Brokers guide to AI agents in freight

"AI agent" has become the catch-all term for any software in freight that does work autonomously. That's broad enough to cover everything from a chatbot drafting an email to a fully autonomous booking engine that runs an entire load. For a brokerage owner trying to evaluate vendors, the term has become unhelpful. This guide gives you a working taxonomy. We'll define the major agent types, what each one does, where they fit in a brokerage, and how they connect.

By the end, you'll have a mental map clear enough to ask better questions on your next vendor call.

The five agent types

Strip the marketing language and AI agents in freight fall into five categories. Each does something specific. They're not interchangeable, and most brokerages will end up running several.

1. Voice agents

Voice agents make and receive phone calls. They hold real conversations, with carriers or customers, by voice. They handle the back-and-forth of a negotiation, the gathering of information, the routine check-in.

Modern voice agents use models that pause naturally, handle interruptions, and adapt to the speaker's tone. They're the most visible part of the AI broker stack because they're the part the carrier on the other end of the phone actually experiences.

What a voice agent does well:

• Outbound carrier outreach on a specific load or lane
• Inbound carrier call handling 24/7
• Check calls during transit
• Customer status updates
• Routine quote conversations

What a voice agent doesn't do alone:

• Decide which carriers to call (that's a scoring agent)
• Generate the rate confirmation (that's a workflow integration)
• Handle the multi-channel orchestration (that's the orchestration layer)

2. Email and messaging agents

Email and messaging agents handle asynchronous text-based communication. They draft and send outbound emails, parse inbound emails, hold multi-turn email threads, and operate in messaging channels (SMS, WhatsApp, in-app chat).

For freight brokerages, email is still the primary channel for many carrier and customer interactions, especially for non-urgent communication, document exchange, and asynchronous negotiation.

What an email/messaging agent does well:

• Quote responses to inbound customer requests
• Carrier outreach where voice isn't appropriate or possible
• Document exchange and follow-up
• Multi-turn negotiations with carriers who prefer email
• Standardized internal communications

What an email/messaging agent doesn't do alone:

• Handle the urgent, high-touch conversations where voice wins
• Replace the need for orchestration across channels

3. Scoring and decision agents

Scoring agents make ranking decisions. They take a load and rank carriers. They take a carrier and score reliability. They take a customer and predict churn. They run silently in the background, feeding decisions to the agents that take action.

They don't talk to anyone. They produce numbers and rankings that drive what the voice and email agents do.

What a scoring agent does well:

• Ranking carriers for a specific load by lane fit, equipment match, performance, and price posture
• Estimating walk-away rates for negotiation
• Detecting fraud patterns in carrier behavior
• Predicting which customers are about to churn based on tender patterns
• Scoring the operational complexity of a load before it enters the queue

What a scoring agent doesn't do:

• Take action. It outputs decisions; other agents act on them.
• Communicate with humans directly.

4. Workflow and integration agents

Workflow agents move data between systems. They watch for events (a load posted in the TMS, an email arriving in a shared inbox, an ELD reporting a status change) and trigger actions in other systems (write to the TMS, generate a document, update the customer portal).

These agents are the plumbing that connects everything. They're not the interesting part of the AI story, but they're the part that determines whether the deployment actually works in production.

What a workflow agent does well:

• TMS event subscription and update writing
• Document generation and routing (rate cons, BOLs, invoices)
• ELD data ingestion and translation
• Customer portal updates
• Internal task creation and assignment

What a workflow agent doesn't do:

• Make decisions about what should happen. That's a scoring or orchestration agent's job.
• Talk to anyone. It moves data.

5. Orchestration agents

Orchestration agents are the conductor. They take a goal (cover this load, respond to this quote, handle this exception) and decide which other agents to call in what order. They watch for the right signals to escalate to humans. They keep state across multi-step processes.

The orchestration layer is what separates a real AI broker platform from a collection of tools stitched together. Without orchestration, you have a voice agent that calls without context, a scoring agent that ranks without action, and a workflow agent that moves data with no goal.

What an orchestration agent does well:

• Sequencing calls and emails on a single load
• Holding state across a multi-day exception
• Routing handoffs to humans with context
• Adapting strategy when an initial approach fails
• Coordinating across multiple loads simultaneously

What an orchestration agent doesn't do alone:

• Anything user-facing. It's the layer that makes other agents work together.

How they fit together on a real load

Walk through a single load from posted to delivered. See how the agent layers stack.

Step 1: Load posted (workflow agent)

A customer tenders a Dallas to Memphis dry van load. The workflow agent watches the customer-facing inbox and the TMS API. It sees the tender, parses it into a structured load record, and writes it into the TMS.

