AI Agents for Trucking& Fleet Operations in 2026

If you run a fleet, you have been planning against a number for more than a decade. Some version of it: the industry is short tens of thousands of drivers. It shapes what you pay recruiters, which vendors get a meeting, which line items your board tolerates, and — the part this article is about — which problems you decide are worth automating.

The number is contested. More importantly, the organization that built it has stopped defending it in the form everyone quotes. That is not a rhetorical point. It changes the shortlist, because a headcount problem and a quality problem call for entirely different software.

Start with the history, because it is cleaner than the argument around it. The American Trucking Associations has published a driver-shortage estimate since 2005, and its own figures moved considerably over seven years.

ATA driver-shortage estimates by year. The 2021 and 2023 figures are reported rather than drawn from a currently published ATA document; the 2018 and 2022 figures come from ATA material directly.

Two things in that table deserve more attention than they usually get. The first is the 2018 forecast. ATA's 2019 release put the shortage at 60,800 at the end of 2018 and projected it would exceed 100,000 within five years and reach 160,000 by 2028. The last figure ATA published was 78,800 for 2022, and reporting indicates the estimate eased from there rather than climbing toward the forecast.

The second is why it eased. Not because a wave of new drivers arrived. Because the freight recession destroyed demand. A shortage measure that shrinks when the economy contracts is measuring the gap between demand and supply, which is exactly what it says it measures — but it means the number tells you as much about freight volume as about labor. That is a fragile foundation for a decade of hiring strategy.

This is the pivotal fact in the whole debate, and it is barely covered outside the trade press.

At ATA's 2025 Management Conference, chief economist Bob Costello — the same economist who announced the record 80,000 figure in October 2021 — reframed the problem in a way that is difficult to reconcile with a headcount narrative.

What we have in the United States is a quality problem around drivers, much more so than an absolute number… It's the quality of the labor. Drug and alcohol testing. Accidents.

— Bob Costello, ATA chief economist, ATA Management Conference 2025 (as reported in trade coverage)

Read that carefully. It is not a softening of the shortage claim; it is a different claim. A quality problem points at screening, qualification, drug and alcohol program administration, and safety performance. A quantity problem points at recruiting funnels and pay. The interventions barely overlap.

We are labeling this as reported — it comes from concordant trade coverage of a conference session rather than from a published ATA document, and ATA's live Labor and Workforce page still states the 78,800 figure for 2022. Both things are true at once, and the tension between them is itself informative: the association's public web copy has not caught up with its chief economist, and neither has the industry that quotes it.

The counter-argument is older than the pivot and better documented. Stephen Burks of the University of Minnesota Morris and IZA, and Kristen Monaco of the Bureau of Labor Statistics, published “Is the U.S. Labor Market for Truck Drivers Broken?” in the BLS Monthly Labor Review in March 2019, building on IZA Discussion Paper No. 11813 from September 2018. It is the most-cited rebuttal to the shortage narrative and it is almost always quoted in half.

The market as a whole appears to work as well as any other blue-collar labor market, and while the truck driver market tends to be ‘tight,’ there do not appear to be any special constraints preventing entry into (or exit from) the occupation. There is thus no reason to think that driver supply should fail to respond to price signals in the standard way, given sufficient time.

— Burks & Monaco, BLS Monthly Labor Review, March 2019

That is the passage everyone quotes, usually as “economists say there is no driver shortage.” It is not what the paper says.

Here is the nuance most coverage omits. Burks and Monaco do not claim there is no problem. They localize it. The difficulty sits in for-hire long-distance truckload — a segment they size at between one sixth and one fourth of all heavy and tractor-trailer truck drivers— and they say explicitly that the dynamics of that segment are “not visible in the aggregate data, and require a distinct analysis.”

That is a considerably more useful finding than either slogan. The aggregate occupation behaves normally: wages respond, people enter and leave, unemployment tracks other blue-collar work. Inside one segment — the one with the worst quality of life and the thinnest margins — something else is happening that averages hide.

Burks and Monaco's conclusion contains a testable prediction, and almost nobody has bothered to test it. Their sentence was that there is “no reason to think that driver supply should fail to respond to price signals in the standard way, given sufficient time.” Seven years is sufficient time. So the question that settles this argument is not rhetorical at all. It is arithmetic: did the price signal ever arrive?

