- 01
Market
Saudi Arabia
- 02
Written for
COO, Head of CX
- 03
Decision
Deploy AI assistance for passengers that answers from live airport data and hands over to staff cleanly
- 04
Languages
EN · AR
The architecture
The path from a passenger’s question to an answer backed by live records
The conversation is the surface; this pipeline is the product. Nothing operational reaches the passenger unless a live system said it first, and the branch to a person is designed with the same care as the answers.
A flow: understand the question in Arabic or English, fetch the live operational record, compose the answer only from what was fetched, then follow the flight — with a fallback branch where the conversation and its records hand over to a staff member.
- 01
Understand, in either language
The assistant identifies what is being asked — flight, gate, bag, transfer — and which flight it concerns, following the passenger between Arabic and English without asking them to choose.
- 02
Fetch the live record
The current answer is read from the operational systems at that moment — the flight’s row in the AODB, the belt from the baggage system, the connection time from the schedule — never from the model’s training.
- 03
Compose only from what was fetched
The reply is built strictly from the retrieved records, carries how fresh they are, and phrases uncertainty as uncertainty — “boarding not yet announced” rather than a plausible invention.
- 04
Stay with the journey
If the passenger accepts, the assistant follows the flight and speaks up on change — a gate move, a belt assignment — so one question becomes an accompanied journey rather than a transaction.
The question that belongs to a person
A missed connection, a lost child, a medical need, or simply a passenger who wants a human: the assistant hands over immediately, passing the transcript and the records it fetched, so the staff member starts informed and the passenger never repeats the story.
The workload
Flight, gate, bag, transfer: the four questions that fill a terminal’s day
Listen at any information desk and the pattern is unmistakable: is my flight on time, which gate and how do I get there, where is my bag, will I make my connection. The long tail — prayer rooms, pharmacies, lost property, visa transit rules — is real but thin. This concentration is what makes assistance automatable at all: four question families, each answerable from a system the airport already runs. It is also what defines the engineering bar, because every one of the four changes by the minute. An assistant that answers them from anything but the live feed is not a lower-quality version of the product; it is a different and dangerous product.
This is why we describe grounding as the whole product. The conversational layer — the greeting, the tone, the phrasing — is genuinely commodity now; any capable model produces it. What separates a deployable assistant from a liability is the unglamorous layer underneath: the interfaces into the AODB and the baggage system, the freshness rules that decide when a cached answer is too old to give, the refusal logic that prefers “let me check with a colleague” over a fabricated certainty, and the audit trail that records which system said what, when, for every answer given. Buying the conversation without building that layer is buying the paint without the car.
The language reality
Gulf passengers switch language mid-sentence, and the assistant must follow
Real conversations in a Gulf terminal do not stay in one language. A passenger opens in Arabic, drops in an English flight code, asks a follow-up in English and closes in Arabic; a family splits the conversation between generations and languages. An assistant that forces a language choice at the start, or loses the thread when the language flips, fails the exact population it was bought to serve. The requirement is code-switching as a first-class behaviour: the assistant follows whichever language each message arrives in, keeps one continuous understanding of the conversation across the switches, holds entity names stable — the flight code, the gate letter, the city — in both scripts, and replies in the language the passenger last used unless asked otherwise. This is a testable behaviour, and we test it with transcripts of how people in this region actually speak, not with translated English test sets.
The surfaces
Kiosk, phone or messaging: where the assistant should meet the passenger
One grounded brain, several bodies. The pipeline is built once; the choice per surface is about the moment each one serves, and the same passenger will cross between them in a single journey.
The kiosk in the hall
For the passenger with no app, a foreign SIM or a dead battery. It carries the walk-up moment — voice and touch, both languages on one screen, tall enough for a wheelchair user and a standing family alike, and positioned where confusion actually happens: arrivals, transfer junctions, reclaim.
The passenger’s own phone
The accompanied-journey surface. Inside the airport app, the assistant knows the marked flight, can speak up on a gate change, and hands its answers to the map for the walk. It serves the prepared traveller — and only the prepared traveller, which is why it cannot be the only surface.
The messaging channel
The lowest-friction surface for this region: the passenger asks from the platform already on their phone, before leaving home or from the taxi, with no install. It suits the before-and-after of the terminal visit — is the flight on time, which terminal, where do arrivals wait.
The staff tablet, assist mode
The same grounded pipeline turned to face the employee: an information-desk agent or roving host queries the live systems in plain language instead of juggling terminals. It is often the wisest first deployment — the assistant’s answers are checked by professionals before passengers ever see them.
Our role
What the assistant answers and what stays with your floor staff
- 01
The grounding layer
We build the interfaces into your operational systems, the freshness rules and the refusal logic — the layer that decides whether the assistant is trustworthy, engineered and tested before any conversation goes live.
- 02
The bilingual conversation
We design and evaluate the assistant against real regional speech — code-switching, dialect, flight codes in either script — with Arabic quality reviewed by Arabic speakers, not by a translation pass.
- 03
The handover machinery
We engineer the escalation paths so the moments that belong to people — distress, disruption, judgement — reach people fast, carrying the context, and we tune what escalates with your floor teams rather than for them.
What the assistant is allowed to say, when it must stand aside, and how your teams greet the conversations it passes them are operational decisions that stay on your floor; we build the pipeline, the languages and the handover so those decisions are enforced by the system rather than hoped for.
Common questions
What buyers ask
How do you stop an airport AI assistant from inventing a gate number?+
By making invention structurally impossible for operational facts, not by asking the model nicely. The assistant is architected so that flight, gate, bag and connection answers can only be composed from records fetched from the live systems at the moment of the question; if the lookup fails or the data is older than the freshness rule allows, the assistant says it cannot confirm and offers the desk — it has no fallback to the model’s memory for those question types. Behind that sits an evaluation suite that replays thousands of question variants against known system states and fails the build if a single answer contains an operational fact the fetched records do not support. The residual risk is a wrong record in the source system, which is an operational data problem the assistant inherits honestly rather than a hallucination.
Can one assistant handle Arabic and English switching within a single conversation?+
Yes, and for a Gulf terminal it is the correct requirement to write into the contract, because it is how the population actually speaks. Modern language models follow code-switching well at the comprehension level; the engineering work is everything around that — keeping one conversation state across the switches, holding flight codes and gate identifiers stable between scripts, choosing the reply language from the passenger’s last message, and rendering Arabic that reads as written Arabic rather than translated English. The proof is in the test set: insist on evaluation transcripts that mix languages mid-sentence the way real passengers do. A vendor who demonstrates only single-language conversations has not demonstrated the product you need.
Will an AI assistant replace the airport information desk?+
No, and planning as if it will designs the deployment badly. The assistant absorbs the repetitive, data-shaped questions — status, gate, belt, connection time — which frees the desk for the work that was always human: the distressed passenger, the missed connection needing rebooking judgement, the situation with no clean answer. The desk also remains the assistant’s escape hatch, and the quality of that handover — transcript passed, records attached, no story repeated — is what makes the whole arrangement feel like service rather than deflection. The staffing question worth studying is not whether the desk disappears but how its work changes shape; treat any promise of a specific headcount reduction with suspicion, ours included if we ever made one.
In this cluster
Next step
Ground it on your feeds
A grounding review that lists the live feeds an assistant would need at your airport and demonstrates one question answered from real data end to end.

