of both inputs. If the BIM model is not on the current revision, or the schedule's activity coding does not match how the trades actually sequence work, the deviation the tool reports is an artifact of stale baseline data rather than a site condition. Vendor behaviour corroborates this: Buildots sells fully managed site captures, and Doxel and OpenSpace likewise sell capture as a service or supply the capture hardware. A category that needs to control its own data collection is telling you where the failure point is. The honest evaluation question is not "how accurate is the AI" but "who owns keeping the model and the schedule current, and what happens to the output when they drift?" No vendor answers that on a public page.

Regulatory status, stated precisely

Four regulatory items get misstated constantly in construction-tech content, and each of them has a version that sounds right and is wrong. Status precision matters here: proposed, comment period closed, final, in force, enjoined and vacated are different states, and a proposed rule is not law.

OSHA heat: proposed, not final — and the next step is another proposal

There is no federal heat standard. OSHA's heat injury and illness prevention rule is a proposed rule. The NPRM was published on 30 August 2024 at 89 FR 70698 under RIN 1218-AD39, covering 29 CFR Part 1926. The comment period was extended to 14 January 2025 — the Federal Register record also lists a 30 December 2024 close, so cite the extended date and note the extension. An informal public hearing ran from 16 June to 2 July 2025, and the post-hearing comment period closed on 30 October 2025.

The current Unified Agenda entry for RIN 1218-AD39 places the rule at Proposed Rule Stage with a legal deadline of None, and a timetable projecting a Supplemental NPRM in December 2026 and final action in October 2027. There is no statutory deadline compelling finalisation. So the sentence "OSHA will finalize the heat rule in 2026" is misleading even when it appears in a departmental plan — the next step is another proposal. Until then, heat is enforced through the General Duty Clause, OSH Act § 5(a)(1), and through state plans. If a vendor sells you heat monitoring on the basis that a federal standard takes effect in 2026, they have not read the docket.

For contrast, here is a construction rule that is final: Personal Protective Equipment in Construction, published 12 December 2024 at 89 FR 100321 and effective 13 January 2025, requiring PPE that properly fits each employee. Knowing the difference between a rule that exists and one that does not is most of compliance literacy.

Recordkeeping: most construction firms file the 300A only

Under the 2023 Improve Tracking final rule and 29 CFR Part 1904, establishments with 100 or more employees must electronically submit Forms 300 and 301 in addition to the 300A summary — but only if their industry appears in Appendix B to Subpart E. Appendix B contains exactly one construction NAICS code: 2381, Foundation, Structure, and Building Exterior Contractors. Appendix A lists NAICS 23 as a whole, which is the 300A-only pathway. So general contractors (236), heavy and civil engineering (237) and specialty trades (238) other than 2381 submit the 300A only. The blanket claim that large construction contractors must file their 300 logs and 301 incident reports is wrong for most of the industry.

The deadline is 2 March of the following year, and OSHA accepts submissions by web form, CSV upload or API. That API is one of the few places in this article where an agent could legitimately automate a filing end to end — and it is still a human-attested record, which is exactly why the attestation stays with a person. 29 CFR Part 1904 is otherwise stable: the only 2025-26 activity was a withdrawal terminating the musculoskeletal-disorder column proposal, which does not affect existing obligations.

Davis-Bacon: three provisions vacated, not struck down and not enjoined

The Davis-Bacon Act applies to federally funded construction contracts in excess of $2,000 under 40 U.S.C. § 3142(a) — a threshold that has not been inflation-adjusted since 1935. Certified payroll is a Copeland Act obligation, 40 U.S.C. § 3145, with 18 U.S.C. § 1001 applied to false statements. 29 CFR 5.5(a)(3)(ii)(A) requires weekly submission and provides that the prime contractor is responsible for the submission of all certified payrolls by all subcontractors; 29 CFR 5.5(a)(3)(ii)(B) provides that full Social Security numbers and addresses must not appear on the weekly transmittals. Records are retained three years after completion of all work on the prime contract — not four. And Form WH-347 is optional: weekly submission and the Statement of Compliance are mandatory, the format is not.

