There's a moment every dispatcher knows. A van has been sitting still near a customer site for twenty minutes. Is the driver unloading? Queuing at a gate? Broken down? On the phone to the depot? The telematics pin says nothing, and finding out means a phone call — someone at the desk stops what they're doing, dials, waits, asks four questions, and types the answer into a system.
Now imagine a fleet of two hundred vans having that moment simultaneously, on the afternoon a motorway closure scrambled half the day. That's the real bottleneck in driver communication: not that information doesn't exist, but that extracting it from the field costs a phone call each time, and the desk can only make so many calls.
This page is about what changes when AI voice agents make those calls instead: how the calls work, what happens to the answers, and what it means for the people running the operation.
The call itself
A Finmile voice agent is a channel an AI worker uses, not a separate product. The worker behind it already has a job context: which vehicle, which stops, which promises are at risk. When a trigger fires — a vehicle stationary longer than it should be, a route deviation, a scheduled checkpoint reaching its time — the worker places an outbound call to the driver's phone, through the vehicle's hands-free system or the driver's headset.
The driver answers and talks. No app to open, no screen to tap, no form to fill at the roadside. The agent asks what it needs to know for the specific case: what's the holdup, how long, anything blocking the next stop? It listens, asks a follow-up when the answer is thin, and reads critical details back — "forty minutes at the loading dock, yes?" — so a misheard number doesn't become a system record.
When the driver hangs up, the call is already working. The transcript, the extracted facts, and the time of the call are attached to the case, and the worker proceeds to whatever the answer implies.
From answer to action
This is where a voice agent differs from a phone call. The point isn't the conversation; it's what the conversation sets in motion.
Say the driver reports the dock queue at the RDC. Within the same worker loop, the system now knows the delay, its cause, and its likely duration. The worker can re-sequence the remaining stops so achievable ones aren't held hostage to a blocked one, pass the affected jobs to a van with room if that protects a promise, reset the ETA and notify the customers whose windows moved, and update the transport record so the evening report reflects the day as it happened.
The dispatcher sees the adjustment, not just the alert. And because the voice agent is one worker among the operation's others, the same conversation can trigger a customer call, a claim note, or an escalation — whatever your rules say that situation requires.
One constraint matters here: a voice agent acts under the same permissions as the worker it serves. It can say what it's authorised to say and do what it's authorised to do, no more. If a situation needs a human — an angry customer, a safety concern, a judgement call — the call hands off with the full history attached.
Scale, spikes and the desk
Voice agents make calls in parallel. A dispatch desk can't. During a regional disruption, when every controller is already on a line, routine status calls are the first thing sacrificed — which is exactly when the operation most needs to know what's happening across the fleet.
Handing the routine layer to agents changes the desk's job rather than shrinking it headcount-first. The calls that reach people are the ones that need judgement. The routine harvest of "where are you, how long, anything blocking" happens continuously, in the background, at whatever volume the day produces.
It also reaches people the desk struggles to reach. Third-party contractors who won't install your app still have a phone. Drivers in noisy cabs can talk where they couldn't type. And every call is logged — transcript, facts extracted, actions taken — so when a customer asks three weeks later why a window slipped, the answer is a record, not a memory.
Under the hood, briefly
Technically, the voice agent is a channel plugged into the worker's loop. Understand, investigate, decide, act, follow through. The call is one way to investigate; the extraction feeds the decision; the customer message or route change is the action; the confirmed outcome closes the case.
Voice is one channel among several, and workers don't prefer it reflexively. Email, WhatsApp and SMS suit some exchanges better; phone suits time-sensitive, eyes-free, two-way ones. The same worker uses whichever the situation and your rules call for.
Deploying it on top of what you run
Voice-capable workers deploy under Finmile Autonomy, across your existing dispatch, telematics and transport systems — connected by API and webhook where available, otherwise through approved databases, files and controlled interfaces. There's no rip-and-replace programme, and drivers need nothing installed: the agent calls the number already on the manifest.
The starting point follows One AI worker. One workflow. Live in 30 days. A common first workflow is the stationary-vehicle check-in: the agent calls on defined triggers, captures cause and duration, and updates the board. Another is arrival confirmation for time-critical drops. Once the pattern proves itself in the audit trail, the same worker's remit grows.
Operations that want a consolidated environment eventually move to Finmile OS, the operating system for the physical operation with Autonomy built in. But the voice layer starts working long before any such decision, on the systems the fleet already uses.
Frequently Asked Questions
Can it call agency and third-party drivers, not just our employed fleet? Yes. It calls whatever number is on the job. Contractors don't need your app, an account, or any setup — just a phone.
What happens when the driver doesn't pick up? The worker follows your rules: retry after a defined interval, try another channel such as SMS, or escalate to the desk. An unanswered call never silently disappears; it stays an open case until something resolves it.
How does it handle accents, noise and bad signal? The agent confirms critical values by repeating them back before committing them to the record, and marks low-confidence exchanges rather than guessing. Ambiguity becomes a follow-up question, not a wrong entry.
Can we hear what it said? Yes. Calls are logged with transcript and extracted facts, attached to the case. Supervisors review them whenever a dispute or question arises.
