Top 10 AI Agents forFreight and Logistics in 2026

Every ranking of freight AI agents you can find scores the same handful of things — accuracy, autonomy rate, loads covered, hours saved, ROI — and not one of those numbers has ever been measured by anyone other than the company selling the product. That is not a cynical framing. It is a finding. We went looking for an independent evaluation of AI agents in freight brokerage: a third-party benchmark, an academic study, an analyst test harness, a trade-association comparison. There is none. A site search of the largest trade publication in the sector for independent AI-agent evaluation returned nothing matching. The closest thing that exists is an industry award with no published methodology and no test set.

So this article does something different. It ranks ten products on attributes you can re-check yourself in an afternoon — who owns the company today, which integrations are named in public documentation, which security attestations are actually published and of which type, whether pricing is disclosed, whether the product is standalone or locked to a network — and it says out loud what it refused to score and why. If you want a number telling you which agent books more loads, no honest source can give you one. If you want to know which vendors will still exist under the same name at renewal, which will hand you a SOC 2 Type II report, and which one of them is not an agent at all, that is checkable, and this is that list.

Three things changed under this vertical in 2026, and each one breaks most of the rosters currently circulating. First, on 14 May 2026 the Supreme Court decided broker liability, unanimously, and the federal preemption defence that protected brokers from negligent-selection claims in several circuits is gone. Second, on 14 July 2026 project44 split into two companies, moving the entire broker and 3PL agent business into a new entity called LSP44 — so a broker being pitched project44 today is being shown the wrong half. Third, four of the best-known names in freight tech have been acquired and all four still trade under their original brands, which is precisely why listicles keep ranking them as independent.

This piece sits inside a wider set of vertical rankings; the cross-industry view of how these products differ by market lives in our overview of the top AI agents across industries in 2026. What follows is the freight-specific version, written for a broker or a 3PL operations lead who has to sign something.

We scored only attributes a buyer can independently re-check, and every one of them was checked on 23 August 2026. Nothing here is scored on performance, because no independent measurement of these products' performance exists. Where a vendor publishes a figure about itself we mark it as such and do not adopt it. Where a cell could not be verified, it says so in those words rather than being quietly filled in.

And here is the list of things we refused to score, which is the more important half: accuracy, deflection, containment, resolution rate, autonomy rate, ROI, time saved, hours returned, loads covered per rep, and every variant of those. Not because they do not matter — they are the only things that matter operationally — but because every published figure in this market is a vendor marketing claim about itself, and printing it as neutral fact would make this article worse than useless to the person who has to defend the purchase internally.

Every cell below is either a citation to something you can open yourself or the literal words describing what could not be verified, checked on 23 August 2026. The ordering reflects how much of the buyer's checklist is publicly verifiable, not how well the product works — a distinction we will keep making because it is the honest one. Two of the ten entries are not agents at all, and they are in the table because a broker evaluating this category will be shown them anyway.

There is a column we deliberately left out, because it would have been identical in all ten rows: independent evaluation. For every product above, as of 23 August 2026, the answer is none found. Rather than print a column of ten identical cells we say it here, once, plainly. If a vendor tells you they have been independently evaluated, ask who by, ask for the methodology, and ask whether the evaluator was paid by the vendor.

The second table is the one that usually decides the shortlist — not what a product does, but who it is for and who it is not for.

Notice how much of the useful signal in both tables is negative space. Four of the ten have no publicly verifiable compliance documentation at all; one displays a badge without stating its type; not a single product in the category publishes a price. That last point deserves saying without euphemism: in a market this active, universal price opacity is a market norm, not a coincidence, and it means every one of your comparisons will be built from quotes you cannot benchmark against anything. Plan the procurement accordingly — run at least three quotes on an identical, written workflow specification, and do not let a vendor scope your workflow for you.

The single likeliest error in any 2026 freight-tech ranking is a stale roster, and the reason is structural: several of the best-known names in this market were acquired during 2025 and 2026, and every one of them still trades under its original brand. The brand continuity is not deceptive — retaining a well-known name after an acquisition is ordinary practice — but it means a writer relying on recognition rather than public record will rank a subsidiary as an independent company. Here is the corrected picture, all of it from primary corporate announcements.