Now the litigation, because most summaries are stale. In AGC of America v. DOL, No. 5:23-cv-00272 (N.D. Tex.), a nationwide preliminary injunction issued on 24 June 2024. Two years later, on 24 June 2026, final judgment vacated three provisions under APA § 706(2)(A) and (C): the truck-driver item at 29 CFR 5.2(2)(iv)(D), the material-supplier carve-back at 29 CFR 5.2(2), and 29 CFR 5.5(e) — incorporation by operation of law — in its entirety. The vacatur is nationwide, not party-limited. The Fifth Circuit appeal was voluntarily dismissed on 26 June 2026. Nothing is pending.

Five corrections follow. The rule was not struck down — three provisions of a 221-page rule were vacated. They are not enjoined; an injunction is not a vacatur. The site-of-the-work expansion survives, because plaintiffs dropped that claim: 29 CFR 5.2(2)(iv)(A), (B), (C) and (E) all remain in force. The 30 percent rule in 29 CFR 1.2 was never challenged. And it is not on appeal.

Two traps that will actively mislead your compliance team.The eCFR still prints all three vacated provisions verbatim, with no editorial note — vacatur is not self-executing, so a practitioner reading the CFR today gets the wrong answer. And DOL's own page still describes the 2024 preliminary injunction rather than the 2026 vacatur. If you are building an agent that reads regulatory text to answer prevailing-wage questions, that is not a hypothetical failure mode. It is a live source of confidently wrong answers, and it is the single best argument in this article for keeping a human between the retrieval and the advice. Do not conflate any of this with EO 14026, the $15 contractor minimum wage revoked by EO 14236 — different statute, different rule.

Drones: Part 108 is [Reserved], and BVLOS runs on individual authorisations

Three of the vendors here sell drone-based capture, and Procore is paying approximately $845 million for one of them partly on that strength — so this constraint is load-bearing. 14 CFR Parts 108 and 109 are marked [Reserved] in the current CFR. There is no Part 108. The FAA published its BVLOS NPRM on 7 August 2025 at 90 FR 38212 under RIN 2120-AL82. Comments closed; an extension was denied on 29 September 2025; the period was reopened on 28 January 2026 until 11 February 2026, limited to electronic conspicuity and right-of-way; a further extension was denied on 10 February 2026. No final rule exists as of 2026-08-23.

The statutory deadline is commonly miscited. It is not § 2209. It is § 930 of the FAA Reauthorization Act of 2024, Pub. L. 118-63, codified at 49 U.S.C. § 44811. The FAA missed the NPRM deadline by roughly eleven months; the final-rule milestone of about 7 December 2026 is not yet due. Do not write that the FAA blew the final-rule deadline — those are different obligations.

What authorises BVLOS today is individualised: a Part 107 waiver, since § 107.31 prohibits BVLOS and § 107.205(c) allows a waiver — though no waiver of that provision will be issued to allow the carriage of property of another by aircraft for compensation or hire — or a 49 U.S.C. § 44807 exemption, whose sunset was extended to 30 September 2033 and is therefore not expiring. Two things are already fully in force over your jobsite: Remote ID under 14 CFR Part 89, where the enforcement-discretion window closed on 16 March 2024, and the operations-over-people rules. A jobsite is full of non-participating people; § 107.39(a)'s "directly participating" exception covers the flight crew, not the trades working below, so a construction drone flight typically needs a Category 1-4 aircraft or a waiver in addition to anything BVLOS-related.

There is a structural parallel worth noticing. 14 CFR § 107.19(b) provides that the remote pilot in command is directly responsible for and is the final authority as to the operation of the small unmanned aircraft system. Aviation law, like engineering law, terminates responsibility in a named natural person. That is the same architecture as the professional seal, arrived at independently, and it is a good sign that the pattern is not going to be engineered away.

One more item to put on your counsel's list rather than in your architecture: California's automated decision-making technology regulations under the CCPA expressly count "allocation or assignment of work for employees" as a significant decision, with a compliance date of 1 January 2027 for businesses already using ADMT for such decisions. If your agent assigns crews to tasks in California, that is a scoping question worth asking now rather than in December 2026.

