How to Build a Concierge Agent That Books Flights, Hotels, and Restaurants
Key takeaways
- A concierge agent completes bookings for the user. An assistant only recommends them. The difference is whether the conversation ends in a confirmation or in a list of links.
- Flights, hotels, and restaurant reservations each have their own supply channels, such as airline NDC connections, global distribution systems, hotel aggregators, and reservation platforms.
- A multi-part itinerary can be approved once, if each line item has a merchant, a description, and a price limit, and execution is checked against those terms.
- Charging the user's own card through a tokenized credential keeps the platform out of the flow of funds, which matters more as booking amounts grow.
- Travel prices and availability change between search and purchase, so the agent must stop and ask again when a line item moves outside the approved terms.
To build a concierge agent that books flights, hotels, and restaurants, connect the agent to a booking channel for each category, turn the user's request into an itinerary of specific line items, get one approval for the whole itinerary, charge a tokenized payment credential for each booking, and report each confirmation in the conversation. The agent owns the work of booking. The user approves what will be bought.
This guide describes general practice for agent builders. Companies such as Amadeus, Sabre, OpenTable, and Resy are named only as examples of the market.
What makes an agent a concierge rather than an assistant?
An agent is a concierge rather than an assistant when the conversation ends in a confirmed booking instead of a suggestion. Many travel and lifestyle assistants produce a good itinerary that the user must then book by hand across several websites. A concierge completes those bookings after the user approves them.
This changes what the product must do. A recommendation can be approximate. A booking must have exact dates, names, prices, and cancellation terms, and the agent must handle what happens when one of them changes.
Which concierge patterns are worth building?
Three patterns cover most concierge requests: trip planning that ends in a booked itinerary, restaurant reservations, and small recurring errands. They differ mainly in how many line items one request produces.
Trip planning that ends in a booked itinerary
The user describes a trip in one message, with dates, a budget, and preferences. The agent proposes an itinerary, the user edits it in conversation, and the whole itinerary is booked after one approval. This pattern has the highest value and the most moving parts.
- What the agent must do well: Read intent from vague input, remember preferences such as aisle seats, and propose an itinerary the user does not need to rewrite.
- What the booking layer must do: Book each flight, room, and table under the terms the user approved.
- A common mistake: Showing twelve options. A concierge recommends one plan and offers to change it.
Restaurant reservations
Reservations are frequent, low in cost, and simple to describe. "A table for four at 8 on Friday, somewhere good" is a complete request. Some reservations need a card on file for a deposit or a no-show fee, so the payment step still applies.
Reservations also build trust. A user who has let an agent book several dinners is more likely to let it book a flight.
Errands and recurring purchases
Household staples, a replacement for something that ran out, and lunch for a team are small and repetitive purchases. Users rarely want to open an app for them, so they suit a concierge well. They use the same approval and payment steps as travel, with fewer line items.
How can a concierge agent book flights, hotels, and reservations?
A concierge agent books flights, hotels, and reservations through supplier APIs, aggregators, merchant-side agent protocols, or browser automation, and each category has its own main channels.
| Category | Common channels | What to watch |
|---|---|---|
| Flights | Airline NDC connections, global distribution systems such as Amadeus, Sabre, and Travelport, and flight booking APIs | Fares change between search and purchase. Ticketing usually needs a travel agency accreditation or a partner that holds one |
| Hotels | Hotel chain APIs, wholesale and affiliate programs from online travel agencies, and channel managers | Rates differ by cancellation terms. Some rates are paid now, others at the property |
| Restaurants | Reservation platforms such as OpenTable, Resy, and Tock, and restaurant websites | Slots are limited. Deposits, party-size rules, and no-show fees vary |
Where an agent uses a consumer website, it should identify itself. Booking sites use bot detection, one-time codes, and challenge steps to block unknown automation. Standards such as Web Bot Auth, and card network programs such as Visa Intelligent Commerce and Mastercard Agent Pay, let an agent present itself as known software that acts for a real user.
How does one approval cover a whole itinerary?
One approval covers a whole itinerary when the itinerary is recorded as a list of specific line items, each with a merchant, a description, and a price limit, and the user approves that list once. Separate prompts for each booking recreate the manual work the concierge was meant to remove.
- 1
Assemble the itinerary as line items: Each flight, room, and table gets a merchant, dates, traveler details, and a price limit.
- 2
Show the whole itinerary in one approval: Render it in your own interface, with the total and the cancellation terms of each line.
- 3
Confirm once: The user approves with a passkey or biometric. The approved list becomes the record of what may be booked.
- 4
Book each line under its terms: Check every booking against its approved line item before the charge.
- 5
Report each confirmation: Flight confirmed, hotel confirmed, table confirmed, in the same conversation.
