The freight broker's guide to AI agents in freight

By Ivan Tsybaev, founder and CEO of Ten8. I co-founded Trucker Path and was its founding CEO, and I spent more than a decade in freight before starting Ten8.

An AI agent in freight is software that does brokerage work on its own. It makes and takes carrier calls, ranks carriers, moves data between systems, and runs a load from tender to invoice. It hands off to a person when something falls outside its rules. That is the definition worth using. The marketing version stretches from a chatbot that drafts an email to a service that books freight end to end, which is why the word stopped meaning much on a sales call.

This guide is the map I wish every broker and freight agent had before their next demo. It covers the five agent types and what each one does and does not do. It walks one load through all of them, marks where a human still has to stand on the load, and shows how to tell a working system from an "AI" badge. One figure frames the stakes. By FleetWorks' estimate, cited by FreightWaves, close to 80% of U.S. truckloads are still booked by hand, through phone calls, emails, and text messages, one load at a time. That manual layer is the work AI agents are starting to absorb, and the market is pricing it in. One estimate from Precedence Research puts AI in logistics at about $26.35 billion in 2025, growing near 44% a year through 2034.

Agent, tool, or IVR: get the words straight

Three different things get called "AI" in freight, and only one of them is an agent.

The line that matters: a tool makes a rep faster at a step, an agent runs the step. An IVR reads from a script, an agent holds a real conversation and adapts. Most of what gets sold as an "agent" is a tool with a script behind it, so keep that distinction close when a vendor uses the word.

How this guide is sourced

Most "AI in freight" explainers cite nothing. Every figure below is tied to a named public source: a regulator, a research body, dated trade press, or a company's own statement. Each is flagged when it is one company's result rather than a general benchmark. Safety and market data move quickly, so treat the numbers as of this writing.

The five agent types

Strip the language back and AI agents in freight fall into five categories. They are not interchangeable, and a working brokerage usually runs several at once.

Voice agents

Voice agents make and take phone calls and hold real conversations by voice. They work the back-and-forth of a negotiation, collect information, and run the routine check-in. Modern voice models pause, handle interruptions, and adjust to the other person's tone. This is the visible part of the stack, because it is the part the carrier on the line actually hears.

Does well: outbound carrier outreach on a lane or a load, inbound carrier calls around the clock, check calls in transit, customer status updates, routine quote conversations.

Doesn't do alone: decide which carriers to call, generate the rate confirmation, or coordinate across channels. Those belong to other agents.

Email and messaging agents

Email and messaging agents handle written, asynchronous contact. They draft and send outbound email, read inbound email, hold multi-turn threads, and work in SMS and chat. For most brokerages email is still the main channel for non-urgent carrier and customer contact, document exchange, and slow negotiation.

Does well: quote replies to inbound requests, outreach where a call is not warranted, document follow-up, negotiation with carriers who prefer to put it in writing. C.H. Robinson said in October 2024 that it had automated more than 10,000 email transactions a day this way, from price quote through check-in.

Doesn't do alone: win the urgent, high-touch conversations where a call is faster.

Matching and scoring agents

Matching and scoring agents rank and decide. Give one a load and it ranks carriers by lane fit, equipment, recent performance, and price posture. Give it a carrier and it scores reliability and fraud risk. They run in the background and feed decisions to the agents that act. They do not talk to anyone.

Does well: ranking carriers for a load, estimating a walk-away rate, flagging fraud patterns, scoring a load's complexity before it enters the queue. The math is not trivial. ATRI put the average cost to run a truck at $2.336 per mile in 2025, and empty miles routinely run around one mile in six, so a load that looks fine on the posted rate can lose money once the real cost comes out.

Doesn't do: take an action or talk to a person. It outputs numbers, and other agents move on them.

Check-call and tracking agents

Check-call agents take the most repetitive job on the desk: answering "where's my truck?". They pull GPS from a telematics feed, use geofences to catch the events that matter (near pickup, loaded and rolling, near delivery), and answer with a live ETA. The better mode is proactive: push the update before the customer asks, which cuts inbound call volume.

Does well: status calls on your SOP cadence, portal updates, exception alerts when a truck falls behind.

Doesn't do alone: decide whether a delay needs a customer escalation. That is a scoring and coordination call.

Coordination agents

Coordination agents are the layer that runs the other four. Give one a goal, such as covering a load, answering a quote, or handling an exception, and it decides which agents to call, in what order, and when to pull in a person. It holds state across a multi-day process. Without it you have a voice agent calling with no context, a scoring agent ranking with no one acting, and a workflow moving data toward no goal.

Does well: sequencing calls and emails on one load, holding an exception open across days, routing a clean handoff to a human with full context.

Doesn't do alone: anything a carrier or customer sees directly. It is the part that makes the other four add up.

Quick reference:

What a voice agent actually runs on carrier sales

For a brokerage, carrier sales is where the money is, and a production voice agent covers four jobs. Ask a vendor about each one by name.