A worked example, and one real project

Here is an illustrative scenario, framed unmistakably as one. Consider a general contractor running four active commercial projects, with a project engineering group of six people processing roughly 240 submittals and 90 RFIs a month across the portfolio. Assume, from your own timesheets rather than from a vendor, that first-pass submittal review takes 35 minutes and that drafting an RFI from a field note takes 25 minutes. That is 140 hours a month of submittal review and 37.5 hours of RFI drafting — about 178 hours, or a little over one full-time equivalent.

Now apply the constraint honestly. Both platform vendors state that a human reviews and approves every output, so the agent does not remove the review — it changes what the reviewer starts from. If a pre-review that flags spec conflicts, missing product data and expired submittal deadlines converts a 35-minute cold read into a 20-minute checked read, the portfolio saves 60 hours a month on submittals. If RFI drafting drops from 25 minutes to 12, that is another 19.5 hours. Roughly 80 hours a month, or about half an FTE — real, but not the headcount elimination the category markets. And it is contingent on something the vendor cannot supply: a document set that is actually current, with the correct spec revision loaded.

Note also what the arithmetic does not include. Someone has to keep the corpus current. Someone has to review the agent's misses, which are more expensive than its hits, because a missed notice deadline is not a productivity problem. And the reviewer's attention changes character: checking a plausible draft is a different cognitive task from reading a document cold, and it is easier to nod through. Build your acceptance criteria around the misses, not the throughput.

One real project, and it is deliberately not an AI story. Our published Pergola Cave 3D configurator case study describes work for a licensed contractor in Burbank who manufactures and installs motorised aluminium pergolas across Los Angeles County, on jobs the case study puts at $15,000 to $50,000-plus. The problem was not a lack of AI. It was that customers could not specify what they wanted, so the sales team was fielding 50 to 70 calls a week answering the same configuration questions, quotes took three to five days, and specifications arrived as vague descriptions that needed interpretation. We replaced a WordPress site with a custom React and Three.js configurator that captures the exact specification at intake. The case study reports quote requests rising from 12-15 a month to 34-42, phone volume falling from 50-70 a week to 18-25, a four-month timeline and a $65,000 project cost. Those are our own reported figures from our own engagement — not an independent evaluation, and we will not pretend otherwise.

The reason it belongs in an article about agents is the lesson underneath it. The win came from structuring the input, not from adding intelligence at the end. Nearly every construction agent failure we have seen traces to the same place: an agent asked to reason over a document set, schedule or model that nobody owns keeping current. Fix the intake first. It is cheaper, it is verifiable, and it is what makes an agent worth buying afterwards.

What breaks first

Construction agents fail in a small number of recognisable ways, and none of them announce themselves. Each row below pairs the failure with the signal that detects it and the rollback that contains it. Design the detection before you deploy, because the default failure mode of a drafting agent is silent plausibility.

Failure mode	How it shows up	Detection signal	Rollback
Stale baseline	Progress or deviation reports drift from what the superintendent sees on site	Reconcile agent output against the current drawing revision and schedule update date on every report; alert when the baseline is older than the last issued revision	Suspend automated reporting; revert to manual walk until the model and schedule are re-based
Prompt injection via project documents	An agent follows instructions embedded in a submittal, transmittal, supplier email or PDF attachment	Log every document the agent read alongside every action it proposed; flag any output referencing an instruction not present in your own prompt	Read-only credentials by default; no outbound communication or financial action without human approval
Confidently wrong regulatory answers	The agent cites vacated or superseded text — the eCFR still prints the three vacated Davis-Bacon provisions with no editorial note	Require a citation with a retrieval date on every regulatory answer; spot-check against the docket rather than the code text	Remove regulatory question-answering from the agent's scope; route to counsel
Silent plausibility in review	Reviewers approve drafts faster over time and catch rates fall; misses cluster in unusual conditions	Sample 10% of approved outputs for blind re-review by a second person monthly; track misses, not throughput	Reinstate cold review for the affected document class
Integration drift	A connector breaks after a platform release and records stop syncing without an error surfacing to the field	Daily record-count reconciliation between systems; alert on a zero-delta day	Fail closed rather than open — stop the workflow rather than proceeding with partial data
Vendor status change mid-term	Pricing, roadmap or integration neutrality shifts after an acquisition closes	Track your vendors on EDGAR full-text search and set a calendar reminder at each renewal	Negotiate change-of-control terms, data export and a defined transition period before signing

