AI Agents
AI agents for mishandled baggage: the claim an agent can own, and the liability call it cannot
A delayed-bag claim is mostly status lookups, receipt checks, and reassurance, work an agent handles well. But "declare this bag lost" and "settle this liability" are decisions with a treaty behind them. Here is how to build a baggage agent on Salesforce that owns the 90% and knows exactly where the WorldTracer status ends and a human judgment begins.
The bag doesn’t come out on the belt. That’s the moment a passenger’s trip goes sideways and an airline’s cost clock starts ticking, because from here, almost everything that happens is a service interaction the airline is obligated to run well, and most of it is repetitive, rule-bound, and emotionally charged in equal measure. File the report. Look up the status. “Where is my bag.” Reimburse the toothbrush and the change of clothes. Look up the status again. “Where is my bag.” It is, in other words, a near-perfect shape for an AI agent: high volume, low complexity per interaction, a passenger who wants an answer at 11pm, and a set of rules that are rules.
And then, threaded through that repetitive work, are a handful of decisions that are not repetitive at all: the moment a delayed bag becomes a lost bag, the moment a reimbursement crosses from “reasonable essentials” into a liability settlement, the moment a claim starts to smell like fraud. Those carry a treaty behind them, the Montreal Convention, and a dollar exposure that makes them the wrong thing to automate. The entire design problem of a baggage agent is drawing that line cleanly: let it own the 90% that’s lookups and reassurance and small reimbursements, and make it escalate the handful of judgments that decide real money. This is the same discipline we’ve written about for airline disruption and travel insurance claims; baggage is its own version, with its own systems of record.
The system of record isn’t Salesforce: it’s WorldTracer
Before any of the AI, get the data topology right, because a baggage agent that doesn’t understand where the truth lives will confidently tell a passenger the wrong thing. The authoritative record of a mishandled bag is not in your CRM. It’s in the global baggage-tracing system (WorldTracer, the SITA-operated platform most of the industry uses) where the bag is matched, traced, and tracked across carriers and airports. When a passenger reports a missing bag, a Property Irregularity Report (PIR) is created; that PIR, with its file reference, is the case identity everything else hangs off.
Underneath that sits the tracking data mandated by IATA Resolution 753, which requires carriers to track every bag at four points in its journey (acquisition at check-in, loading, transfer, and delivery on arrival) producing a verifiable audit trail of custody. That’s the data that answers “where is my bag”: not a guess, but the last scan point. The industry credits 753 tracking with a large reduction in mishandling since its rollout, and it’s the reason a modern baggage agent can give a real location instead of a shrug.
So the architecture is a bridge, not a database. Salesforce holds the customer: the passenger, their unified profile, their loyalty tier, their contact preferences, the case and its interactions. WorldTracer holds the bag: the PIR, the trace status, the match events. The agent’s job is to join them: take the passenger in front of it, find their PIR, pull live status, and act on the combination. That means the load-bearing integration is a governed, reliable call out to the tracing system, exactly the kind of MuleSoft or HTTP-callout API action that turns an external system into something an Agentforce agent can safely use. Get that action right (authenticated, error-handled, returning a clean status) and the agent has something true to say. Skip it and you’ve built a chatbot that guesses.
What the agent owns: status, essentials, and the 21-day clock
Map the claim to its lifecycle and the agent’s territory becomes obvious. A mishandled-baggage claim moves through stages (delayed, then either recovered or, if it isn’t found, eventually declared lost) with damaged and pilfered as parallel tracks. The agent can own most of the delayed stage end to end.
Filing and enriching the PIR. For a delayed bag, the first job is opening the report accurately: flight, bag tag, description, delivery address, contact details. An agent conducting a structured intake conversation, grounded in the passenger’s booking and profile so it isn’t re-asking for the PNR it already has, produces a cleaner PIR than a stressed passenger filling a web form at a baggage-claim carousel. Clean intake matters downstream: a bad address or a vague description is what turns a two-day delay into a two-week one.
Answering “where is my bag.” This is the single highest-volume interaction, and it’s pure retrieval: pull the current WorldTracer trace status for the PIR and translate it into plain language. “Your bag was scanned onto a connecting flight in Frankfurt and is scheduled to arrive on tonight’s service; we’ll deliver it to your hotel by tomorrow midday.” An agent does this at 3am, in the passenger’s language, as many times as they ask, without a queue, and every one of those interactions is a call the contact center didn’t have to take.