Two observations for a buyer. The first is consolidation: DAT Freight and Analytics acquired Trucker Tools in December 2024, Outgo in May 2025 and Convoy in July 2025, so three separately branded products a broker might have shortlisted as alternatives to one another now sit inside a single company. That is not an argument against any of them. It is an argument for asking, at the diligence stage, which of your shortlisted vendors are commercially the same counterparty, because the negotiating leverage you think you have may not exist.

The second is the project44 split, which is the correction most likely to matter to you this quarter. On 14 July 2026 project44 separated into two businesses: project44 for enterprise shippers, and LSP44 for 3PLs, freight forwarders and brokers. Jett McCandless is chief executive of both, and LSP44 launched with its own sales, engineering and leadership. There was no merger with FourKites — a rumour we specifically checked and can report as false; the two remain separate competitors. If someone hands you a 2026 ranking that lists project44 as the broker agent product, you now know how recently it was researched.

Each entry follows the same shape: what it does, what is verifiable, what is not disclosed, who it fits, who it does not, and who owns it. Where a vendor publishes a performance figure about itself we name the figure as a vendor claim and refuse to adopt it, because that is the whole point of the exercise.

One boundary entry that did not make the ten but that you will be shown: Aljex, the broker TMS owned by Descartes Systems Group. It is the useful reference point for a classic TMS with an ecosystem rather than an agent vendor, and it appears as a documented integration target on Vooma's list. We could not verify any Descartes AI-agent claim for Aljex and therefore do not assert one. Similarly, Cargado is a cross-border US–Mexico load board and marketplace, not an agent; adding chat and pricing features to a load board does not move it into this category, and a ranking that includes it as an agent has not drawn the line carefully.

On 14 May 2026 the Supreme Court held, unanimously, that a negligent-hiring or negligent-selection claim against a transportation broker is not preempted by the FAAAA, because it falls within the statute's safety exception. The case is Montgomery v. Caribe Transport II, LLC, No. 24-1238. Certiorari was granted 3 October 2025, argument was heard 4 March 2026, and the decision came down 9–0 with the opinion written by Justice Barrett and a concurrence by Justice Kavanaugh joined by Justice Alito. If you read one paragraph in this article before signing anything, read this one: it is the reason autonomous carrier selection is now a liability question and not merely an efficiency one.

The statutory mechanics are short. 49 U.S.C. 14501(c)(1) provides that a State may not enact or enforce a law, regulation, or other provision having the force and effect of law related to a price, route, or service of any motor carrier — and the provision expressly reaches brokers and freight forwarders, which is why brokers had a preemption argument at all. Subsection (c)(2)(A) is the safety exception: paragraph (1) shall not restrict the safety regulatory authority of a State with respect to motor vehicles. The whole case turns on whether a state negligence claim about which carrier a broker picked is an exercise of that safety authority. The Court held that it is.

Requiring C.H. Robinson to exercise ordinary care in selecting a carrier therefore 'concerns' motor vehicles—most obviously, the trucks that will transport the goods.

— Justice Barrett, writing for a unanimous Court in Montgomery v. Caribe Transport II, LLC, No. 24-1238 (U.S. May 14, 2026), Section II of the opinion

A note on that quotation, because it looks odd at first glance and the oddity has caused confusion elsewhere. The case caption names Caribe Transport II, but the opinion discusses C.H. Robinson throughout its reasoning, because Robinson is the broker-respondent inside the caption's et al. This is the body of Barrett's opinion, not the syllabus — a distinction that matters, because a Supreme Court syllabus carries the standard disclaimer that it forms no part of the opinion and should never be quoted as the holding. We cite Cornell's Legal Information Institute because supremecourt.gov and Justia both refuse automated requests; if you want the slip opinion PDF, pull it in a browser.

