Now, Perk lets ChatGPT and Claude execute travel workflows

Perk is moving bookings, expenses and events into AI assistants, while keeping permissions, approvals and audit trails underneath.

Perk just made the corporate travel dashboard optional

Perk lets ChatGPT and Claude execute corporate travel workflows. Now comes the awkward part: deciding who controls the policy, permissions and inevitable 2 a.m. flight change.

My flight is delayed at JFK, and the espresso tastes like burnt litigation. I need a new flight, another hotel night and a clean expense trail before this becomes finance’s problem and, shortly after, my problem again.

The news that Perk lets ChatGPT and Claude execute corporate travel workflows sounds like standard AI-integration confetti. Then I realize I may never need to open Perk. I can tell Claude what happened and let it perform the approved actions through Perk’s systems.

Today, I’d open several tabs, hunt for a confirmation number, compare fare rules and send someone a “quick question” that is neither quick nor really a question. We have built an entire economy around employees copying six-character codes between badly behaved web apps.

Finding another flight is easy. Google Flights has handled that since dinosaurs roamed Terminal B. Safely changing a corporate reservation, triggering approval, extending the hotel and attaching the receipts to an expense is where Perk earns its money.

As a founder, I have to admit something mildly painful: tech companies spent years polishing dashboards that employees would happily avoid. Perk may now be helping make its own interface optional.

Smart move.

the dashboard loses its visa

Travel search and travel execution live in different tax brackets.

A consumer chatbot can suggest three hotels in Rome and remind me that Trastevere is “charming.” Grazie. A corporate travel agent needs my company’s nightly hotel cap, my Miles & More number, the cabin policy and the name of whoever approves the extra €300 when the preferred flight is sold out.

According to PhocusWire’s report on Perk’s expanded MCP capabilities, employees can manage travel, submit expenses and organize events through ChatGPT and Claude. Those jobs usually occupy separate menus and workflows, plus at least one password reset performed while muttering words HR would classify as “concerning.”

The breadth matters. Perk has opened actions across the trip lifecycle instead of gluing a friendly chat box above a list of flights.

I’ve watched plenty of conversational product demos where the assistant answers beautifully, hands me a recommendation and leaves every consequential click to me. It’s a charming intern saying, “Here’s what I found,” before vanishing for lunch.

Perk’s setup goes further. I speak to an assistant already open on my screen; Perk handles the managed-travel machinery behind the conversation.

Others are moving in the same direction. A Business Travel News listing for Amex GBT’s Egencia connector describes flight and hotel search, booking and trip management inside Claude. Egencia keeps its marketplace and policy controls underneath.

Another BTN product summary says FCM is preparing MCP support for ChatGPT, Claude and Gemini. The announced system draws on traveler profiles, trip history, loyalty data and company policy, while FCM’s infrastructure applies the corporate guardrails.

The layers are already taking shape. ChatGPT, Claude or Gemini gets the conversation. Perk, FCM or Egencia gets the transaction.

I welcome this because I have zero appetite for another “all-in-one” work app. I already have Slack, Linear, Google Workspace, GitHub, Figma, several banking portals and enough browser tabs to qualify as a cry for help.

If I spend half my day in ChatGPT or Claude, asking the same assistant to change Tuesday’s flight feels obvious. Opening a dedicated travel portal and manually rebuilding the context feels like using a fax machine with nicer CSS.

Corporate software founders have historically treated interface ownership like beachfront property. We wanted employees to log in, see the logo and admire the sidebar we debated for three months.

Employees never admired the sidebar. They wanted Tuesday’s flight fixed.

permissions become the product

MCP stands for Model Context Protocol. I think of it as connective tissue that lets an AI assistant use defined tools in an external system, rather than improvising from whatever the model happens to remember.

The protocol gets the conference slides. Authority gets the purchase order.

Under the Perk MCP integration described by PhocusWire, requests still pass through Perk authentication and user permissions. If Claude requests an out-of-policy itinerary, the booking enters the company’s existing approval workflow. The chatbot cannot sweet-talk its way past finance.

That distinction becomes obvious the second procurement enters the room. Everyone loves the fluid demo until someone asks who can authorize a $4,000 itinerary from Los Angeles to Zürich.

Suddenly, permissions are molto sexy.

I learned this building connected products at Ad Astrum. With ALYT, our US home-automation system, I lived through hub hardware, the mobile app, cloud services and app-store politics. A beautiful control screen meant very little if the wrong user could unlock the wrong device.

I still keep one sentence close when I work on systems that can act:

Hardware, software, and cloud products fail at the seams between those three.

The lesson travels surprisingly well from smart homes to business travel. ChatGPT interprets the request. Perk verifies the user and decides which actions that person may trigger.

FCM’s announced architecture follows that logic. Its conversational system reportedly combines a traveler’s profile with trip history, loyalty information and corporate policy. FCM-controlled infrastructure applies the rules.

Egencia’s Claude connection also keeps each transaction attached to Egencia’s managed marketplace. Claude gets the conversational surface; Egencia retains inventory access, personalization and corporate travel policy compliance.