Design the failure case as well. If the hotel price rises above its approved limit, the agent should not book it. It should report the new price and ask. Decide in advance whether the other lines proceed, and tell the user which bookings succeeded.
Who pays, and what does that mean for the platform?
The user should usually pay with their own card, charged directly by each merchant through a tokenized credential, so the platform never collects or holds the user's money. A platform that collects a trip total and then buys the trip itself starts to act as a seller of travel, which can bring licensing, tax, refund, and working capital obligations.
| Model | How it works | Considerations |
|---|---|---|
| User's own card, tokenized | Each merchant charges the user's card through a vault token or network token | The platform does not hold funds. The user keeps the dispute rights of their card |
| Platform-funded virtual card | The platform pays with a single-use virtual card | Suits corporate travel and funded programs. The platform carries the cost and the refunds |
| Platform as merchant of record | The platform charges the user and pays suppliers | Gives control of pricing. Adds seller-of-travel rules, refunds, and float in many jurisdictions |
With the user's own card, the charge is a normal card purchase. If a booking is not delivered, the user can dispute it through their bank, and the platform is not the party that owes the refund.
What should you build, and what should you get from infrastructure?
Agent builders usually build the parts users judge, such as conversation, memory, and interface, and use infrastructure for identity, consent records, payment credentials, and execution at merchants.
| Layer | Usually built by | Why |
|---|---|---|
| Conversation and taste | Agent builder | This is the product. It decides what the agent recommends |
| Memory and preferences | Agent builder | Seat choices, dietary rules, and favorite areas make the concierge useful |
| Approval interface | Agent builder | The approval should look like the rest of the product, without a redirect |
| Agent identity | Infrastructure | Merchants and card networks need to recognize the agent |
| Consent record | Infrastructure | The approved terms must be stored and checked during execution |
| Payment credential | Infrastructure | Card data must stay out of the agent and out of the builder's systems |
| Execution at the merchant | Infrastructure | Each supplier has its own checkout, rules, and failure modes |
Why is travel harder than food delivery?
Travel is harder than food delivery because checkouts are longer, prices change more often, and each booking has more required data and more rules. Fares move between search and purchase. Hotel rates differ by cancellation terms. Flights need exact traveler names, dates of birth, and sometimes passport data. Loyalty numbers and seat choices add more fields.
Amounts are also larger, so mistakes cost more. A wrong dinner order costs a meal. A wrong fare can cost a nonrefundable ticket. This is why the approval must show cancellation terms and why execution must stay within the approved limits.
How should you sequence the build?
Sequence the build from small, frequent bookings to large, multi-part ones, so users and your system build trust before the amounts grow.
- 1
Start with small purchases: Food, errands, or retail produce real transactions and real user behavior quickly.
- 2
Get the approval right: Make it fast, specific, and native to your product. Users remember this step.
- 3
Add reservations: They are frequent and low in cost, and they build the habit of delegation.
- 4
Add single flights and hotels: Handle price changes, traveler data, and cancellation terms for one booking at a time.
- 5
Add multi-part itineraries: One approval for a whole trip is the most valuable feature, and it depends on the steps above.
- 6
Test the failure cases: Price changes, sold-out rooms, and exceeded limits decide whether users trust the agent again.
Frequently asked questions
Can an AI agent book flights and hotels?
Yes, an AI agent can book flights and hotels through airline NDC connections, global distribution systems, hotel and travel agency APIs, or browser automation. Flight ticketing usually needs a travel agency accreditation or a partner that holds one, and the user should approve each booking before the agent pays.
How does a user approve a multi-part trip without approving each charge separately?
A user approves a multi-part trip once when the itinerary is recorded as a list of line items, each with a merchant, a description, and a price limit. The user approves the whole list with one passkey or biometric confirmation, and each booking is checked against its own line.
What should happen if a travel price changes after the user approves?
If a travel price rises above the approved limit, the agent should not book it. It should report the new price to the user and ask for a new approval, and it should tell the user which other bookings in the itinerary have already been confirmed.
Does a company become a seller of travel if its agent books trips?
A company can become a seller of travel if it collects the trip total from the user and pays suppliers itself. If each supplier charges the user's own card directly through a tokenized credential, the platform does not collect or hold funds. Rules differ by jurisdiction, so get legal advice for your model.
Which channels can an agent use to book restaurant reservations?
An agent can book restaurant reservations through reservation platforms such as OpenTable, Resy, and Tock, where access is available, or through a restaurant's own website. Some reservations need a card for a deposit or a no-show fee, so the user should approve those terms too.
Why not integrate directly with every airline and hotel?
Direct integration with every airline and hotel takes a long time, because each supplier has its own API, commercial terms, data requirements, and failure modes. Most agent builders use aggregators, distribution systems, or shared agent infrastructure to reach many suppliers through fewer connections.
Related articles