Two adjacent developments get overstated in freight-tech content and both need status precision. In July 2026 a Dallas County jury returned a verdict of $604 million in a case arising from a 2021 Interstate 20 crash in Mississippi in which three people were killed and two injured, apportioning 23% to C.H. Robinson. The theories the jury accepted were vicarious liability — it found the driver a borrowed employee operating under the broker's control — and negligence in undertaking the responsibilities of a motor carrier. Notably, the jury did not find negligent carrier selection. Status matters here: that is a verdict, final judgment has not been entered, and the broker has announced an appeal. It is not settled law and should not be cited as if it were. Separately, a Texas Supreme Court decision the following day held that passive shippers owe no duty to the driving public merely for engaging a federally regulated carrier — a shipper-side holding that does not protect brokers, and should not be presented as broker-favourable.

There was a genuine circuit split on the preemption question before Montgomery — the Sixth Circuit ruled against a large brokerage on it in July 2025, and a certiorari petition followed. Reporting on the precise alignment of the other circuits is inconsistent enough that we will not print a circuit-by-circuit map. The accurate summary is that circuits disagreed, the Supreme Court resolved it 9–0, and the earlier decisions are now background rather than live law.

What this means for software, concretely: every autonomous carrier-booking decision your agent makes is a decision a jury may review under state negligence law, and you no longer have a federal argument for getting that review dismissed early. An agent that books a carrier without a recorded vetting step is not saving money. It is manufacturing uninsured liability, and it is doing so at machine speed and machine volume, which is the part that should concentrate the mind.

The regulatory question is never whether the agent may broker. It is whose registered authority the agent is operating under, and whether that registrant exercised the care the law expects. Broker registration licenses a person and demands experience. 49 U.S.C. 13904 requires the Secretary to register a person as a broker of property only on finding that the applicant has sufficient experience to qualify the person to act as a broker for transportation and is fit, willing, and able to be a broker for transportation and to comply with this part and applicable regulations. Each broker must employ an officer with at least three years of relevant experience, or satisfactory evidence of knowledge of the rules and industry practice, and must update its registration within 30 days of a change of address, contact information, officers or process agent.

What counts as brokering is defined in 49 CFR 371.2: a broker is a person who, for compensation, arranges, or offers to arrange, the transportation of property by an authorized motor carrier, and brokerage service is the arranging of transportation or the physical movement of a motor vehicle or of property. Motor carriers and their employees or bona fide agents are excluded when arranging transport they are authorised and legally bound to provide; a bona fide agent is part of the carrier's normal organisation, works under the carrier's direction, and is barred by a pre-existing agreement from independently directing traffic to other carriers. FMCSA issued final regulatory guidance on those definitions on 16 June 2023. That guidance predates commercial LLM agents and does not address software agents at all. It is silent — not permissive, not prohibitive — and anyone who tells you FMCSA has blessed or barred agentic brokering is characterising a document that does not say either thing.

If your agent dials carriers or drivers, assume it is inside the TCPA and build accordingly. In Declaratory Ruling 24-17, adopted unanimously and released on 8 February 2024, the FCC held that AI-generated voices are artificial voices for purposes of the Telephone Consumer Protection Act. The ruling expressly refuses any carve out of technologies that purport to provide the equivalent of a live agent — which is a direct answer to the argument every voice-AI vendor eventually makes. The practical obligations that follow are prior express consent, prior express written consent where the call is marketing, identification and disclosure of the party responsible for the call, and a working opt-out.

Freight has a specific wrinkle worth naming honestly: a great deal of check-call activity is business-to-business, placed to dispatchers and carrier reps, and some of it lands on a driver's personal mobile. We did not verify any business-to-business treatment of carrier dispatch calls in this pass and we are not going to imply one exists, because getting that wrong is precisely the kind of advice an operator could act on and be sued for. Take it to counsel with your actual call patterns in front of you. What we will say architecturally is that consent state should live outside the agent, in a system of record the agent reads and the dialer enforces, with suppression applied at dial time rather than by the model deciding whether to call. That design survives a compromised or confused agent. A consent rule written into a system prompt does not.