Prompt injection deserves its own paragraph because it is the failure this industry is least prepared for and it is unsolved. An agent that reads submittals, supplier emails, subcontractor invoices, transmittals and RFI attachments is consuming untrusted documents supplied by parties with an active financial interest in the outcome — which is a materially worse threat model than an internal knowledge assistant. The OWASP Top 10 for LLM Applications names the failure modes; in the 2025 edition published as current, prompt injection is LLM01, improper output handling is LLM05 and excessive agency is LLM06. Be careful with the numbering: three different numberings are live on the internet right now, including a legacy 2023 page still serving different identifiers, so name the risks rather than trusting an ID you saw in a summary. Treat all of it as blast-radius reduction rather than a solved problem — restricted permissions, no autonomous financial or contractual action, human approval before anything leaves the organisation, and a complete audit trail. The NIST AI Risk Management Framework(AI 100-1, January 2023) and its Generative AI Profile (AI 600-1, July 2024) give you a voluntary structure — Govern, Map, Measure, Manage — to hang that on. Voluntary, not mandatory; say so when a vendor implies otherwise. Procore's trust centre states its policies are based on the NIST frameworks, which is a claim you can actually test against the standard.

The safety-observation category deserves a specific caution. Fall protection under 29 CFR 1926.501 is the most frequently cited OSHA standard across all industries in FY2025, and five of OSHA's Top 10 are construction standards; falls, slips and trips accounted for 370 of the 1,032 fatal injuries among construction and extraction occupations in 2024. A vision system that flags an unprotected edge is addressing a real hazard. But detection is not abatement, and an alert nobody acts on within the shift is a liability record rather than a safety control. The workflow design question — who receives the alert, what they are required to do, and how the closure is documented — matters more than the detector. We work through that design in more depth in our guide to AI agents for construction safety and compliance.

The human-in-the-loop boundary

Write the boundary down before you deploy, in the same document as the scope. Both platform vendors state their own version of it, so you are not inventing a standard — you are adopting one they already published.

Task	May the agent act alone?	Who must review or sign	Basis
Draft an RFI, fill a daily log, pre-review a submittal, assemble a submittal register	Prepare only	Project engineer or PM approves before anything leaves the organisation	Procore: your field and office teams review and approve every output
Select among generated design alternatives or build sequences	Generate options only	The professional directs the system and controls the final product	Autodesk: the professional still gives instructions and makes selections
Answer an RFI that changes engineered work; seal a drawing or calculation	Never	A licensed professional in responsible charge, personally aware of the project and able to answer questions on the decisions made	NCEES Model Law 110.20(E); Model Rules 240.20(D); NCARB Model Law and Regulations, July 2024, 401(2)
Adopt a schedule as the contract schedule	Never	The contractor of record	Spearin risk allocation; the contractor remains responsible for means, methods and the accuracy of the schedule it submits
Sign a certified payroll	Never	A human signs the Statement of Compliance under penalty of law	Copeland Act, 40 U.S.C. 3145; 18 U.S.C. 1001; 29 CFR 5.5(a)(3)(ii)(A)
Submit the OSHA 300A electronically	May prepare and queue the submission	A named person attests to the record before it is filed	29 CFR Part 1904; OSHA accepts web, CSV or API submission
Issue a notice, a change order or a claim	Never	Whoever holds contractual authority under the executed agreement	No AI-specific provision exists in AIA A201-2017 or ConsensusDocs; the default allocation applies
Fly beyond visual line of sight for site capture	Never, absent an authorisation	The remote pilot in command is directly responsible and is the final authority	14 CFR 107.19(b); Part 108 is [Reserved]; BVLOS requires a 107.205(c) waiver or a 44807 exemption

The shape to internalise is this: an agent may prepare, assemble, retrieve, flag and queue. A named human adopts, signs, seals, submits and serves. If a vendor's roadmap moves any row in the third column into the second, ask which licence the software holds and which insurer stands behind it. There is no answer to that question, which is why it is the right question.

