Insider

🪪 An AI agent can now book the trip and read the card statement. Only 38% of travel companies are confident they'd catch the fraud — would you?

This week an expense platform became one of the first finance connectors inside Claude, and a TMC shipped an assistant that books a round trip in under five minutes by chat. 92% of travel companies see the agentic upside; 16% are very concerned about the fraud. Before you hand an agent your data an

Egor Karpovich

24 Sep 2026 — 7 min read

This week

On 15 September, Expensify announced it had become one of the first expense-management tools available as a connector inside Anthropic's Claude for Small Business — an AI assistant that can now read a company's expense, card and receipt data and answer "which expenses from last month are still missing receipts?" in a sentence. The day before, Gant Travel put an agentic assistant into its self-service platform that searches, books and manages business travel through conversation — applying traveler profiles, company policy and booking rules automatically, and completing an eligible round-trip air booking in under five minutes. Also on 14 September, a new Accertify study of the travel industry landed with the uncomfortable counterpoint: 92% of travel companies see significant or moderate opportunity in agentic commerce, only 16% are very concerned about the fraud risks tied to AI agents, and just 38% are very confident they can manage the emerging threats. The agent layer stopped being a demo this week and became a connector — one that reads your financial data on one side and touches a payment credential on the other. The question is no longer whether your program will have agents in it. It is what yours is allowed to see, spend and approve — and whether you wrote that down before someone else's default did.

Why it matters

1. The opportunity curve and the control curve have separated. In the Accertify data the gap is not subtle: 92% see the upside, 38% are confident they can police it, and 58% name bots and credential stuffing as the top threat — the exact attack surface that looks identical to a legitimate agent hitting your booking flow with stored credentials. Seventy-two percent are expanding login controls and 64% account-change controls, which is the right instinct aimed at the old problem: those defences ask who is signing in, not what is this agent authorised to spend. Meanwhile 66% of travel respondents still describe fraud and security as a cost centre, against 12% who call it a revenue enabler. Budgets follow that framing, and right now it is pointing away from the thing that just changed.

2. You cannot govern an agent on data you haven't consolidated. GBTA's research this spring, across buyers in the US, Canada and Europe, found that only 12% have data consolidated from a single source and 63% lack consolidated reporting — while 83% say they want a conversational booking experience. Those two findings are in direct conflict. An assistant answering "what did we spend in Frankfurt last quarter?" is only as correct as the record underneath it, and an agent booking against policy is only as governed as the policy it can actually read. Point a conversational layer at four disconnected systems and you do not get visibility; you get a confident answer you cannot audit. The sequencing is unforgiving: consolidate the record, then automate on top of it.

3. The regulatory clock is running on the European side. The EU AI Act moved into general application on 2 August 2026, with Article 50 transparency duties for providers of synthetic content phased to 2 December 2026, and the Chapter III obligations for Annex III high-risk systems applying from 2 December 2027. That gives European programs a real window — but a window is not a pass. Whatever your agent does with employee travel data in 2027 will need a defensible record of what it was permitted to do, what it decided, and who could override it. Logging that after the fact is archaeology; capturing it at the point of action is compliance.

What this means for your program

Write the agent's spending mandate before you write the integration. Decide now, in policy language: which cabin, which value threshold, which suppliers, which countries an agent may transact in without a human. Card-network plumbing is already shipping — Visa folded its Intelligent Commerce and Trusted Agent Protocol work into Ramp's corporate card and expense stack across 50,000+ customers in March, and Mastercard's Agent Pay binds a tokenised credential to a specific agent, merchant scope and consent policy. The rails will arrive with defaults. Yours should arrive first.
Give agents their own credential, never a human's. The single worst pattern is an agent operating on a shared corporate card or an employee's personal one, because the audit trail collapses into "someone spent this." Issue a dedicated, scoped card per agent or per workflow, with its own limit and its own freeze, so the blast radius of a bad instruction is a number you chose in advance.
Consolidate the record first — you are in the 88%. If only 12% of buyers have data from a single source, the honest first project is not an agent; it is one governed record where bookings, card transactions and receipts already reconcile. Every agentic use case worth having reads from that record. Build the answer before you buy the question.
Instrument the override, not just the automation. Auto-approval is the point of the exercise, but the value sits in the exceptions: log which requests an agent approved on rule, which it escalated, and which a human reversed. That log is your 2027 assurance file, your fraud tripwire, and the only honest measure of whether the automation is actually working.