One more design note. Identification and disclosure are not a script the model improvises; they are a fixed opening the platform plays or asserts on every call, logged with a timestamp and a recording reference. When a complaint arrives eighteen months later, the artefact you need is a per-call record showing consent state at dial time, the disclosure delivered, and the opt-out honoured. Build that on day one. Retrofitting it into a running voice deployment is unpleasant and expensive.

What follows is an illustrative scenario, not a client. We have invented the brokerage and its volumes to make the arithmetic concrete; every external figure inside it is sourced, and we mark clearly which numbers are assumptions we made up and which come from the record. We are doing it this way deliberately: an article that refuses an unsourceable industry statistic and then asserts an unverifiable client outcome has destroyed its own standing.

The scenario. Consider a 40-person brokerage moving 900 loads a month, mostly dry van truckload, with eight carrier reps and four track-and-trace coordinators. Assume, as an input we are inventing for the illustration, three status touches per load. That is 2,700 outbound contacts a month, or roughly 135 per working day, before exceptions. The brokerage is evaluating an outbound voice agent for check calls and appointment scheduling, and a separate question about letting the agent cover loads autonomously.

The savings arithmetic we refuse to do.We could multiply 2,700 contacts by a per-call minutes-saved figure and hand you a payback period. We will not, because every minutes-saved figure published in this market — five minutes per quote, ninety minutes to seconds, a capacity uplift band, a phone-tree success rate — is a vendor's claim about itself, and building your business case on a supplier's number means the supplier controls both sides of your ROI equation. The baseline you should use is yours and it takes a week to get: pull last month's call detail records, sample the durations, and time the workflow end to end for five working days. Then hold that number back from the vendor until after they have quoted.

The liability arithmetic we can do, because it is sourced. This is where the numbers are real and where they should change your design.

• The bond is not a shock absorber: Federal financial security for a property broker is $75,000, set in the statute itself at 49 U.S.C. 13906(b)(3) and unchanged. ATRI's October 2025 research puts average annual theft losses at more than $1.84 million for logistics service providers. The bond covers roughly four per cent of a single average year's theft exposure for an LSP. It was never designed as loss coverage, and no one should treat it as one.
• One verdict, against the bond: In the July 2026 Dallas County case, a jury apportioned 23% of a $604 million verdict to a broker. Twenty-three per cent of $604 million is $138.9 million — approximately 1,850 times the $75,000 federal financial-security requirement. That verdict is on appeal and final judgment has not been entered, so it is not a number to plan against as settled law. It is a number that tells you the shape of the tail.
• The exposure per bad tender: ATRI's survey work, as recounted in December 2025 congressional testimony, reported an average loss per incident of $29,108 for motor carriers and $95,351 for logistics companies. At 900 loads a month, a single fraudulent tender per quarter at the logistics-company average is roughly $381,000 a year in direct loss, before any liability claim and before the cost of the investigation. That figure is hedged — it comes via a testimony account rather than the report itself — but the order of magnitude is the point.
• What the vetting gate is worth: You cannot compute a fraud-reduction percentage honestly, because the only ones published come from vendors who sell fraud prevention and at least one of them carries the escape hatch 'when process is followed' inside the claim. What you can compute is the cost side: a single-workflow agent build with an enforced vetting gate falls in the $28k–$70k band. Compare that to one average logistics-company theft incident, and the comparison stops being close.

What we would actually build for this brokerage, in order.First, the check-call and appointment workflow, autonomous, with per-call consent state checked outside the model and a fixed disclosure. That is the reversible, high-volume, low-stakes surface and it is where a real baseline gets established. Second, POD chasing and document extraction into the TMS with idempotent writes and scheduled reconciliation, because that workflow pays for itself in dispute avoidance and produces the audit artefacts you want anyway. Third — and only third — carrier sourcing, with the agent permitted to source, quote and prepare a tender, and the tender blocked in code until the carrier-identity record returns a current pass. That third step is where the money is and it is also where Montgomery lives, which is why it goes last and behind a gate.