The Bureau of Labor Statistics publishes average hourly earnings by industry, including NAICS 484121 — general freight trucking, long-distance truckload. That is not a proxy. It is the exact segment Burks and Monaco isolated as the place where a real problem might live. We pulled the series from the BLS public data API alongside CPI-U for the same window.

Computed from BLS Current Employment Statistics and CPI-U via the BLS public data API, June 2026 vintage. Earnings are for production and nonsupervisory employees.

Be precise about what that does and does not establish. It does not prove that no carrier struggles to seat a truck — the 10.5% decline in long-distance truckload employment since December 2022 is real, and the enforcement changes covered later in this article removed drivers from the pool for reasons that have nothing to do with wages. What it rules out is the specific mechanism the shortage narrative asserts: a persistent excess of demand over supply at the offered wage. That condition has a signature, and the signature is rising real pay. Across seven years, in the exact segment where the problem is supposed to live, it never appeared.

Two honest caveats on the series. These are earnings for production and nonsupervisory employees on carrier payrolls, which excludes owner-operators — a large share of for-hire long-distance capacity, whose net earnings track freight rates rather than wage rates. And average hourly earnings blend overtime and shifts in pay mix, so the series is not a clean measure of the rate paid for a given hour of work. Neither caveat moves the result by anywhere near thirty percentage points. This is the closest thing to a clean market test that public data allows.

If the labor market works, the obvious objection is: then why does turnover look like that? The OOIDA Foundation's April 2025 analysis, The Churn: A Brief Look at the Roots of High Driver Turnover in U.S. Trucking, gives the sharpest available answer, and it is the piece most technology buyers have never read.

The argument runs through market structure. Deregulation under the Motor Carrier Act of 1980 opened entry, thousands of carriers came in, and long-haul truckload became close to perfectly competitive. In a market like that, no individual carrier can raise driver pay and pass the cost through to shippers without being undercut by someone who did not. So carriers do not. Citing Burks and Monaco, the OOIDA Foundation concludes that firms “accept high turnover as a cost-minimizing response.”

That sentence reframes the entire conversation. High turnover is not evidence of a broken market. It is what a functioning, highly competitive market produces when the cost of replacing a driver is lower than the cost of retaining one. Two structural mechanisms keep it that way, and both are worth understanding before you scope any fleet software.

• Paid by the mile, regulated by the hour: Compensation is per mile. Duty status is per hour. A driver may be on duty 60 to 70 hours in a week while being paid for only 40 to 50 driving hours. Everything in the gap — detention at a dock, waiting on a load, a delayed gate, a shop visit — is time the driver spends and nobody pays for. Because it is free to the shipper, there is no market pressure to compress it.
• The FLSA Motor Carrier Exemption, 1938: The Fair Labor Standards Act excludes most interstate truck drivers from federal overtime protection. In nearly every other hourly occupation, long hours cost the employer a premium. Here they do not. There is no financial penalty anywhere in the system for consuming a driver's week, which removes the ordinary corrective that would otherwise price detention and waiting time into freight rates.

This is also the single clearest place where software earns money rather than saving time. If unpaid on-duty hours are the mechanism, then the workflow that documents them, files the accessorial claim, and collects on it inside the shipper's window is not administrative housekeeping. It is revenue recovery, and it is the first workflow we scope for most asset carriers.

Turnover numbers get quoted loosely, and the loose versions are usually wrong. Here is what is actually defensible, with its scope attached.

Driver turnover by segment. The 92.7% and 77.6% figures are long-run averages reported by the National Academies citing ATA data — not a current quarter.

The contrast in that table is the finding, and Burks and Monaco sharpen it further than the National Academies figures do. Working from ATA's own quarterly survey data across 1995 to 2017, they record averages of 94.0% for large truckload, 79.2% for small truckload, and 11.7% for less-than-truckload. Hold those three numbers side by side for a moment. Same CDL, same labour pool, same country, same regulatory regime — 94.0% versus 11.7%. That is a job-quality gradient, not a supply constraint. A labour shortage does not respect segment boundaries this cleanly. An operating model does.