I expect the model layer to become interchangeable. A company may standardize on Claude this year, move to ChatGPT after a new enterprise agreement, then test Gemini six months later when legal has its annual identity crisis.

Replacing the layer containing negotiated rates and approval history is harder. That system also holds traveler identity, supplier access, duty-of-care information and the transaction record. Nobody migrates those on a Friday afternoon for fun.

A separate PhocusWire analysis of online booking tools identified four requirements for AI-driven adoption: transparency, security, reliable data and human oversight. I’d put reliable data first. An AI that confidently changes the wrong reservation has still changed the wrong reservation.

I run my own Docker stack on Linux, including this Ghost site, ERP, mail, analytics and automation. Self-hosting has taught me that connecting two systems is the cute opening act. The actual job is deciding what the connection can read, change or delete.

“Connected” looks great in a launch post. “Narrowly authorized and fully auditable” survives security review.

A sleek interface displaying ChatGPT and Claude managing travel workflows, featuring icons for booking, itineraries, and destinations.

Alt text: How Perk lets ChatGPT and Claude execute corporate travel workflows through MCP, company permissions and approval rules.

booking covers eight percent of the job

Travel companies remain weirdly obsessed with search demos. I can find a nonstop flight from JFK to Heathrow in seconds. That problem has been sufficiently solved for years.

The administrative swamp begins after selection.

Perk’s announced integration covers booking travel, submitting expenses and organizing events. An assistant can follow the work across departmental borders instead of producing an itinerary and waving goodbye.

Say I’m arranging a 20-person offsite in Lisbon. “Find me a hotel” covers maybe 8% of the job.

I need a room block. Three people arrive a day early, two leave late, and somebody needs accessibility information. One colleague responds after the deadline with, “Can I bring my partner?” Finance wants every approval tied to the right cost center. The restaurant would appreciate the dietary restrictions before all 20 people are physically seated and asking whether the sauce contains dairy.

This is where AI corporate travel booking becomes corporate travel automation. The assistant has to survive contact with colleagues.

Oversee offers a useful example. The announced AgentSee conversational feature reportedly handles bookings, exchanges, cancellations and fare-rule checks. Human consultants can review or modify AI-generated work in plain language. Eligible tasks can run from start to finish.

Exchanges expose fake automation fast. An assistant can recommend the 6:30 p.m. flight with enormous confidence. The useful part is knowing whether my existing ticket can be exchanged, how much the penalty costs and whether touching the outbound segment will invalidate the return.

I don’t care if Claude can write a sonnet about my aisle seat. I care whether it strands me in Frankfurt.

BCD Travel’s investment in Amgine points toward the same unglamorous work. According to the product scope summarized by Business Travel News, existing deployments included automating group-booking workflows and parsing thousands of emails so agents could spend more time helping travelers.

Thousands of emails.

That sounds less cinematic than an AI planning the perfect weekend in Kyoto, but this is where the money lives. Travel operations run on schedule changes, free-text requests, supplier replies and waiver rules. Somewhere, an agent is manually extracting the one sentence that changes a booking.

I built a connected espresso machine for Pascucci, the Italian coffee brand. Brew telemetry reaching the cloud was technically cool. The difficult work sat between physical behavior and software assumptions: what the machine did, what the data claimed and which action should happen next.

Travel has the same ugly seam. An itinerary exists in a database, a disruption happens in the physical world, and an employee sends a vague message: “I need to get home tonight.”

Expenses make Perk’s scope much stronger. If the assistant connects the original booking to a later change, then carries the receipt into the expense submission, it closes the gaps where employees lose hours and finance teams lose their will to live.

Better recommendations produce marginal gains. Deleting the clerical sludge around an average recommendation saves the afternoon.

canceled flights eat happy-path demos

Corporate travel operates in a hostile environment. Flights disappear. Hotels oversell. Fare restrictions emerge at the worst moment. Executives discover urgent family obligations five minutes after boarding closes.

AI adoption is rising anyway. An American Express research summary published by BTN said traveler use of AI had nearly doubled year over year.

The research also found interest among business travelers and decision-makers in letting AI manage the process from research through expense reporting. Respondents still wanted the ability to review, approve and change recommendations.

Same.

I want the machine doing more work, with a visible checkpoint before it spends $6,800 or sends me to London, Ontario.

Silicon Valley has a bad habit of treating human involvement as product failure. In corporate travel, escalation is often the correct result. Reality has edge cases; Ryanair appears to manufacture them recreationally.

Kayak for Business reportedly lets conversational tools build itineraries, answer questions and modify trips while observing company policy and traveler preferences. Its announced support model includes escalation to human agents when the system cannot resolve an issue.

Oversee draws a similar boundary. Consultants can inspect and modify generated work. Only eligible tasks proceed automatically from beginning to end.

Werner Vogels, Amazon’s chief technology officer, has used a line for years that every travel-tech product manager should tattoo somewhere discreet:

Everything fails, all the time.

