Is AI Carrier Vetting Reliable?

What can AI vetting actually verify?

The reliable part of AI carrier vetting is the part that runs on public, checkable data. FMCSA's QCMobile API exposes a carrier's authority status, insurance on file, safety rating, out-of-service history, and crash counts. A software check against those fields is the same check a careful dispatcher would run by hand on SAFER — the difference is that the software runs it on every inbound reply, at the moment the reply lands, instead of on the two or three carriers the dispatcher had time to look up.

On top of the FMCSA layer sit identity signals: does the email domain match the carrier's FMCSA-registered name, how old is the domain, does the phone number resolve to the company on file, has the signature on an established thread suddenly drifted. None of these is proof of fraud alone; matched at scale across every reply, they separate the inbox into clean rows and rows that need a human look. In Keelway's operating data, a posted load draws roughly 40 carrier replies, and about 3 of them carry double-broker risk signals (Keelway operating data, 2026). Finding those 3 by hand means reading all 40.

That is the honest scope of the word "reliable" here: AI vetting is dependable at retrieving and cross-checking data that exists. Where the data does not exist yet — that is the next section.

What can't AI vetting catch?

Four failure modes are structural, and no vendor — including us — has engineered around them:

• Brand-new fraud patterns. Pattern-matching systems learn from patterns that have already happened. The first wave of a genuinely new scheme — a document format nobody has spoofed before, a fresh social-engineering angle — lands before it is a pattern. Detection catches wave two onward.
• Collusive insiders. If someone inside the brokerage or at a trusted partner is part of the scheme, the data all looks consistent because the insider makes it consistent. Email-signal analysis assumes the fraud is on the other side of the inbox.
• The clean-record carrier that goes bad tomorrow. A carrier with five years of clean authority, valid insurance, and a matching domain can accept your load Friday and re-tender it Saturday. Nothing in any database predicts a first offense. This is precisely the case for keeping a human relationship layer — a dispatcher who knows the carrier's dispatcher.
• FMCSA data lag. An insurance policy cancelled this morning can still show active in federal records; a revoked authority takes time to propagate. An AI check is only as current as the database behind it (more on this below).

A vendor that claims its AI "catches all fraud" is describing a system that cannot exist. The right question is not "does it catch everything" but "which layer of my vetting does it make cheap enough to run on every single reply."

What happens if the AI flags wrong?

Both error directions matter, and they cost different things.

A false positive — a legitimate carrier flagged as risky — costs a delayed booking. The dispatcher reviews the flag, makes a call, clears the carrier, and the load covers a few minutes later. Annoying, recoverable, and visible: the broker sees every flag and can overrule it.

A false negative — a fraudulent carrier scoring clean — is the expensive direction, and it is why carrier vetting should be layered rather than delegated. The AI data check at the inbox is layer one. The dispatcher phone call before tender ("where's your terminal, what do you run, how many drivers today") is layer two. Rate-confirmation language prohibiting re-tendering is layer three. Tracking during the load is layer four. Each layer exists because the others miss things.

This is also why the booking decision itself should stay human. Keelway's model is explicit: the AI reads, verifies, ranks, and flags — the broker books. An AI that auto-books carriers removes the one layer that can catch what the data layer missed.

Is FMCSA data itself reliable?

AI vetting inherits the limits of its sources, and the primary source is FMCSA. Three known limitations are worth stating plainly:

• Self-reporting. Parts of a carrier's FMCSA profile — fleet size, power units, mileage — are reported by the carrier itself and are only as honest as the filer.
• Update lag. Insurance cancellations, authority revocations, and address changes take time to appear in the records. A check that ran clean this morning can be stale by the afternoon — which is an argument for re-checking at the moment of each new quote rather than relying on a packet filed months ago.
• Thin files on new authorities. A freshly granted MC has essentially no safety history, no crash record, and no inspection trail. "No negative data" and "verified good" are different findings, and a new authority quoting aggressively on premium lanes deserves the opposite of relaxed scrutiny — see what is double brokering for why fraud rings prefer fresh MCs.

None of this makes FMCSA data useless — it is the best public record that exists, and a carrier that fails an FMCSA check is a carrier you do not book. It means a passing check is a floor, not a verdict. Software that re-scores a carrier on every new quote narrows the lag window; nothing eliminates it.

Do brokers actually trust it?

Adoption is real, but it is task-narrow — brokers trust AI with the screening volume, not with the booking decision. The volume is the part no human process survives: Highway blocked 495,267 fraudulent carrier email attempts in Q2 2025 alone (Highway Freight Fraud Index, 2025), and TIA's Watchdog program recorded more than 1,600 fraud reports between September 2024 and February 2025 (TIA State of Fraud in the Industry, April 2025). Against that base rate, a brokerage screening carriers by spot-checking a few SAFER pages per day is sampling, not vetting.

The trust pattern we see with working brokers is conditional and specific: they trust the AI's FMCSA pull because they can click through to the record, they trust an extracted rate because it links back to the source email, and they keep the phone call and the final call for themselves. That is the right shape of trust for this technology in 2026. For the full sourced dataset on inbox volume and fraud, see the freight broker email statistics page; for the manual process the AI layer sits under, see how to vet a carrier.

Where does Keelway fit?

Keelway is an AI platform that automates carrier email triage for freight brokers — turning 40+ carrier replies per posted load into a ranked, vetted shortlist in under a second. The vetting layer checks every inbound reply against FMCSA's QCMobile API at the moment of reply — authority, insurance, safety rating, out-of-service history, crash counts — plus email-domain and identity signals, with the Fraud Shield add-on running the full server-side check in under 600 ms on a fresh-cold MC (Keelway product data, 2026). One adjacent number, quoted in its proper context: rate extraction runs at greater than 95% accuracy on numeric quotes (Keelway product data, 2026) — an extraction figure, not a fraud-detection figure.

Keelway's position on the reliability question is the one this page argues: human-in-the-loop, by design. The AI ranks and flags; it does not book. Every flag links to its source — the FMCSA record, the email header, the domain lookup — so a dispatcher can audit any row in seconds and overrule the machine when judgment says otherwise. See what Keelway is for the full product picture.