Two caveats keep that comparison honest. The first is vintage: those are 1995–2017 averages, not current rates, and the series moves with the freight cycle — its trough was Q1 2010, at 39% for large carriers and 35% for small ones, during the worst freight downturn inside the window. The second is definitional, and it comes from ATA itself: the association's own explainer notes that this measure counts drivers moving between carriers, not drivers leaving the industry. A 94% churn rate in large truckload is largely the same population recirculating through the same set of jobs, which is a very different thing from 94% of those drivers being lost to trucking.

For pay context, the Bureau of Labor Statistics Occupational Employment and Wage Statistics release for May 2025 puts heavy and tractor-trailer truck drivers at a median annual wage of $58,640, a median hourly wage of $28.19, across employment of 2,062,040. That supersedes the $57,440 May 2024 median this article previously carried.

Labor arguments go in circles until someone puts the cost structure on the table. The American Transportation Research Institute's 2026 update to An Analysis of the Operational Costs of Trucking, released July 15, 2026 and covering 2025 data, is the best benchmark the industry has, and it settles the argument more decisively than either side's rhetoric.

ATRI, An Analysis of the Operational Costs of Trucking: 2026 Update, covering 2025 data.

A marginal cost of $2.336 per mileis the highest per-mile figure in the report's history. Stripping fuel out to isolate underlying cost pressure gives $1.854 per mile, up 4.2% — meaning the increase is not a diesel story. Here is where every cent of that mile goes.

ATRI per-mile line items, 2025 data, from the 2026 update.

Read the top of that table twice. Driver wages at $0.818 and driver benefits at $0.210 come to $1.028 of a $2.336 mile — 44% of the marginal cost of running a truck. Labour is not one cost among many here. It is nearly half the mile, and it is the line a carrier has the least ability to compress without losing the driver. Then the margins.

Operating margins by segment, ATRI 2026 update. Carriers also executed their largest reduction in freight capacity since the start of the freight recession in 2022.

Two structural details sit underneath those margins. Carriers cut fleet size by 2.4% in 2025 — and roughly 10% of the trucks that remained sat unseated, generating no revenue while still carrying payments, insurance, and depreciation. An idle truck is the most expensive asset in a sub-1% business. That combination, shrinking fleets plus unseated equipment inside the survivors, is the real shape of the 2026 capacity contraction, and it is a different phenomenon from the wave of carrier failures the trade press keeps predicting.

The 2026 market looks, at a glance, like a recovery. It is not, and reading it correctly is the sharpest analytical call an operator has to make this year.

Freight-market indicators to the week ending 7 August 2026. Spot and contract linehaul rates and capacity ratios are DAT; shipment and expenditure indices are Cass, June 2026 report. The June all-in rate comparison is search-summarized trade reporting and remains labelled as reported.

A spot rate above the contract rate is a genuine signal. It happens when shippers cannot cover freight at contracted capacity and have to go to the open market, and the last comparable inversion was February 2022, at the peak of the post-pandemic crunch. Dry van spot linehaul at $2.28 a mile, up 40.5% year over year, now sits above contract linehaul at $2.25. Truck posts are down 31.3%, and the load-to-truck ratio has gone from 5.45 to 10.38 in twelve months — nearly double the loads chasing each available truck.

Then the pair of numbers that changes everything. Cass's June 2026 report has shipments down 4.1% while expenditures rose 11.2%. Shippers are paying substantially more to move measurably less. Rates rising while volumes fall is not demand returning; it is capacity leaving. Cass says so itself, without hedging.

Volumes are still down because capacity is declining.

— Cass Freight Index Report, June 2026

One timing note, because this trips people up in August: those are the Junefigures. The July Cass report is not yet published. If you see June's shipment and expenditure numbers attributed to July in a vendor deck or a market summary, the deck is recycling data it did not check.

And one correction we are making to our own earlier framing, because it circulates constantly and it is wrong. The capacity leaving the market is not leaving through bankruptcy court. FMCSA's own registration data shows operating-authority revocations in the first quarter of 2026 running at their lowest level since the fourth quarter of 2021, with a net influx of carriers over the period. The widely repeated claim that a thousand carriers are failing every week is fiction. What is contracting is the truck count and the seated-driver count inside surviving carriers — the 2.4% fleet reduction and 10% unseated equipment from the previous section. The squeeze is on margin and on utilisation, not on authority count. If you are modelling who your competitors will be in six months, that distinction is the whole model.

