CARE Ai answers airline customer service calls on your contact-centre line with a realtime AI voice agent that looks up approved guidance before answering policy questions, reads flight status from a connected reservation or operations system, and connects callers who ask for a person to a representative with a guest brief. The server, not the AI, runs decision guides and the handoff rules, and administrators test it with scripted calls over the real telephone network.
Email me the brochure, CARE Ai
CARE Ai is FlightAtom's AI voice agent for airline customer service teams. It answers the support line with a realtime AI voice, looks up the airline's approved, customer-visible guidance and decision guides before answering policy questions, and reads flight status from a connected reservation or operations system. When a caller asks for a person, it hands them to a representative with a guest brief, the transcript, and the sources used. Production use begins with operator activation.
A realtime voice model holds the conversation and looks up approved guidance. The server runs decision guides, supplies flight status as fixed sentences, and applies the handoff rules: a person when the caller asks, an offer when the assistant cannot help.
Current product interface with simulated calls, synthetic records and neutral presentation branding. Open a screenshot to inspect the workflow.
A resolved call. The voice assistant gave flight status in the operations lookup's fixed sentences and the bag allowance from approved guidance. The call timeline and the cited passages it used are shown. Synthetic demonstration data.
Before picking up, a representative sees the handoff reason and a brief built by fixed rules from the caller's words and tone of voice: a time-pressed flag, the topic, why the caller was handed over, a mood estimate with a tone line and a quote from the caller, and the caller's last words. Synthetic demonstration data.
Consented callback requests with the confirmed number, reason, preferred time and recorded consent. A representative claims each one and closes it with an outcome. Synthetic demonstration data.
An administrator tries a guest question against approved, customer-visible guidance in hybrid search. The cited passage shows its source, version and section, and the saved knowledge checks are below it. Synthetic demonstration data.
A release publishes a replacement policy and withdraws the one it supersedes together. It lists each exact revision with its audience, owner, scope and dates. The author cannot publish the release; another administrator must. Synthetic demonstration data.
Missed answers are grouped into gaps that reopen if the failure returns. Resolving a gap needs an approved document revision and a saved regression question. Synthetic demonstration data.
Each scripted test call is graded from its call record on first audio, reply latency, knowledge searches, flight-status lookups and handoff. Here a three-line custom test passes and a flight-status test warns on reply latency. A question that approved knowledge cannot answer fails, with a note to add a document. Synthetic demonstration data.
Every call keeps its transcript, the guidance it looked up, why it was handed over, and the representative's notes in one record.
Before answering a policy question, the assistant searches the airline's approved, customer-visible guidance and covers the pause with the holding line "Let me check that for you." Callers speak with a realtime AI voice they can interrupt at any point, and administrators choose its voice, greeting, pronunciations, and speaking style. The default greeting introduces a virtual assistant, and the assistant is instructed to say it is an AI, and offer a person, whenever a caller asks whether it is human.
A caller reaches a representative by asking in their own words, pressing 0, or accepting an offer. When the assistant cannot help, for example after a withheld answer, a lookup limit, a question about the caller's own booking or a decision guide that ends with a person, it offers a representative rather than transferring on its own. Words of distress or urgency and changes of topic are recorded for the representative's brief, and an emergency cue is answered once with the advice to hang up and call the local emergency number. A dropped voice-model connection is reopened with the conversation so far, a call that cannot be restored still goes to a person, and every handoff records its reason.
Representatives take calls in a browser workspace with a built-in phone. Before picking up, they see a brief built by fixed rules: the topic, why the caller was handed over, possible-emergency and time-pressure flags, a mood estimate from the caller's words and tone of voice, with the reasons shown, verification state, what the assistant tried, and the caller's last lines. The caller stays on the same call, and the assistant leaves once the representative is connected.
Customer policies need a publisher other than their author, related changes of up to 50 revisions are released or withdrawn together by a second administrator, and two different approved policies on the same topic and scope cannot both be published. Up to 100 saved test questions rerun as a retrieval benchmark that reports passes, failures, and search time. Policies import from PDF, Word, text, CSV, and JSON files, with English OCR for scanned pages.
Administrators place scripted test calls to the support line over the real telephone network, and each call is graded from its record for the greeting, response time (a pass needs a median reply within two seconds and none slower than four seconds), the handoff reason, and word recognition. Scripts include a request for a person, a caller who asks whether it is human, a spelled booking reference, a frustrated caller, and baggage questions in British, Indian, and Australian English voices, and administrators can add up to three lines of their own. A suite can run every night.
Missing or conflicting knowledge and withheld answers become a gap queue, grouped by contact reason. A gap closes only when it is linked to an approved document revision and a saved test question that expects that document, and it reopens if the failure returns. Reports show contact reasons, daily trends, why conversations were handed over, and human-reviewed answer accuracy with its sample size, without reading transcripts or phone numbers.