• Inbound carrier calls, around the clock. Capacity questions, calls on posted loads, payment-status follow-ups. These get answered at 11 PM and 5 AM Saturday, not only during a shift.
• Outbound outreach in parallel. Dialing several carriers at once against your network and your board posts.
• Multi-turn rate negotiation. Holding a real negotiation inside a floor and ceiling you set per lane: countering, holding, and closing without a rep on the line.
• Check calls and track and trace. Status calls on your SOP cadence, with updates written back so the customer portal stays current.

An agent that can't pick up an inbound call or hold a multi-turn negotiation isn't doing carrier sales. It is a tool with a phone number.

How the agents move one load from tender to invoice

Walk a single load from posted to paid and the layers stack up.

• Load posted (coordination and matching). A customer tenders a Dallas to Memphis dry-van load. The system reads the tender from the inbox and the TMS, parses it into a structured record, and writes it in. Coordination kicks off coverage.
• Ranking and posting (matching and workflow). The matching agent ranks every carrier in your network for this load. In parallel, the load goes to the boards so inbound carriers can find it.
• MC verification (scoring). Before any conversation turns into a booking, the scoring agent checks authority, insurance, and safety record, and flags a recently registered MC or a lane history that doesn't fit.
• Outreach and inbound (voice and email). The voice agent dials the top outbound carriers. The email agent works the long tail. The inbound voice agent takes carriers calling on the board post, all against the same floor and ceiling.
• Negotiation (voice and scoring). A carrier picks up. The voice agent holds the call and asks the scoring agent for the walk-away number in real time. It counters, holds, and lands inside your parameters, or it walks.
• Booking (workflow and coordination). The deal closes in range. The rate confirmation generates in your template, the carrier's W-9 and insurance get pulled and checked, and the rate con goes out for signature. The TMS flips to covered.
• Check calls (voice and workflow). Through transit, check calls run on your cadence. GPS data feeds the portal, which updates on its own.
• Exception (scoring, voice, and coordination). A check call finds the driver running late. The scoring agent decides whether the customer needs an escalation. If so, the email agent sends a revised ETA, and past a threshold it hands the load to a human rep with a summary.
• Closeout (workflow). The POD arrives, gets attached, generates the invoice, and pushes it to accounting. The load closes.

On a clean booking, no person touches the load between tender and POD. On a messy one, a person handles a few minutes of judgment instead of running the whole thing by hand.

What AI agents do not do

Vendors skip this part, so read it closely. An agent stack runs the operational layer. It does not run the business.

• Top-account relationships stay with your people. An agent will not save a strategic customer who is unhappy.
• Hard exceptions, such as a claim, a stuck load, or an angry shipper, go to a human with context. The agent's job is to hand off cleanly, not to bluff.
• Lane strategy and pricing policy are yours. You set the floors and ceilings; the agent works inside them.
• Nuanced compliance calls, like an edge-case cross-border move or a regulation that needs interpreting, need a person.

This matches what the trade press keeps finding. Inbound Logistics, reviewing which broker tasks automate, put quoting, load matching, tracking, data entry, and initial carrier vetting in the high-likelihood column, and complex negotiation, crisis management, and relationship building in the low-likelihood column. A vendor who tells you the AI replaces your brokers is selling you a problem. The honest version: it runs the repetitive coverage work so your people spend their hours where judgment pays.

Where fraud and compliance fit

Carrier vetting is a scoring-agent job, and it is one of the strongest reasons to run this layer. The threat has moved faster than manual vetting. Verisk CargoNet put 2025 cargo theft across the U.S. and Canada at close to $725 million, up about 60% year over year. Strategic theft, the deception-based tactics that hit brokers including double brokering and identity theft, grew roughly 1,475% between early 2022 and the end of 2024, by CargoNet data cited by Travelers, and now makes up about 33% of all cargo theft, up from a small share three years ago. Double brokering alone is estimated to cost $500 million to $700 million in freight a year, and in a FreightWaves and TriumphPay survey, 85% of brokers and carriers said they had been hit by it.

The hard part is that the standard checklist (active authority, valid insurance, clean safety record, matching contact info) now passes carriers that later turn out to be fraudulent. Transport Topics documented "ghost carrier" cases where a real MC number, its phone, and its FMCSA credentials are bought or taken over, so every check clears. In March 2026 the FMCSA issued a bulletin warning carriers not to buy, sell, or lease USDOT or MC numbers.

A scoring agent helps in three ways a human queue cannot match at volume. It checks many more signals at once on every interaction. It re-verifies continuously instead of only at onboarding: on a banking change, a new dispatcher contact, a jump in posted rates. And it watches patterns across loads, which is how one fraudster running several brokerages in sequence gets caught. The decision to confirm fraud, call the customer, or file a bond claim still goes to a person. Ten8 runs this checking inside its brokerage operation and vets carriers through providers like Highway. The fuller version is in our fraud playbook.

What it costs, and how brokers pay for it

Vendors price this a few different ways, and the model changes what you are comparing.