For an operator, the practical consequence is that this tightening is not a demand cycle you can ride. It can reverse with an enforcement change or a rulemaking outcome, on a timeline that has nothing to do with freight fundamentals. Price accordingly, and do not add fixed cost against it.

One regulatory change dominates 2026 operationally, and it is not the one most technology coverage discusses.

English Language Proficiency. The Commercial Vehicle Safety Alliance voted in May 2025 to restore English language proficiency to the North American Standard Out-of-Service Criteria, effective 25 June 2025, following an executive order issued in April 2025. The underlying qualification standard at 49 CFR 391.11(b)(2) was already on the books. What changed is the consequence: a driver who fails the check can now be placed out of service at the roadside rather than simply cited.

Reported ELP enforcement volumes before and after restoration to the out-of-service criteria. These figures are search-summarized — the Federal Register and CVSA pages block automated retrieval, so treat them as reported and confirm against the primary text before you cite them in a board deck.

The jump from 33 out-of-service orders to 19,045 is the operational story of 2026 for a carrier. An out-of-service order does not just cost a citation; it strands a load, a truck, and a driver, and it feeds a safety record. A notice of proposed rulemaking codifying ELP as an out-of-service violation published in August 2026, with a comment period running into the autumn. That, too, is reported rather than verified against the docket text.

Hours of service. The core rules are unchanged: an 11-hour driving limit inside a 14-hour window after 10 consecutive hours off duty; a 30-minute break after 8 cumulative hours of driving; 60 and 70-hour weekly limits with a 34-hour restart; sleeper-berth splits of 8/2 and 7/3. Two pilot programs are active but not permanent — a flexible sleeper-berth pilot testing 6/4 and 5/5 splits, and a split-duty-period pilot allowing a pause of the 14-hour clock of between 30 minutes and 3 hours.

Cargo theft is where AI agents have the clearest defensive case, and also where the statistics are worst. Start with data that has an owner.

Verisk CargoNet full-year 2025 data, reported. The last two rows are figures we are explicitly declining to use.

The pattern is the useful part: fewer, larger, more organized hits. Total events were essentially flat, but confirmed thefts rose 18% and the average loss per theft rose 36%. That is a shift from opportunistic to targeted — from a trailer taken from a lot to a load obtained by deceiving the people who arranged it.

The second quarter of 2026 pushed that pattern to its extreme, and it is the single most important thing to understand about freight crime this year.

Q2 2026 freight fraud and cargo crime reporting. Incident counts fell while losses rose — read the two together or you will read either one wrong.

Incidents fell 26% year over year to 677, and losses more than doubled to $304.6 million. That is an average of $564,009 per incident. If your dashboard tracks incident count, 2026 looks like a year of improvement. It is not. Fewer, far larger thefts is the whole story, and the access vector underneath it is business email compromise — a compromised or convincingly spoofed mailbox operating inside a transaction that everyone involved believes is legitimate.

Double-brokering is the vector most often described as growing fastest, and the two numbers usually attached to it do not survive a source check. A “$700 million to $1 billion” annual loss estimate is routinely credited to the Transportation Intermediaries Association; we could not trace it to any TIA document. A claim that FMCSA received more than 8,000 double-brokering complaints in 2025 appears only in low-quality secondary sources. An earlier version of this article carried both with a reported label. On a second pass we are removing them rather than hedging them: a figure that cannot be traced to its stated author is not a figure, and the label was doing too much work.

What is verifiable is the regulation, and two concrete levers came into force. The broker and freight-forwarder financial-responsibility rule reached its compliance date on 16 January 2026. It holds the $75,000 floor but rebuilds the enforcement mechanics around it: if available financial security falls below the floor and is not restored within seven business days, FMCSA moves to suspend the operating authority. That is the first mechanism in the rule's history capable of removing an under-secured broker quickly rather than through a protracted proceeding. (An earlier version of this article reported a rise to $150,000 effective July 2026. That is not the rule that took effect, and we are correcting it here.) Separately, FMCSA's Motus registration system is replacing the 1994-era licensing and insurance platform, with Phase I live in December 2025 and Phase II in the second quarter of 2026. Both changes operate on identity infrastructure, which is precisely the layer double-brokering exploits.