A read-only connection to the airline's Radixx reservation system or myAirOps lets the assistant describe up to three matching flights in sentences the server writes: whether each is on time, delayed, departed, landed or cancelled, with its scheduled and expected or recorded times. When a status cannot be confirmed, the caller is offered a representative, and passenger and crew fields are never read.
Set staffed hours, closed dates, an active-call ceiling, and an overflow number. A controlled pilot limits AI answering to approved hours and a daily call allowance, sends other calls to the staffed fallback, and cannot start until a named owner records a reference for each acceptance check, from telephony and knowledge to privacy, load, and fallback. Callback requests are saved only with the caller's explicit permission and become owned work items with recorded outcomes.
Staff sign in with their Microsoft work account through single sign-on, and Viewer, Editor, and Admin roles decide what each person can see and change. Voice diagnostics keep codes and timings with masked phone numbers, never what was said, and a recorded call cannot start until a recording notice is configured, which then plays before the caller reaches the assistant. Backups and recording archives are encrypted, and deletion waits for an administrator to preview and confirm the retention policy, with legal holds for calls, documents, and backups.
CARE Ai uses a realtime voice model for the conversation and keeps the controls in the server: which guidance applies, which step of a procedure comes next, how many lookups a turn may take, and the rules that hand a caller to a person. People control what is published.
Search combines keyword ranking with a locally hosted semantic model, so documents are searched without an external inference service. Results are filtered by audience, effective dates, route, and fare, and each passage is cited by document, version, section, and page. If two approved policies on the same topic and scope ever overlap with different text, search returns no passage and cites both for review.
Decision guides are evaluated by the server step by step: the assistant asks the question the server returns instead of choosing a branch, and a guide that ends in a handoff queues the caller. Each caller turn allows at most three knowledge lookups with a five-second deadline each, and a caller who has only said hello cannot trigger a knowledge search or a decision guide.
For deployments that put verification ahead of speed, a strict mode holds each spoken answer about baggage, fees, fares, or flight status until every sentence matches an approved passage or a fixed safe sentence word for word. A reply that does not match is withheld and the caller goes to a representative.
Agree a representative sample and a baseline from your current process. These are proposed evaluation measures, not published performance results.
Approved airline guidance, a policy exception, and a caller asking for a person.
Trace the source used for the answer, review the representative handoff, and follow a consented callback to its recorded outcome.
Answer support against approved sources; handoff completeness; queue wait; callback follow-up time.
Yes. The default greeting introduces a virtual assistant, and the assistant is instructed never to claim to be human and to tell a caller who asks whether they are talking to a person that it is an AI, then offer a person. Asking that question alone does not transfer the call; asking for a person does.
It is instructed to ask for missing details or hand the caller to a representative rather than guess, and the handoff records that approved knowledge was missing. The gap then stays in the knowledge-gap queue until it is linked to an approved document revision and a saved test question.
In CARE Ai, administrators place scripted test calls to the support line over the real telephone network, including a request for a person, a spelled booking reference, a frustrated caller, and baggage questions in British, Indian, and Australian English voices. Each call is graded from its record for the greeting, response time, handoff reason, and word recognition, and a suite can run every night.
A guest brief built by fixed rules: the topic, the handoff reason, any urgency flags, a mood estimate from the caller's words and tone of voice, with the reasons shown, the verification state, what the assistant tried, and the caller's last words, alongside the transcript, notes, and the knowledge used in the conversation. The caller stays on the same call while the representative takes over.
No. The voice assistant does not look up or change bookings, issue refunds, or take payments, and a caller who asks about their own booking is offered a representative. Where the airline connects its Radixx reservation system, a representative can look a booking up read-only after the caller confirms the booking's phone number with a text-message code. Changes, refunds and payments stay with authorized airline staff and their reservation systems.
No. A callback request is saved only with the caller's explicit permission and the number to call, and becomes a work item with an owner, notes, and a recorded outcome. A representative arranges the call through the airline's approved process.
Yes. A universal app for iOS 17 and later gives staff their conversations, notes, guest briefs, consented callbacks, approved-evidence search, and team presence, and gives administrators an audit view and a live view of calls and representatives. It is in pre-release testing and is not yet in the App Store; calls are answered in the browser workspace.
No. CARE Ai has its own staff workspace, call workflows, knowledge governance, and administration. AURORA Ai is not required. An optional XENIA Ai connection adds concierge handoff chat and concierge administration.
As your fleet, routes, rules or teams change, we adapt CARE Ai to match, and you approve each change before it goes live.
A focused walkthrough, with the people and decisions that matter to your team.