There is no public pricing standard, so the useful move is to reduce every quote to a cost per load and set it against what the work costs you now. The formula is simple: take your carrier-sales rep's fully loaded annual cost, divide by the loads that rep books in a year, and you have the human cost per load an agent layer has to beat. An agent layer earns its fee when the cost per load it removes, times your monthly loads, clears that fee. Then add the margin on loads you now cover that you were dropping after hours. On a Fura Freight deployment, Ten8 reports carrier calls answered rising from 60% to 100% and gross margin per load moving from 8.2% to 14.9%; those are one customer's reported results, not a benchmark.

How to tell a real agent stack from an "AI" badge

Every TMS has an "AI" button now. Four checks separate a working system from a label.

• Ask for a recording of a real production carrier call, not a scripted demo. A voice agent that can only handle a happy path shows it here.
• Ask for the autonomous booking rate: what share of routine loads go from posted to covered with no rep clicking.
• Ask which integrations are live versus roadmap, and how long a connection to your TMS takes. A real platform gives you a number. Common live targets are McLeod, Aljex, Tai, Turvo, and Rose Rocket on the TMS side, DAT and Truckstop on the boards, Highway for vetting, and Twilio, RingCentral, Microsoft Teams, and Slack for comms.
• Pilot one full workflow, not one feature. Run an end-to-end booking on a single lane, including the failure modes. Run two weeks in shadow mode first, where the AI decides and your reps compare, then go live for 30 days and measure coverage, margin, and load count.

Buy the coordination layer, not just one agent. Plenty of vendors sell a single agent: voice only, email only, scoring only. Stitching them together is the hard part, and it is where deployments die.

Build it, buy it, or run under a broker that already operates it

For a freight agent or a small brokerage, there are three ways to get this layer, and the right one depends on how much of the operation you want to own.

• Build it. Assemble your own agents: a voice vendor here, a scoring tool there, your own coordination glue. You keep full control and take on the integration work, the maintenance, and the cost of being the systems integrator.
• Buy a platform. License a stack that ships coordination plus the agents. Faster to stand up, and you still run the brokerage: the authority, the factoring, the credit risk, the customer relationships.
• Run your book under a broker that already operates the stack. You bring your customers; the broker carries the MC authority, the factoring, the credit risk, and the operations, and takes a share of the margin. This is the model Ten8 runs as its Agent Network. Agents keep their customer relationships and their share of the margin, and the AI-run operation is how the cost per load stays low enough to make the split work. It trades independence for not having to build or babysit any of the above.

None of these is free. Building means you own the maintenance. Buying means you own the deployment. Running under a broker means you give up some control over how the operation is run. Pick the trade you can live with.

Where this goes

Within a few years most brokerages will run an agent stack of some kind, and the ones that start now build a data advantage that compounds. The move is not to automate a single feature. Pick one lane, run the full workflow (coordination plus the agents) for 30 days, and measure what changes in coverage, margin, and load count. If you would rather not build or buy it, running your book under a broker that already operates the stack is the third door. See how Ten8 runs the operation at ten8.ai.

FAQ

What is an AI agent in freight?

Software that runs a brokerage task on its own, such as carrier calls, carrier ranking, data entry between systems, or coordinating a load end to end, and hands off to a person on exceptions. That is different from an "AI feature" that only speeds up a rep inside one step.

Do AI agents replace freight brokers and agents?

No. They run the repetitive coverage work: carrier calls, negotiation inside your parameters, check calls, data entry. Relationships, hard exceptions, lane strategy, and pricing stay with people. The rep job moves from working the phone to overseeing the system.

How do AI agents handle carrier fraud beyond a basic MC-number check?

The scoring agent checks many signals on every interaction and re-verifies continuously, on a banking change, a new dispatcher contact, or a jump in rates, not just at onboarding. That catches "ghost carrier" identities that pass the standard checks, where a real MC number and its FMCSA credentials are bought or taken over. The agent flags the pattern, a person makes the fraud call.

What does it cost to run AI agents in a brokerage?

Models vary: per seat, a flat platform fee, per result, or a margin share if you run your book under a broker that operates the stack. There is no public standard, so reduce every quote to a cost per load and set it against what the work costs you today.

Do AI agents keep the transaction records the FMCSA requires?

They can, and you should confirm it. Under 49 CFR 371.3 a broker must keep a record of every transaction for three years. A workflow agent that writes rate cons and settlement data back to your TMS should be capturing that; ask the vendor how the records export for an FMCSA review.

Can AI agents handle LTL and intermodal freight, or only truckload?

The voice and email layers are channel-agnostic, but the load logic differs by mode. LTL adds class, density, and appointment rules; intermodal adds ramp cutoffs, chassis, and per diem. Ask which modes a vendor's scoring, quoting, and documents actually support in production, not just truckload.

What happens to my team and my agents' pay if AI runs the coverage work?

The rep role shifts from working the phone to overseeing the system and owning exceptions and relationships. Commission usually follows the same freight and margin, so the lever is loads covered per person, which rises when routine calls run themselves. Under a broker's agent network, your share is set by the margin split instead.