The defensive posture that follows is unambiguous, and it is where an agent genuinely helps. Fraud in freight is concentrated at onboarding: cloned DOT numbers, spoofed carrier identities, forged insurance certificates, and mailboxes that look exactly like the ones you corresponded with last week. The counter is not a smarter human reading the same packet faster. It is systematic cross-checking of every identity signal against independent sources at the moment of onboarding, with a human adjudicating anything flagged.

Autonomy occupies more space in trucking-technology conversation than its current operational footprint justifies. Here is the state of it, with the economics attached, because the economics are what most coverage leaves out.

Autonomous trucking status, August 2026. Aurora's target and quarterly financials come from its Q2 2026 results; Kodiak, Gatik, Bot Auto, Waabi and Plus details come from company disclosures and filings.

Aurora deserves credit for a genuine engineering milestone. It is the only company running driver-out Class 8 freight on public highways at any scale, and validating the roughly 1,000-mile Fort Worth to Phoenix lane means operating a lane longer than a single driver can legally cover under hours-of-service limits. That is a real capability, not a demo. It is also younger than almost every account of it suggests.

Two more figures get mangled routinely. The roughly 440,000 driverless miles is cumulative since launch, not a quarterly rate — aggregators and vendor decks consistently present it as the latter, which inflates the implied run rate by more than an order of magnitude. And the fleet behind it is about 25 trucks across all generations. Now put the economics beside that: Q2 2026 revenue of about $2 million against a net loss of roughly $270 million. That gap is the story of the sector, and it is not a criticism of the engineering. A capability costing its operator two orders of magnitude more than it earns is not yet a service you can build a lane strategy on.

The distinctions further down the table matter just as much, because they are where near-term availability gets inflated. Kodiak runs 35 driverless trucks and none of them are on public roads — they operate on private lease roads in the Permian Basin, hauling frac sand at under 20 miles per hour. Public-road operation is targeted for “early 2027,” and that qualifier is not optional when you repeat the number. Its Q2 2026 “net income of $12.5 million” is almost certainly non-cash warrant revaluation rather than operating profit; do not let anyone describe it as profitability. Gatik is genuinely driverless, in Class 6 box trucks on middle-mile runs — real, and not a substitute for a tractor-trailer. Bot Auto's April 2026 humanless run was a single 231-mile demonstration. Waabi has no verified driver-out operation anywhere. And Plus is not a public company — its SPAC was terminated on 20 April 2026, despite a great deal of 2026 coverage that still describes it as listed.

And the sector has a failure record worth remembering. TuSimple wound down its North American operations, delisted, and became a different company. Embark went from a roughly $5 billion valuation to acquisition in about sixteen months. Capital and coverage are not the same thing as durable operations.

The takeaway for a carrier: this is not a 2026 capacity plan. Watch it, do not budget against it, and be sceptical of anyone selling adjacent technology on the premise that autonomy is about to relieve a driver constraint.

Now the practical core. Everything above narrows the field considerably, and the shape of the industry narrows it further.

Reported FMCSA registration data puts the population at roughly 580,000 active motor carriers, of which about 91.5% operate ten or fewer trucks, with an even higher share among for-hire carriers and a large majority of for-hire operations running a single truck. We are labeling that as reported — the percentages vary between citations and we could not read the underlying ATA trends report directly — but the direction is not in dispute. This is an industry of very small businesses.

A small carrier cannot outbid anyone on rate. At $2.336 a mile in marginal cost and sub-1% truckload margins, it has no pricing room, no procurement leverage, and no ability to absorb a bad onboarding decision. What it does have is a back office where a disproportionate share of the owner's week disappears. That is where the margin actually improves, and it is where agents earn their cost.

Frenchy Digital's workflow shortlist for carrier and broker AI agents, 2026, with the human-review boundary for each.

Notice what is not on that list. Nothing here dispatches autonomously, prices a contract without a human floor, disciplines a driver, or makes a safety determination. Every entry is administrative work under human review, and every one of them produces an artifact a person signs. That is not caution for its own sake — it is because the failure modes in this industry are commercial and legal rather than merely inconvenient. An agent that binds capacity produces a contract you are party to. An agent that scores driver behavior into discipline produces an employment decision.