Reimbursing interim essentials. While a bag is delayed, passengers are entitled to reimbursement for reasonable essential expenses: toiletries, a change of clothes, the basics. This is the agent’s most valuable owned action, and also where the guardrail has to be sharp. Within a clear, pre-defined policy (per-day caps, eligible categories, receipt required) the agent can review a submitted receipt, check it against the policy, and approve a reimbursement without a human. Outside that policy (an unusually large claim, an ineligible category, a request that exceeds the daily cap) it escalates. The rule of thumb: the agent approves what the policy already decided; a human decides anything the policy didn’t.
Watching the clock. There’s a hard temporal rule that structures the whole claim: if a delayed bag isn’t recovered within 21 days, it is deemed lost, which flips the claim from “reimburse essentials” to “settle the loss.” The agent should track that clock, keep the passenger proactively informed as it runs, and stage the claim for the transition, but, as we’ll see, it should not be the one to pull the trigger.
Where it stops: the treaty, the lost declaration, and the fraud call
Now the escalations, and these aren’t edge cases to handle later, they’re the core of a safe design. Each one is a decision where the cost of an agent being confidently wrong is measured in real money or legal exposure.
Declaring a bag lost. The transition from delayed to lost is not a status update; it’s the event that triggers a liability payout. It carries the 21-day rule, but also carrier discretion, ongoing trace efforts, and sometimes a bag that’s about to be found. An agent should surface that the threshold is met and the claim is ready. It should never decide, on its own authority, that a passenger’s belongings are gone and the airline now owes for them. That’s a human judgment with a financial consequence, and it belongs in the human-in-the-loop approval gate you design before you need it.
Settling liability under the Montreal Convention. For international carriage, an airline’s liability for mishandled baggage is capped by the Montreal Convention: currently 1,519 SDR per passenger (the Special Drawing Right limit, raised from 1,288 SDR effective December 28, 2024, and worth roughly US$2,000). Domestic itineraries are governed separately by national regulators with their own caps. A final settlement means valuing the loss, applying the right cap for the itinerary, offsetting any interim essential expenses already paid, and often weighing a declared special value. That’s an adjudication, not a lookup, the same reason insurance claims adjudication stays with a human even when the intake is automated. The agent assembles the file; a person signs the settlement.
The filing deadlines that create legal rights. The Convention also sets complaint windows (broadly, 7 days for damaged baggage and 21 days for delayed baggage from the date it’s placed at the passenger’s disposal) and missing them can extinguish a claim. The agent’s role here is protective: recognize the claim type, know the clock, and make sure a passenger’s complaint is captured in time. But interpreting whether a late or ambiguous complaint is still valid is a call with legal weight, and that’s a human’s.
The fraud signal. Baggage claims are a known fraud vector: inflated contents, receipts that don’t add up, the same passenger with an implausible pattern. An agent can and should flag anomalies, but it must not adjudicate fraud. A wrongful fraud accusation is its own kind of expensive, and pattern-matching that looks like fraud is often just an unlucky traveler. Flag, route to investigations, let a human decide, the same posture as AML alert triage, where the agent works the queue but never closes the case as guilt.
Building it on Salesforce: the concrete stack
Here’s how the pieces assemble into something you could ship.
The claim is a Service Cloud case. The PIR maps to a case (or a custom Baggage Claim object) carrying the file reference, claim type, status, the 21-day clock as a field, and the running total of reimbursements paid. Standard case automation handles routing, SLAs, and escalation. Nothing exotic, the exotic part is what the agent does on top.
Data 360 unifies passenger and bag context. The agent needs the passenger’s booking, itinerary, loyalty status, and prior claims in one place so it grounds on a real person, not a stranger. That’s a unified profile job, with data graphs relating the passenger to their bookings, bags, and open claims so the agent can traverse from “this passenger” to “their delayed bag’s current status” without re-interrogating them.
WorldTracer is an API action. The trace lookup, status refresh, and PIR updates are governed actions the agent invokes: built as API-backed actions through MuleSoft or, for a simpler integration, a Flow HTTP callout. Wrap the external call so the agent gets a clean, typed result and a graceful failure path when the tracing system is slow or down, because it will be, and “I can’t reach the tracing system right now, here’s what I’ll do” is a far better answer than a hallucinated location.
