The Dispatcher Problem Nobody Talks About And the AI Architecture We Built to Solve It

In March 2026, Anthropic published a labor market analysis that ranked transportation as the second-least AI-impacted sector in the economy. I read that and felt two things simultaneously: unsurprised, and genuinely troubled.

Unsurprised, because I've spent fifteen years building technology for this industry. I know exactly why AI hasn't landed here the way it has elsewhere. Troubled, because the volume, velocity, and complexity of last-mile logistics is only accelerating, and the humans running these operations are bearing the full weight of that growth, largely without meaningful AI intervention.

That gap is most visible in dispatching. A typical dispatcher runs a ten-hour shift firefighting chaos the rest of the organisation never sees. The average enterprise spends $20 million annually on dispatching operations (a function that has seen almost no meaningful AI intervention), because solving it requires AI that can actually do the work, not just inform the human doing it.

Most enterprise AI deployments optimise at the edges. Route suggestions. Visibility dashboards. Predictive ETAs. These are useful tools. But they share a common limitation: they make the human more informed without making the human less necessary.

Dispatching is different. The work is high-stakes, high-frequency, and deeply contextual - geocoding failures, capacity mismatches, drivers who don't show up, customers who need to be reached before a white-glove delivery can be booked, PODs that come back non-compliant. You can't make that meaningfully easier with a dashboard. You either automate it or you don't.

The reason most enterprises haven't automated it isn't lack of ambition. It's that solving it properly requires AI that can reason across multiple systems simultaneously, take action without constant human direction, and do so with the reliability that logistics operations actually demand. A wrong decision in a knowledge-work context costs time. In dispatching, it costs SLA commitments, customer trust, and real money. The bar here has always been higher, and it should be.

This year we launched PILOT, an AI dispatcher that performs the full arc of a dispatcher's day, end-to-end, in under 30 minutes of active human time.

Here's what that actually looks like in practice:

• Validate. PILOT scans every order in tomorrow's dispatch plan before the dispatch day begins. It is trained to identify 33 distinct data issues across consignments like geocoding errors, inaccurate service times, overweight orders, special instruction flags, and more, surfacing problems that would otherwise surface at the worst possible time: mid-execution.
• Fix. Flagging issues is old school, PILOT fixes them. Splitting overweight consignments, recalculating service times using historical delivery data by product type and geography, sending address correction links to customers with IVR follow-ups if there's no response.
• Schedule. For white-glove or high-value deliveries requiring customer confirmation, PILOT reaches out automatically, first with an email, then a follow-up call, and books the slot without dispatcher involvement.
• Route. PILOT triggers route optimisation across all valid orders, generating efficient multi-stop routes with drive time and distance estimates. The dispatcher retains full visibility and can intervene at any point.
• Staff. PILOT checks driver availability. When zero confirmed drivers are found, it calls through the roster one by one, secures confirmations, assigns routes autonomously, and logs everything.
• Monitor. During live execution, PILOT tracks deliveries, detects driver no-shows at loadout, evaluates SLA impact, and recommends rescheduling or rescue routing in real time. It audits PODs automatically by flagging location mismatches, missing signatures, invalid barcodes,  and generates end-of-day reports on demand.

By the time a human dispatcher would have finished their first coffee, the plan is clean and the day is running.

PILOT isn't a single model. It's an orchestrated layer of 30+ specialised sub-agents, each one a microservice handling a discrete dispatching function: geocoding, scheduling, routing, driver management, POD compliance, exception handling. These aren't prototypes. They've been built, tested, and refined across 250+ enterprise deployments over fifteen years.

The reason we built it this way matters. A monolithic AI model trained on general operational data doesn't have the precision last-mile logistics requires. The exceptions a dispatcher handles follow patterns, but those patterns vary by industry, geography, product type, and customer segment. An agent architecture lets us bring domain-specific intelligence to each problem, rather than averaging it away.

But the architecture decision that took the longest to get right wasn't the sub-agent design. It was governance.

Enterprise AI at this scale only works if operations leaders trust it enough to actually let it run. PILOT operates with human-in-the-loop approval gates at key decision points. Beneath that: IAM role-based access for agents, reasoning trace logs, tool call monitoring, semantic drift detection, policy enforcement, and a circuit breaker with a kill switch. PILOT is strictly scoped & operates only within its configured hub, city, and company parameters. You cannot prompt it outside its dispatcher function.

This is why customers deploy it at scale rather than keeping it in a sandbox. Deployments with Blue Dart, Maersk, and Tractor Supply are showing 80% reductions in dispatcher hours, 17.5% lower cost per delivery, and three-to-five times the throughput from dispatching operations.

The roadmap I shared at MileZero, FarEye’s annual innovation-led event, reflects a deliberate maturity model. We think about AI in transportation across three levels: assistive, autonomous with human verification, and fully autonomous. PILOT today operates at Level 2. The dispatcher retains visibility and can intervene at any point. That's the right place to be for most enterprises right now.

What we're building toward is expansion into adjacent functions. The Finance agent (handling invoice reconciliation, rate card management, dispute resolution) is in development now, plugging into the same orchestration layer. Later this year, PILOT will autonomously manage the financial settlement layer of a dispatch operation, not just the execution layer.

By 2027, the vision is a Super agent that operates across planning, analytics, and customer queries — the cross-functional intelligence that today requires three or four roles working in coordination. The architecture we've built is explicitly designed to scale there without starting over.

The dispatch and network execution market represents $52–70 billion in annual spend across North America and Europe alone. That spend exists today, budgeted, approved, allocated to running operations largely the same way they've been run for a decade. It's the largest operational cost in logistics that has had, until recently, no credible path to meaningful AI-led reduction.

The organisations moving first into autonomous dispatching aren't technology adventurers. They've done the maths. When cost per delivery drops 17.5% and dispatcher hours fall 80%, the risk calculus changes.

The conversation in last-mile logistics has already moved from when to how fast. The technical foundation is there. The governance model is there. The proof points, at enterprise scale, are there. What happens next is a question of deployment readiness, not capability, and that gap is closing faster than most organisations have planned for.

PILOT is fifteen years of operational knowledge, encoded into an architecture that works the way a dispatcher works — at a scale no dispatcher can match, and a reliability level no manual process can sustain.

If you're wondering what PILOT looks like inside your own operation, let's walk through it together. Book a meeting here.