One more note on the quality framing from earlier. If ATA's chief economist is right that the binding issue is quality of labor — screening, drug and alcohol program administration, accidents — then safety event triage and qualification-file management are not peripheral back-office chores. They are the operational expression of the actual constraint. That is a direct consequence of taking the reframing seriously instead of reciting the old number.

Everything above is diagnosis. This section is the evidence base, and you should know its shape before you build a business case on it, because it is considerably thinner than the category's marketing implies. Across a deliberate search we found exactly one company converting AI agents into audited financials, one fully documented failure at scale, one precisely measured problem with no measured solution, and one head-to-head efficacy benchmark — which was adjudicated false in a legal proceeding.

The one audited example. C.H. Robinson is, as far as we can establish, the only company in freight turning AI agent deployment into numbers that land in audited financial statements rather than in a case study.

C.H. Robinson Q2 2026 disclosures. Note the caveats below before reusing any of these figures.

Two caveats keep this honest. C.H. Robinson's own productivity figure appears as 45%, 50%, and more than 60% across different 2026 disclosures, with no public reconciliation of the three — use the direction, not the decimal. And a claim circulating widely, that 92% of its managed shipments are automated, appears in no C.H. Robinson source we could find. We do not publish it and you should not repeat it.

The failure, equally well documented. Convoy raised at a $3.8 billion valuation on the premise of algorithmic freight matching, and shut down in October 2023. Flexport bought the technology for $16 million. DAT bought it from Flexport in July 2025. That is roughly 99.6% value destruction, with every step on the public record. The lesson is not that the technology failed — parts of it clearly worked, which is why it was purchased twice and is running inside a load board today. The lesson is that a technology advantage in freight brokerage did not survive contact with a rate cycle, because the business it was bolted to was structurally low-margin. That is the same arithmetic this article applied to carriers, turned around and pointed at the vendor.

Detention: a precisely measured problem with no measured solution. This one deserves its own table, because it is the sharpest example of the evidence asymmetry running through this whole category.

ATRI driver detention research, using 2023 data — the most recent comprehensive measurement available.

The last row deserves a sentence of its own. A driver held four hours at a dock does not get those hours back, so the driver drives faster to recover them — and ATRI could measure it directly in the GPS traces at 14.6% faster after a detention event. Detention is not only a pay problem or a productivity problem. It is a speed problem, which makes it a crash-risk problem, which connects the paid-by-the-mile structure from earlier in this article straight through to a safety outcome.

And the benchmark that was ruled false. In February 2026, a JAMS arbitration panel ordered Motive to pay Samsara $30.3 million. The finding concerned a Motive-commissioned benchmark claiming its system detected unsafe driving behaviour “86% of the time versus 21%” for the competing product. The panel found the claim “literally false” and ordered a corrective statement.

One category with no data at all. AI-assisted driver recruiting is marketed hard and has nothing behind it: no named carriers, no published outcomes, no independent evaluation. The “$8,000 to $12,000 cost per hire” figure that anchors most of these pitches traces only to vendor blogs, and we refuse it. That does not mean the tooling cannot work — application chasing and document collection are ordinary automation problems and appear on our own workflow shortlist. It means the ROI model has to be built entirely from your own hiring costs, because there are no industry numbers to borrow.

Every article about AI in logistics discusses model capability. In practice, model capability has not been the limiting factor on any fleet engagement we have run. Integration has.

Transportation management systems, dispatch platforms, ELD providers, accounting packages, and shop systems are mostly closed or semi-closed. Reading data out of them is usually solvable — an API, a reporting extract, a nightly file, occasionally a screen. Writing back is the hard problem, and it is where projects stall four weeks after the demo went well.

Integration reality by system class for carrier and broker technology stacks, 2026.

The scoping rule we apply to every engagement: the write path gets confirmed in week one, in writing, before anything is designed.Not “the vendor has an API” — the specific endpoint or file format, its rate limits, whether it requires a partner tier or a certification, and who at that vendor will confirm it. If nobody can answer, the phase-one design assumes a reviewed export instead of a live write, and the proposal says so rather than discovering it later.