The numbers

92% of travel companies see significant or moderate opportunity in agentic commerce, but only 16% are very concerned about fraud risks tied to AI agents and just 38% are very confident in managing emerging fraud threats — with 58% naming bots and credential stuffing as the top threat, 72% expanding login controls, and 66% describing fraud and security teams as cost centres against 12% who see them as revenue enablers, per Accertify's travel-industry fraud-cyber research reported 14 September 2026. (Business Travel Executive, Accertify)
Of 3.878 billion transactions analysed between 1 December 2025 and 15 March 2026, just 7,741 were agentic — roughly 0.0002%, or about one in 500,000. The exposure today is small; the governance you write today is what meets it when it isn't. (Accertify)
Only 12% of corporate travel buyers have data consolidated from a single source and 63% lack consolidated reporting — while 83% want a conversational booking experience and 58% say AI has had little or no impact on their program so far, per GBTA research fielded in March 2026 among buyers in the US, Canada and Europe. (Business Travel Executive)
33% of travel buyers say their organisations are already experimenting with autonomous AI, 49% of suppliers and TMCs say the same, and 51% of buyers plan to use agentic AI for expense reconciliation — with data privacy, security and integration with existing systems named as the top concerns, per GBTA's Business Travel Outlook Poll of 591 respondents fielded 2–15 October 2025. (GBTA)
The payment rails are already agent-aware: Visa's Intelligent Commerce and Trusted Agent Protocol are being deployed across Ramp's corporate card, expense, travel booking and bill-pay platform and its 50,000+ customers (announced 31 March 2026), while Mastercard's Agent Pay issues agentic tokens bound to a specific agent, merchant scope and consent policy. (PR Newswire, PYMNTS)

Two tools built exactly for this moment

An agent is only as safe as the two things you hand it: the record it reads and the credential it spends. Get those two right and agentic travel is an efficiency story. Get them wrong and it is an incident with a very clean audit log pointing at you.

Travel Code Expense Management → — one governed record where bookings, card transactions and receipts already reconcile: OCR receipt capture, automatic policy checks against your own rules, approval routing and Excel export. It is the 12%-problem solved at the source — so when an assistant answers "what did we spend in Frankfurt last quarter?", the number behind the sentence is one your auditor can reconstruct.
Travel Code Net-60 Card → — scoped virtual cards with per-swipe visibility, budget rules that auto-freeze at 80%, and 60-day settlement at 0% interest. Give a workflow — or an agent — its own credential with its own ceiling instead of a shared corporate card, and the blast radius of a bad instruction is a number you set in advance, visible the moment it posts.

The bottom line

This week the agent layer quietly crossed from demo to connector: an expense platform inside a general-purpose assistant, a TMC booking round trips by chat, card networks shipping agent credentials into corporate spend stacks. Almost none of your transactions are agentic today — one in 500,000, by the only large dataset anyone has published. That is precisely why this is the cheap moment to decide the rules: what an agent may see, what it may spend, whose card it holds, and what it must escalate. Programs that write that mandate now will onboard agents as a controlled supplier. Programs that don't will discover their mandate was written for them — by a default setting, in a connector, on a Tuesday.

What's new at Travel Code

Did you know you can issue a card to an AI agent — deliberately, with its own rules? The virtual-card flow now spells out exactly what you're choosing: every card type carries a plain-language name, a short description and chips for its key traits (3DS support, issuing region, Apple Pay / Google Pay compatibility), and admins can hide card types platform-wide or open a specific type to selected organisations only. There is also a dedicated AI Agent Cards tile with ready setup instructions for companies issuing cards to their agents via API — so an agent gets a scoped credential of its own instead of quietly borrowing a human's. (Changelog →)

Sources

Travel Code helps corporate travel and finance teams see, control, and reconcile every dollar of trip spend — in one platform. travel-code.com · Expense Management · Net-60 Card