AI Agents
AI agents for aircraft turnaround: the ground operation an agent can own
A turnaround is roughly 100 tasks run by about 70 organizations against one clock. The agentic-airport pitch says AI will orchestrate all of it. Here's the honest split: the coordination and exception-handling an agent can own, why A-CDM milestones are the real data model, where Salesforce fits, and the safety line an agent never crosses.
An aircraft lands with 45 minutes before it’s due out again. In that window, roughly 100 separate tasks have to happen, run by about 70 different organizations that don’t report to each other.
Fuel, catering, cleaning, water and waste, baggage off and baggage on, boarding, the pushback tug. Every one has to slot into a sequence. And the whole thing races one clock: the time the aircraft has to be off the block, or it loses its departure slot.
That’s the shape of a problem an agent is built for. It’s also where the “autonomous airport” pitch overreaches. An agent isn’t going to marshal the aircraft or sign the load sheet.
What it can own is the part that actually breaks turnarounds: coordination. Keeping every party’s view of the clock honest, spotting the task that’s slipping before it cascades, and acting on the gap while there’s still time to recover. The question isn’t whether AI runs the ramp. It’s which slice of the ground operation an agent takes off a human’s plate without touching the parts that keep people safe.
I’ll draw that line, name the data model that makes it possible, and show where a Salesforce stack fits.
The clock everyone races is already a standard
You can’t orchestrate a turnaround with a language model and good intentions. The industry already runs it on a shared framework: Airport Collaborative Decision Making (A-CDM), a joint effort of ACI EUROPE, EUROCONTROL, IATA, and CANSO. Any agent that touches turnaround has to speak its language, or it’s guessing.
A-CDM links the inbound flight to the outbound one and tracks progress through a fixed sequence of events, 16 milestones in the EUROCONTROL specification. The airport operator, the airline, the ground handler, and air traffic control share one picture instead of four.
Two timestamps carry most of the weight. TOBT, the Target Off-Block Time, is the handler’s and airline’s best current promise of when the aircraft will be ready to leave the stand. TSAT, the Target Start-Up Approval Time, is what ATC gives back in reply, weighing TOBT against flow and capacity restrictions to set the departure sequence. Miss the TOBT and the TSAT moves, and the slot behind it belongs to someone else.
This tells you what the agent reasons over. It doesn’t invent a schedule. It watches milestones against TOBT and catches the divergence early. That’s the high-volume, rule-bound, time-critical monitoring a human ramp coordinator does under load and loses track of when six aircraft turn at once. The data model isn’t a prompt. It’s the milestone stream.
What an agent can own on the ground
Start with the tasks that are coordination, not physical work, because those are the ones an agent can absorb.
The first is milestone reconciliation. Seventy organizations means seventy sources of truth, and the expensive failures live in the gap between them. The caterer thinks they’re done, the system still shows the truck en route, and nobody updates TOBT. An agent that ingests each party’s status and checks it against the milestone sequence flags the inconsistency the moment it appears, instead of a coordinator discovering it at boarding.
The second is exception detection and the first response to it. When fueling runs long or a baggage cart is stuck, the question is always the same. Does this threaten the TOBT, and if so, what’s the smallest intervention that recovers it?
An agent watching the live sequence can spot the pattern before a human does, alert the roles that need to move, and draft the knock-on updates: a revised TOBT to the airline, a heads-up to the gate, a re-sequenced task order for the handler. In an operations center these systems act as a co-pilot, running what-if scenarios and proposing gate swaps, stand re-sequencing, and turnaround prioritization for a human to approve.
The third is the notification and paperwork tail that surrounds every delay. A slipped turnaround throws off a flurry of updates across systems and people, and most of it is mechanical. Log the reason code, notify the downstream stakeholders, update the passenger-facing time, open the record operations and finance reconcile later. That’s the load an agent removes so the coordinator spends attention on the recovery decision, not the data entry.
None of this is hypothetical. Fraport is deploying agentic AI at Frankfurt through its AI@Fraport initiative, including an Intelligent Dispatching Agent aimed at turnaround optimization. Computer-vision vendors like Assaia already watch the ramp to time each turnaround event automatically. The direction of travel is real. The discipline is scoping the agent to the coordination layer and leaving the physical and safety-critical work where it belongs.
Where Salesforce fits, and where it doesn’t
If you’re building this on a Salesforce stack, the mistake is imagining Agentforce replacing the airport operational database or the A-CDM platform. It doesn’t, and it shouldn’t. Those are systems of record for a safety-regulated operation. The agent sits above them as a coordination and engagement layer.
Three pieces do the work.
Data 360 is where the milestone streams, handler statuses, and flight data unify into one profile of the turnaround the agent can reason over. An agent grounded on fragmented data invents answers, and the data foundation decides the agent’s quality far more than any instruction does.