I’ve spent 20 years operating Ad Astrum in the North American market, and polished AI demos still get me. That’s my honest concession. Even after shipping IoT systems for Megafon and E.ON, I occasionally watch an agent complete one immaculate happy-path workflow and think, “Okay, maybe this one solved it.”

Then production arrives wearing steel-toed boots.

I’d judge these systems on workflow completion and out-of-policy detection. I want approval latency measured separately, along with change success, cancellation success and escalation frequency. Every action the agent took should appear in a clean audit record.

Fluent prose belongs near the bottom of the scorecard.

Liability gets ugly quickly. Claude may display the option while Perk executes it. If the reservation is wrong, does Anthropic own the interpretation? Does Perk own the confirmed transaction? Perhaps the employer’s policy authorized it, or the employee’s “looks good” counted as final approval.

Procurement will want a named owner. “The agent got confused” will enjoy a short and unsuccessful career as an incident report.

The winning system will stop before the expensive mistake, ask for approval and route the messy exception to a person. It will also preserve enough evidence to reverse the action without reconstructing a 2 a.m. conversation from screenshots.

Claude gets the habit

Travel platforms are making an uncomfortable trade. Opening workflows to external assistants makes the service easier to use. Their logos, interfaces and upsells fade from view.

Perk is opening execution through ChatGPT and Claude. FCM has described support across ChatGPT, Claude and Gemini. Egencia has connected travel management to Claude, while Kayak for Business has announced connections with companies’ own AI assistants.

The stack is easy to picture. ChatGPT or Claude receives my request during the workday. Perk or Egencia applies policy and handles the transaction. A human agent takes the exception that refuses to fit inside a neat workflow.

I may never know which system touched each step. Frankly, I won’t care unless something breaks.

Invisible infrastructure can capture enormous value. Visa disappears when I buy coffee. AWS disappears when I open half the internet. Logistics networks remain hidden until my package spends four days in Secaucus.

A travel platform with years of approval history, negotiated corporate content and traveler profiles has serious weight beneath the chat box. Add duty-of-care data and support operations, and replacement becomes risky.

The danger starts earlier in the funnel. If I ask ChatGPT to plan the trip, it shapes my preferences before Perk receives a request.

Once several travel platforms expose similar MCP actions, the assistant may compare execution partners and route the transaction elsewhere. Commodity flight and hotel inventory offers thin protection.

Perk and its peers need policy logic that works under pressure. Corporate rates matter, but dependable servicing and expense connections matter more once the flight gets canceled. The audit history keeps finance and legal calm after everyone else has gone to bed.

I have argued about this architecture over dinner more often than any healthy adult should. My Italian side wants to discuss carbonara. My founder side drags the table into API economics before the secondo arrives.

Perk can give away the interface and keep a valuable business. Payments companies and cloud providers have already proved the power of an invisible execution layer.

There is a catch. If ChatGPT owns the daily habit, OpenAI owns the doorway. Perk has to remain the safest destination behind it.

By 2029, corporate travel sales pitches will focus on how much of a trip the system can safely execute, how rarely it escalates and whether every action survives an audit. The prettiest flight-results page will rank somewhere below the office coffee.

That prediction has a date. Feel free to bother me in three years.

the confirm button goes underground

Within a few years, opening a standalone corporate booking tool will feel like opening a bank website to pay for coffee. The option will remain. Occasionally I’ll need it. Each visit will feel weirdly manual.

The consequential product screen will sit behind the conversation. It will decide what the assistant can do, when a manager must approve the request and which human takes over when reality becomes messier than the prompt.

Perk is betting that employees can live in ChatGPT or Claude while Perk controls identity, policy, booking execution and expenses. I’d make the same bet.

The first major failure will still be ugly. An agent will choose the wrong airport, violate policy or strand somebody overnight. Nobody on the incident call will care how elegant the MCP connection looked.

They’ll ask the question waiting at the end of every enterprise software meeting:

Who, exactly, was responsible?

By 2029, the most valuable button in corporate travel will be invisible. I’ll trust the company that can prove who clicked it.

Frequently asked questions

What can ChatGPT and Claude do through Perk?

Perk’s MCP integration allows employees to manage travel, submit expenses and organize events through ChatGPT and Claude. Perk remains the execution layer, applying authentication, user permissions, company policy and existing approval workflows when an itinerary or action falls outside the traveler’s authorized limits.

How does Perk prevent AI assistants from violating corporate travel policy?

Requests made through ChatGPT or Claude still pass through Perk authentication and user permissions. Out-of-policy itineraries enter the company’s existing approval workflow, while Perk retains control over identity, policy enforcement, booking execution, expense handling and the transaction record.

Will AI assistants replace human corporate travel agents?

AI assistants can complete eligible bookings, exchanges, cancellations and expense workflows, but human escalation remains necessary for exceptions and disruptions. Effective systems stop before expensive mistakes, request approval when required, preserve an audit record and route unresolved situations to a person.