The build-or-buy question underneath this — whether to configure one of the ten products above or build the workflow yourself over your own TMS and rating stack — turns mostly on how unusual your workflow is and how much of the vetting logic is yours. We work through that decision in detail in our build versus buy analysis for AI agents. For most brokerages under a few thousand loads a month the answer is buy the voice layer and build the gate.

Freight agents fail in six recognisable ways, and five of them are detectable before they become expensive if you instrument for them at build time. The sixth — vendor corporate change — is not detectable by monitoring and is handled contractually.

Prompt injection deserves more than a table row, because freight is a hostile-input environment by default and the exposure here is worse than in most verticals. Your agent reads inbound carrier email, rate confirmations, bills of lading, proofs of delivery and load-board messages. A meaningful share of that content is written by parties actively attempting fraud — double-brokering operations, identity-spoofing rings, and people who make a living from the underground market in operating authority. This is not a hypothetical adversary who might one day try an injection attack. It is an adversary already sending you documents.

Prompt injection is unsolved, and it will not be solved by the model getting better. The honest framing is blast-radius reduction and nothing more. The controls that survive a compromised agent are structural rather than instructional: a vetting gate the agent physically cannot bypass because the block is enforced in the tendering system rather than in a prompt; a rate ceiling enforced in code; a payment hold requiring a second, independent signal; a tool allowlist the agent cannot expand at runtime; per-action audit logging that attributes every action to the agent and the human on whose behalf it acted; and a carrier-identity system of record the agent can read but never write. The control that does not survive is a well-written system prompt, no matter how carefully worded. If a vendor's answer to injection is that their model is careful, or that they detect ninety-something per cent of attacks, you have learned what you need to know — in security, ninety-something per cent is a failing grade against an adversary who only needs to succeed once.

Three tiers, and the bottom one is not negotiable after 14 May 2026. This table is the artefact we would put in front of a broker's counsel and insurer before a single line of agent code is written, because it is the document that says what the machine is allowed to decide.

The shape of the rule is worth stating in one sentence, because it is the sentence that should survive being copied out of this article: an agent may prepare, source, quote and queue a tender; a person with your registered authority must approve the selection of any carrier that has not cleared a recorded, current vetting check. Everything above that line is workflow. Everything below it is the thing a jury reviews.

Two implementation notes that decide whether the boundary is real or decorative. First, the gate must be enforced at the tendering system, not in the agent's reasoning. If the block lives in the model's instructions, it is a preference; if it lives in the code path that issues the tender, it is a control, and only the second one is defensible in a deposition. Second, the vetting record must be immutable from the agent's side. An agent that can add a carrier to the approved list and then tender to it has a vetting gate in name only, and that arrangement is worse than having no gate at all, because it produces a document trail that looks like diligence and is not.

These are Frenchy Digital's bands for building agent workflows, published so you can compare them against whatever a vendor or another agency quotes you. They are our pricing, not an industry benchmark, and you should read them exactly the way you read every other number in a procurement: as one firm's figure, from a named source, on a stated date.

Senior-led delivery is $150–$225 per hour. Ongoing retainers run $2,500–$9,500 per month. Every build carries a 30-day post-launch warranty, and full source-code and IP ownership transfers to you at completion — which matters more in this category than most, because a workflow encoding your carrier-vetting policy is a competitive asset and should not sit inside someone else's configuration console. We return a fixed-price phased proposal within 5 business days of the discovery call. Frenchy Digital is a senior-led, Black-owned agency in Los Angeles; you can reach us on +1 (424) 272-5601 or book directly at calendly.com/frenchydigital/discovery-call.

One budgeting point that gets missed and then hurts. Integration drift is an operating cost, not an incident. Rating APIs change, TMS schemas move, EDI maps get versioned, and carrier-vetting providers adjust their response payloads. Put a line in the annual budget for it rather than treating each occurrence as an emergency, and require per-integration health checks with alerting from day one so that drift shows up as a monitoring event rather than as a month of quietly failed loads.