Cost and timeline

If you decide a bought product does not fit the workflow — which is common when the workflow crosses an ERP, a project management platform and a document set nobody owns — here are our published bands. We do not quote outside them, and we do not estimate a vendor's price for you.

Engagement	Range	Timeline
Discovery + workflow audit	$9k–$22k	2–4 weeks
Single-workflow agent	$28k–$70k	4–9 weeks
Multi-workflow platform with system integration	$70k–$180k	9–16 weeks
Enterprise / multi-site / regulated build	$180k–$420k+	14–24 weeks

Senior-led work runs $150–$225/hr and retainers run $2,500–$9,500/month. Every build carries a 30-day post-launch warranty, and full source-code and IP ownership transfers to you. After a discovery call we send a fixed-price phased proposal within 5 business days. We are a senior-led, Black-owned agency in Los Angeles, reachable at +1 (424) 272-5601 or at our discovery call booking page. If a scoped build is not the right first step, our AI agent creation service page sets out what a discovery-only engagement covers.

A note on sequencing that will save you money. Buy the discovery before you buy the platform. The most expensive construction AI failures we see are not bad models; they are twelve-month platform commitments made before anyone established who owns the document set, which drawing revision is authoritative, and whether the schedule's activity coding matches how the trades sequence work. Those three answers determine whether any product in the table above will work for you, and none of them cost six figures to find out.

One market observation for context, stated carefully. Total construction put in place ran at a seasonally adjusted annual rate of $2,166.5 billion in June 2026 — that is an annual rate, not year-to-date spend — down 3.2% (±1.5%) from June 2025, a statistically significant year-over-year decline; first-half 2026 actual spend of $1,046.9 billion was 3.5% below the first half of 2025. Contractors are buying efficiency tooling into a contracting market, not a boom. That is an argument for scoping tightly and proving one workflow, not for a platform migration.

Not sure whether to buy or build?

A discovery and workflow audit tells you which of the ten products above actually fits your document set and your ERP — and whether a scoped build is cheaper than the licence.

Schedule Free Consultation (424) 272-5601

Message on WhatsApp

Red flags when evaluating a vendor

Every one of these is something you can check in a first call, and every one of them appeared somewhere in this roster during research. None of them is disqualifying on its own. Several of them together tell you what the implementation will feel like.

An ROI calculator instead of a case study: A vendor-built estimator is a lead form with arithmetic. It encodes the vendor's assumptions and returns them to you as a finding. Ask instead for three references on projects of your type and delivery method, and call them.
A pricing page that is a qualification questionnaire: One vendor's pricing page asks company type, project size and geography before showing any figure. That is not pricing; it is segmentation. Two of the sixteen products examined publish an actual number.
A trust centre you cannot read: A client-rendered certification list behind a request-access flow means you cannot verify a security claim before a sales conversation. It may well be genuine — but it is a fact about the buying process you should price in.
Certifications announced in a blog post rather than on a trust page: This cuts both ways: one vendor here holds SOC 2 Type II and announces it only in a news item, with /security returning 404. Search the news page before concluding a certification does not exist.
Parent-company attestations presented as product attestations: After an acquisition, the parent's ISO and SOC reports are not automatically in scope for the acquired product. Ask, in writing, whether the specific product is named in the SOC 2 Type 2 scope.
Integrations asserted on a marketing page with no API reference: Thirteen of the sixteen products examined publish no public developer documentation. 'Live syncs with Procore' on a homepage and a documented, versioned API are different commitments.
Agent language over a detection or optimisation product: If the demo shows a dashboard that surfaces a risk to a human, that is detection and alerting. Useful — but price it as monitoring software, not as labour replacement, and do not let the word 'agent' set the budget.
An accuracy figure with no methodology, sample or control group: Ask three questions: what was the denominator, who ran the evaluation, and can you show me the protocol. In this market the answers are usually unavailable, unavailable and no. That is the correct time to move on to verifiable attributes.
Silence on the model and data path: Ask which model providers process your project documents, where the data goes, what is retained and for how long, and whether your content trains anything. One vendor here discloses its model providers by tier; most do not.
No answer on change of control: Four of the biggest names in this market changed ownership or name in the last eight months, and one is mid-transaction right now. Ask what happens to pricing, roadmap and your data export rights if the company is acquired during your term.