MuleSoft is the integration fabric. It pulls the A-CDM feed, the airport operational database, the handler’s task-tracking, and the airline’s operations systems into that foundation and pushes updates back out, the same agent fabric pattern any multi-system agent needs.
Agentforce is where the agent itself lives, scoped to the coordination actions: read the milestone state, detect the divergence, notify the roles, draft the TOBT revision for a human to confirm.
At scale it’s rarely one agent. A turnaround touches enough distinct domains that a multi-agent design fits better, a supervisor coordinating specialists for baggage, catering, fueling, and passenger impact. Each agent keeps a smaller context and its actions stay auditable, rather than one monolith trying to hold the whole ramp in its head.
The same coordination problem shows up in airline crew scheduling and across supply-chain and logistics operations. The architecture rhymes: unify the data, ground the agent, keep the human on the decisions that carry risk.
The line an agent never crosses
Turnaround is a safety operation, and that draws a hard boundary the pitch decks skip.
An agent doesn’t authorize pushback, sign the load and trim sheet, clear the aircraft for fueling with passengers aboard, or make any call a regulated role holds by certification. Those decisions carry legal accountability that sits with a named human, and no amount of model confidence transfers it. The agent’s job stops at recommending and preparing. A human dispatcher, ramp agent, or load controller acts.
The framing holds across every operational agent worth shipping. The agent handles the high-volume coordination so the human’s judgment goes to the decisions that matter, the same split that makes an airline disruption agent valuable without letting it strand a passenger. An agent that revises a TOBT for a human to approve is doing real work. An agent that unilaterally releases an aircraft is a liability with a chat interface.
The failure modes that decide whether this works
Three things sink a turnaround agent, and all three are foreseeable.
The first is stale or conflicting data. If the milestone feeds lag or disagree and nobody reconciles them, the agent reasons over a picture that’s already wrong and confidently proposes the wrong recovery. This is why the data layer is the project, not the prompt. An agent is only as current as the streams it watches.
The second is acting without the authority to be trusted. If the agent fires alerts nobody asked for, or drafts updates the responsible party never agreed to, the humans route around it within a week. The exemptions and approvals have to be explicit. What the agent may send on its own, what it must stage for a human, and who that human is.
The third is treating the standard as optional. An agent that ignores the A-CDM milestone logic, or pushes a TOBT the platform rejects, adds noise to a system that already has enough. The agent has to be a well-behaved participant in the collaborative process, not a parallel one.
The takeaway
The autonomous airport makes a better headline than a build plan. The build plan is narrower and more useful.
Build an agent that owns the coordination layer of a turnaround. It watches A-CDM milestones against TOBT, catches the slip before it cascades, and takes the notification and paperwork load off the coordinator so their attention goes to the recovery. Ground it on unified operational data, wire it to the systems of record without pretending to replace them, and stop it firmly short of any decision a certified human owns.
That’s the turnaround agent that survives a bad-weather morning. It’s a great deal more valuable than one that claims to run the ramp.
Understanding the basics
What is aircraft turnaround in airport operations?
Turnaround is everything that happens to an aircraft between arrival and its next departure: deplaning, cleaning, catering, fueling, water and waste servicing, baggage unload and load, boarding, and pushback. It involves roughly 100 tasks performed by about 70 different organizations, all sequenced against the aircraft’s target off-block time. Because those organizations don’t share a chain of command, coordination is the part that most often fails, and it’s the part an agent can help with.
Can an AI agent run an aircraft turnaround autonomously?
No, and it shouldn’t try. The physical and safety-critical work, and any decision a certified role holds (authorizing pushback, signing the load sheet, clearing fueling), stays with humans who carry the legal accountability. An agent can own the coordination layer: reconciling milestone data across parties, detecting a task that’s slipping before it threatens the departure slot, and drafting the notifications and revised times for a human to approve.
What is A-CDM and why does it matter for an agent?
Airport Collaborative Decision Making is the industry framework that links inbound and outbound flights and tracks a turnaround through a fixed set of milestones (16 in the EUROCONTROL specification), so the airport, airline, ground handler, and air traffic control share one view. Its key timestamps, TOBT (Target Off-Block Time) and TSAT (Target Start-Up Approval Time), are what an agent reasons over. An agent that ignores A-CDM is guessing at a schedule the operation already runs on.
How would you build a turnaround agent on Salesforce?
Unify the milestone streams, handler statuses, and flight data in Data 360 so the agent reasons over one current picture. Use MuleSoft to integrate the A-CDM feed, airport operational database, and handler and airline systems in and out. Build the agent in Agentforce, scoped to coordination actions with human approval on anything that changes an operational commitment. At scale, a multi-agent design with a supervisor and domain specialists keeps each agent’s context small and its actions auditable.
If you’re weighing where an agent could take load off your ground operation without touching the safety-critical parts, talk to us about scoping an operational agent. The value is in the coordination layer, and getting the data foundation and the human-approval boundaries right is the whole job.