Ten things that should slow a purchase down, and what to ask instead. None of them means a vendor is bad. Each one means a specific question has not been answered yet.

The most useful question in this entire category is also the simplest, and we would open every vendor call with it: which of these five things are you — an autonomous agent, a mixed system where some workflows act and others draft, a copilot that assists a person, infrastructure that other products consume, or an optimisation engine? A vendor who answers that question crisply and then tells you which of the five they are not is a vendor worth spending an afternoon with. A vendor who treats the question as a framing exercise has told you something about how the rest of the diligence will go.

The honest list. Everything below is a gap in what we could establish, stated so you know exactly where this article's confidence ends. Several of these gaps are the direct result of automated access restrictions rather than absent information, and we say which.

• No .gov page was read directly: fmcsa.dot.gov, bts.gov, supremecourt.gov, transportation.gov, ecfr.gov and gao.gov all returned HTTP 403 to our fetcher. Every agency-sourced claim here comes either through the Federal Register API and govinfo, which we did reach, or through trade press, which we mark as such. The Montgomery holding is corroborated by two independent sources that agree on every field, and the quoted sentence is confirmed as appearing in the body of the opinion rather than the syllabus, but the slip-opinion PDF itself was not opened by us.
• No SEC filing was read, so there is no broker-margin figure in this article: Attempts to reach C.H. Robinson investor materials returned rate-limit and gone responses. We therefore print no average-broker-margin number of any kind. If you need one, take it from a named 10-K or 10-Q for a named issuer and a named period, and note that a category average across truckload, LTL and managed transportation is close to meaningless even when sourced.
• No detention-hours figure appears here: We did not reach a DOT Inspector General or FMCSA detention study, so we print no detention average, including the widely repeated ones. Where detention matters to an agent workflow we describe the mechanism — the agent can log arrival and departure timestamps and generate the accessorial claim — without asserting an industry average.
• The cargo-theft refusal rests on ATRI, not on a trade-association contradiction: Some coverage refuses the inflated figure by asserting that a trade association publishes two irreconcilable totals differing roughly fivefold. We could not document that from the association's own materials and therefore do not assert it. The refusal here rests on ATRI's contrary primary research and on the visible arithmetic behind the larger number, which is stronger ground anyway.
• FleetWorks is ranked with three unverified cells: Its site returned no fetchable body to us, so integrations, compliance posture and pricing are all unverified as of 23 August 2026. We ranked it because trading status is confirmed by a dated third-party partnership announcement, and we marked every cell we could not fill.
• LSP44's compliance posture is unverified: The entity is six weeks old at the time of writing and has no separately fetched trust page. We specifically decline to assert SOC 2 for it, and you should decline to assume it.
• Two bill numbers are deliberately omitted: The anti-double-brokering Senate bill and the foreign-dispatcher House bill both reached us through trade press; congress.gov and govtrack returned 403. Rather than print a bill number we could not verify, we describe what each would do and tell you to look them up. A wrong bill number is worse than no bill number.
• No independent benchmark exists for this category: That is a finding, not a gap in our research. We searched for a third-party benchmark, academic study, analyst evaluation or trade-association test of AI agents in freight brokerage and found none. Every performance figure we encountered was published by a vendor about itself, and we have named them as such throughout rather than quietly omitting them.

None of this argues against buying. It argues for buying in the order the risk actually runs: establish your own baseline before anyone quotes you, put the reversible high-volume workflows in first, build the vetting gate as a system control rather than a model instruction, and keep carrier selection on the human side of the line until you can show a jury the record. The vendors on this list are, on the whole, doing real work — the category is not vapour. What it lacks is any external measurement, and until that exists the discipline has to come from the buyer.

And the sentence to keep, because everything above is downstream of it: an AI agent can negotiate the rate, but it cannot absorb the liability for choosing the carrier. Since 14 May 2026 that liability is no longer shielded by federal preemption, and the only thing standing between an autonomous booking decision and a jury is the quality of the record your system leaves behind.