Limitations: what we could not verify

This is a desk audit conducted on 2026-08-23, and it has real limits. Naming them is part of the method, and it is the section we would read first if someone else had written this.

No independent benchmark exists for any product in this roster. Not one has a third-party evaluation of output quality. Every accuracy, ROI and time-saved figure in this market is published by the seller about itself. If that changes, this ranking should change with it.
Absence of a readable certification is not proof of absence. Togal.AI holds SOC 2 Type II and announces it in a news post while /security returns 404. Checking one conventional URL would have produced a wrong answer, and the same may be true for others here.
Compliance and pricing were not verified for every vendor. Autodesk, DroneDeploy, EarthCam, ALICE, Bild AI and Handoff have unverified compliance cells. Those cells say not verified, not none.
The exact date of the Autodesk Construction Cloud to Forma rename is contested across trade sources — 17 February versus 31 March 2026 announcement, 24 versus 25 March effective — and Autodesk's own announcement page returned HTTP 403 to automated retrieval. The rename itself is confirmed on Autodesk's product page.
Buildots' funding totals are irreconcilable across sources ($121M versus $166M), so no total is printed. Trunk Tools' reported total of roughly $70M is vendor-reported, not audited.
What happened to SafeAI could not be determined. The domain redirects to pronto.ai, which does not mention SafeAI. Acquisition, asset sale or wind-down — unknown.
Vendor-status databases actively disagreed with vendors' own sites. One aggregator listed Doxel as inactive while Doxel's site was live, hiring and carrying a 2026 copyright. A search summary asserted nPlan was publicly traded under a Stockholm ticker; the UK statutory register shows a private limited company. Both would have shipped as errors if taken on trust, and both are exactly the failure mode this article is about.
Buildots' acquisition of Genda is trade-press-sourced and is not mentioned on Buildots' own site; the Buildots Field product line is consistent with the capability but is not confirmation of the deal.
The construction productivity question is genuinely unsettled rather than merely under-reported, and we have presented both sides rather than picking the convenient one.
Regulatory status was checked against Federal Register documents, the eCFR and agency pages on 2026-08-23. Rulemaking moves. Re-check the docket before relying on any status here, particularly the OSHA heat SNPRM projected for December 2026 and the FAA BVLOS final-rule milestone around 7 December 2026.
BLS blocks automated retrieval, so those figures were obtained through the BLS public API and a rendering proxy against canonical bls.gov URLs, which are the URLs cited.
We did not test any product. This is an audit of what a buyer can verify before signing, not a bake-off — and a bake-off is precisely what this market still lacks.

If you take one thing from this page, take the habit rather than the ranking. Open the pricing page. Open /security. Search EDGAR for the vendor name. Ask for the SOC 2 report and the DPA in writing. Ask whether the product is in scope of the parent's attestation. Ask what happens at change of control. Every one of those questions is answerable in an afternoon, and together they tell you more than any accuracy claim in this category — because there is no independent benchmark behind the accuracy claim, and there is a public record behind every one of these.

Want an agent scoped to one workflow, not a platform migration?

We build single-workflow construction agents with the approval boundary designed in from day one — submittal pre-review, RFI drafting, invoice coding, certified-payroll assembly. Senior-led, fixed-price, full IP transfer.

📅 Schedule Consultation 📞 WhatsApp

1517 S Bentley Ave Unit 204, Los Angeles CA 90025

Top 10 AI Agents for Construction Firms in 2026:Ranked Only on What You Can Verify Yourself

Table of Contents

Key Takeaways