Step 2: Carrier ranking and load board posting (scoring and workflow agents)

The orchestration agent kicks off coverage in parallel. The scoring agent ranks every carrier in the brokerage's network for this load: lane fit, equipment match, recent performance, price posture, capacity signal. At the same time, the workflow agent posts the load to the relevant load boards so inbound carriers can find it.

Step 3: MC verification (scoring agent)

Before any carrier conversation results in a booking, the scoring agent runs the MC check — authority, insurance, safety record, and historical signals. New MCs younger than six months and MCs with mismatched lane history get flagged for a closer look.

Step 4: Outreach + inbound handling (voice and email agents in parallel)

The orchestration agent dispatches the voice agent to the top outbound carriers in parallel. Simultaneously, the email agent sends targeted outreach down the long tail. At the same time, the inbound voice agent is fielding inbound calls from carriers who saw the load on the boards — pulling the load, communicating details, and negotiating against the same floor/ceiling.

Step 5: Negotiation (voice agent and scoring agent in tandem)

A carrier picks up. The voice agent holds the conversation. The carrier counters at $1,920 for the full load. The voice agent calls the scoring agent in real-time to compute the walk-away ($1,855). The voice agent counters at $1,750. Two more turns. Lands at $1,800 for the full load.

Step 6: Booking (workflow and orchestration agents)

The orchestration agent recognizes the deal closed within parameters. It triggers the workflow agent to generate the rate confirmation in the brokerage's template, pull the carrier's W-9 and insurance, verify insurance is current, and send the rate con through DocuSign. The TMS updates to "covered."

Step 7: Check calls (voice and workflow agents)

Through transit, the orchestration agent triggers check calls at intervals defined in the SOP. ELD data feeds in via the workflow agent. The customer portal updates automatically.

Step 8: Exception handling (scoring, voice, and orchestration agents)

A check call detects the driver is delayed by 3 hours. The scoring agent assesses whether this delay needs customer escalation (it does). The orchestration agent triggers the email agent to update the customer with a revised ETA. If the delay grows beyond a threshold, it escalates to a human rep with a context summary.

Step 9: Closeout (workflow agent)

POD arrives. The workflow agent parses it, attaches it to the load record, generates the invoice, and pushes it to accounting. The load is closed.

In a typical clean booking, no human touches the load between tender and POD. In a load with exceptions, a human handles 5 to 15 minutes of judgment work, instead of the 45 minutes it would take to run end-to-end manually.

Where each agent type lives in your stack

Many brokerages already have some of these agents in place, even if they don't think of them that way.

• TMS workflow features = workflow agents (early-stage)
• Email automation = email agent (early-stage)
• Rate intelligence platforms = scoring agents
• Click-to-call dialers = pre-voice agents
• Manual rep coordination = human orchestration

The shift to an AI-native stack is replacing each of these with their AI-agent equivalent and adding the orchestration layer that ties them together.

Buying advice

Two patterns we'd recommend.

Buy the orchestration layer, not just the agents

Plenty of vendors sell single agents (a voice agent only, an email agent only, a scoring engine only). Stitching them together is hard. The orchestration layer is what turns isolated capabilities into a working AI broker. Buy from a vendor that ships the orchestration layer plus the agents, not just the agents.

Start with one full workflow, not one feature

Don't pilot "the voice agent." Pilot "an end-to-end booking workflow on one lane." That tests the full agent stack including the failure modes. A pilot of just the voice agent will look great in isolation and then break when you try to wire it into the rest of your operations.

How Ten8 covers each layer

Ten8 ships all five agent types as one platform. Voice agents for outbound and inbound calls. Email agents for asynchronous outreach. Scoring agents for carrier ranking, fraud detection, and walk-away calculation. Workflow agents for TMS, document, and ELD integration. Orchestration agents that hold state across the booking lifecycle and across exceptions.

The five agents share data and learn from each other. The voice agent's outcomes feed the scoring agent's models. The scoring agent's rankings inform the orchestration agent's strategy. The workflow agent's data closes the loop.

That coupling is what you're buying when you buy an AI freight platform, rather than a single AI feature. The intelligence is in how the agents work together, not in any single one.

Where this goes

In five years, every brokerage will run an agent stack of some kind. The brokerages that move now have a structural data advantage that compounds. The brokerages that wait will be on a flat starting line in 2030 against competitors with five years of behavioral data fed back into their models.

If you want to see what an integrated agent stack looks like on your freight, book a demo. Integration with your TMS takes about two weeks, then shadow mode on one lane for two weeks before any production traffic.