Guardrails encode the escalation line. The distinctions above (approve within policy, escalate beyond it; flag fraud, don’t adjudicate it; stage a lost declaration, don’t make it) are exactly what Agentforce guardrails and a well-designed escalation and handoff topic exist to enforce. The reimbursement action should be hard-capped in configuration, not just discouraged in a prompt; the lost-declaration and settlement paths should route to a human by construction. A guardrail you can point to is worth more than an instruction you’re hoping the model follows.
Reimbursement runs as a deterministic action. When the agent approves an in-policy essential-expense reimbursement, the payout itself should be a deterministic action (a Flow or Apex invocable that records the amount against the claim, enforces the running cap, and triggers payment) not something the model narrates. The LLM decides whether the receipt qualifies; the platform decides how the money moves and keeps the audit trail. That division is what makes the whole thing safe to run at volume.
Takeaways
- The bag lives in WorldTracer, the customer lives in Salesforce. The agent’s core job is bridging them: find the passenger’s PIR, pull live IATA 753 trace status, and act on the combination. The load-bearing piece is a reliable API action, not a database.
- Own the delayed stage. Structured PIR intake, “where is my bag” status answers, and in-policy essential-expense reimbursements are high-volume, rule-bound work an agent handles well, and the interactions your contact center most wants off its plate.
- Escalate the money decisions. Declaring a bag lost, settling liability under the Montreal Convention’s 1,519 SDR cap, judging an ambiguous filing deadline, and adjudicating fraud are human calls. The agent stages and flags; a person decides.
- Watch the 21-day clock, don’t own its ending. Track the delayed-to-lost transition and keep the passenger informed, but route the lost declaration itself through a human approval gate.
- Encode the line in guardrails, not prompts. Hard-cap reimbursements, route settlements and fraud to humans by construction, and run payouts as deterministic actions with an audit trail. The escalation design is the product.
A mishandled-bag claim is a good agent use case precisely because it’s mostly not a decision. It’s lookups, reassurance, and small reimbursements a policy already settled. Build the agent to own that, honestly and well, and it clears an enormous amount of anxious, repetitive work while a passenger sleeps. Just draw the line where the treaty starts. The agent that knows the difference between reading a trace status and declaring a loss is the one you can let run.
Understanding the basics
Can an AI agent handle an airline baggage claim end to end?
For a delayed bag, it can handle most of it: filing and enriching the Property Irregularity Report, answering “where is my bag” from live WorldTracer trace status, and reimbursing reasonable essential expenses within a defined policy. What it should not handle end to end is the transition to a lost claim and the liability settlement that follows: those trigger a payout under the Montreal Convention and belong to a human. The right design lets the agent own the repetitive service work and escalate the decisions that cost real money.
What is a PIR and WorldTracer, and why do they matter for the agent?
A Property Irregularity Report (PIR) is the record opened when a bag is mishandled. It’s the claim’s identity, with a file reference everything else attaches to. WorldTracer is the SITA-operated global baggage-tracing system most airlines use to match and track mishandled bags across carriers and airports, underpinned by IATA Resolution 753 tracking at four custody points. They matter because the authoritative status of the bag lives there, not in Salesforce. A baggage agent works by bridging the two: the passenger and claim in the CRM, the bag’s real status in WorldTracer via a governed API action.
How much is an airline liable for lost baggage?
For international carriage under the Montreal Convention, liability for mishandled baggage is currently capped at 1,519 SDR (Special Drawing Rights) per passenger: raised from 1,288 SDR effective December 28, 2024, and worth roughly US$2,000. Domestic itineraries are governed by national regulators with their own limits. A settlement values the loss, applies the correct cap for the itinerary, and offsets any interim essential-expense reimbursements already paid, an adjudication that should be made by a person, not an agent. The agent’s role is to assemble the file and stage it for human sign-off.
Where should a baggage agent escalate to a human?
At every point where a decision has financial or legal weight: declaring a delayed bag officially lost (it triggers a payout), settling liability under the Montreal Convention, judging whether an ambiguous or late complaint still meets the filing deadlines (broadly 7 days for damage, 21 days for delay), and anything that signals fraud. The agent can track the 21-day lost-baggage clock, flag anomalies, and prepare the claim, but the lost declaration, the settlement, and the fraud call are human judgments, enforced by guardrails and a designed handoff rather than left to the model.
Building a service agent for a travel operation where the systems of record live outside Salesforce, and the hard part is drawing the line between what the agent owns and what a human decides? That line is the whole design, and it’s exactly the work we do. Talk to us.