Every item here comes from an evaluation we have run or a deck a client forwarded us. None are hypothetical.

The Frenchy Digital red-flag list for trucking and fleet AI buyers, 2026.

Ask a fleet AI vendor for two things before the pricing conversation: a named write path into your system of record, and a retrospective run against your own historical data. A vendor that can produce both has built something. A vendor that produces neither has built a demo.

— Frenchy Digital buyer's principle

These are the bands Frenchy Digital uses to scope operations AI work in 2026. They assume integration feasibility and baseline measurement are in scope from the start, because retrofitting either one is what turns a fixed-price project into a time-and-materials argument.

Frenchy Digital cost bands for trucking and fleet AI agent engagements, 2026.

Senior-led delivery runs $150 to $225 per hour, and ongoing retainers run $2,500 to $9,500 per month covering model and dependency upgrades, evaluation expansion, incident response, and a quarterly technical review. Every engagement carries a 30-day post-launch warranty, and you receive a written scope with a fixed-price phased proposal within 5 business days of the discovery call.

One budgeting note specific to this industry. The integration layer is largely a fixed cost paid once and reused. The first agent pays for the connection to your TMS, the document extraction tuned on your real images, and the review queue. The third agent inherits all of it. Carriers that sequence two or three workflows through one integration get considerably better economics than carriers that pilot three disconnected point products in parallel — and at sub-1% margins, that difference is the difference between a project that pays back and one that does not.

A short, unflattering list. If you are building a business case, build it on this rather than on a vendor's projection.

• The write path is the project: Reading from a TMS, ELD, or accounting system is usually straightforward. Writing back is gated by partner programs, certifications, and rate limits that are not visible until you ask. Most fleet AI projects that fail, fail here — not on model quality.
• Document quality in freight is genuinely bad: PODs photographed in a cab at night, faxed BOLs, scanned rate confirmations with handwriting. Extraction accuracy measured on clean PDFs does not survive contact with this. Insist on evaluation against a sample of your own worst documents before you sign.
• Prompt injection is unsolved: Any agent reading external documents or emails is exposed. The mitigation is architectural blast-radius reduction — allowlisted tools, no inferred arguments, explicit human intent for anything with commercial effect — not a better system prompt. Assume the exposure exists and design so it cannot reach a booking or a payment.
• Fraud detection has no shared benchmark: There is no published base rate, no agreed definition of a detection, and no independent evaluation. Every accuracy claim in the category is a vendor claim. Run a retrospective on your own historical onboardings or discount the number entirely.
• Nobody measured the before-state: The most common reason a carrier cannot tell whether an agent paid for itself is that the baseline was never captured. Hours spent on the workflow, claims filed and collected, days to invoice, onboardings rejected — capture these for two weeks before anything is automated, or the ROI conversation becomes an opinion.
• Small-carrier economics are unforgiving: At sub-1% operating margins there is no room for a project that pays back in three years. If a workflow cannot show a countable result inside two quarters, it is the wrong first workflow, however appealing the demo.
• The category's efficacy evidence barely exists: One company converts AI agents into audited financials. The only head-to-head dashcam benchmark ever marketed was ruled literally false in arbitration. No vendor publishes verified detention-recovery outcomes, and AI driver recruiting has no deployment data at all. Build the business case on your own measured baseline, because there is no industry number to borrow.
• The statistics in this industry are unreliable: Current turnover figures are not publicly available. ATA's own two freight-fraud totals differ roughly fivefold. Cargo theft totals in circulation are off by up to fifty times. Market data providers block verification. We have labelled every reported figure in this article as reported, and you should apply the same discipline to anything a vendor hands you.
• Agents do not fix structural problems: Nothing here changes mileage pay, the FLSA Motor Carrier Exemption, or the competitive structure of long-haul truckload. Software can document detention and collect on it. It cannot make detention cost the shipper anything, and it is worth being clear-eyed about that boundary.

None of this argues against building. It argues for building one workflow with a measured baseline, a named write path, and a human review boundary that is written down before the first line of code. The carriers that get value from agents are the ones that instrumented the before-state and picked a workflow where the result is countable in a quarter.