The Meter Starts Before Hello

This week, Shiji Insights published a line of mine: AI doesn’t come with a subscription fee; it comes with a meter, and that meter runs every time a guest says hello.

A day later, PhocusWire showed me the meter starts earlier. It runs at the look.

The piece, Return on tokens: Who pays for the free travel agent?, is by Eugenia Dunaeva, an independent director at TravelAI who spent 11 years at Expedia on both the buy and the sell side. Most travel executives can recite their look-to-book ratio, she observes, yet hardly any can say how many tokens a booking takes.

Read with a hotelier’s eye, her numbers say something more pointed. The “free” travel agent isn’t free. Every booking it makes carries a cost of sale. Today, investors pay it. Costs of sale don’t stay unpaid for long, and in travel they tend to land on the hotel.

The free agent’s cost of sale

Dunaeva builds on a measure named by Packy McCormick and Markie Wagner in Not Boring: return on tokens, or ROT.

Her worked example shows what that means for a travel agent. It fields 1,000 requests and 100 become bookings that stick, earning $30 each. That is $3,000 in revenue. At 60 cents of model and tool cost per request, the token bill is $600, because it is paid on all 1,000 requests, not just the 100 that converted.

Divide that bill by the bookings, and you get a number I call Token Cost Per Booking, or TCPB. Here it is $6 a booking, a fifth of the $30 her example assumes the agent earns, before payment, distribution and support take their share.

That $6 is the platform’s cost of sale. And it moves fast. Change only the agent’s behavior in her example (the two scenarios are Pertlink’s, not hers):

1,000 requests at 60 cents each, $30 earned per booking that sticks. First row: Dunaeva, PhocusWire. Other rows: Pertlink illustration.

The spare rooms are not hypothetical. With free cancellation, Dunaeva notes, an agent can book the best room, then the second best, then the third, and release the extras at the last minute.

Nobody made a mistake in either case. The meter ran. The more an agent looks and holds, the more it costs to sell a room, and the harder its backers will look for someone else to pay.

Who pays: we have seen this film before

Right now, Dunaeva points out, only the agents' investors pay for the looking. Instinct is free to travelers, and so is Meta’s Muse, up to a usage limit. Neither takes a commission yet. Booking Holdings’ new agent, Vayo, is free in beta and hasn’t decided how it will make money.

There are only three places that cost can go. Investors won’t fund it forever. Travelers won’t pay, because free is the hook. That leaves the supplier. Instinct’s founder, Noah Shinn, has already floated the idea that boutique hotels would pay up to 30% per booking.

Hotels have seen this film. Online travel agencies and metasearch were free for travelers too. Hotels funded them through commission and paid per click. Personal agents are the next installment, with a meter on every look.

The route to the hotel may not even be direct. Expedia has signed up to sell hotels inside Muse, as merchant of record. If agent platforms start charging their partners, the cost has an obvious path to the hotel: inside a commission it already pays.

Why a GM should care now

“That’s the platform’s bill, not mine” is a fair reply today. It won’t be in the next contract negotiation.

A platform’s cost of sale sets the floor under what it will charge. It doesn’t set the price. Platforms price at value, which is why the figure floated is 30% and not $6. But a platform that loses money on every booking at zero commission has only one direction to go.

This is where TCPG comes in. For more than a year, I have argued that hotels need a number they don’t keep: Token Cost Per Guest. It is the AI cost of serving one guest across the whole journey, from the first pre-arrival question to the reply to their post-stay review. It is fully loaded, including orchestration, retries, and the person who checks the output. And it is never read alone: a rising TCPG next to rising revenue is a good month; next to flat outcomes, it is waste.

McKinsey now counts AI cost the same way. Its new AI value curve, from QuantumBlack, AI by McKinsey, prices a task done with AI as its model and token costs plus the human review, validation, and rework it still needs.

TCPB is its front end: the pre-arrival slice of the same spend, divided by bookings that become stays.

The agent platform will know its TCPB. The question is whether you know yours. Your own direct channel has one too: the booking assistant, the pre-arrival messaging, the rate tools. Know that number, and you can set what a direct booking costs you in tokens, besides what an agent platform will ask. Without it, there is nothing to negotiate with.

Less room for error than most

That same McKinsey report explains why hotels can’t afford to guess. Its value curve compares the cost of a task done today with the full cost of doing it with AI, industry by industry. Hospitality and food services sit in the bottom third globally, and last of the industries McKinsey charts for the United States. Waiters and waitresses fall below zero: at today’s wages, applying AI to their tasks costs more than the work does now.

The report’s second finding matters for Asia. AI economics are local. AI costs are broadly similar everywhere, but labor costs are not, so a use case that pays off in one market may not pay off in another. Pertlink’s reading: for hotels across much of Asia-Pacific, where wages sit well below US levels, the margin AI has to work with is thinner still.

Put the two arguments together. An industry whose AI margins are already thin is about to meet agents whose cost of sale someone will have to pay. A few dollars of TCPB passed down the chain matters more to a hotel than to a bank.

Make yourself cheaper to sell

Here is the part most hotels miss. The agent’s cost of sale isn’t fixed. Part of it depends on you.

Every extra page an agent has to read, every form it has to fill in twice, every block it has to retry, adds to its bill. A hotel whose rates, rooms and policies are clear and readable by machines is cheaper to sell than one the agent has to fight. Cheaper to sell is a negotiating position.

Most hotels haven’t taken that position, because they haven’t decided anything. On 18 October Pertlink sent an AI agent to book a one-night stay on the websites of the world’s ten largest hotel groups. Six let it reach checkout. Two stopped it with bot-management software rather than any published AI policy, and a repeat test showed the outcome could change from one day to the next, and from one agent to another. Two kept booking behind apps and SMS logins.

Hotels haven’t closed their doors to AI agents. Most haven’t decided whether to open them. A firewall setting isn’t a policy, and it can’t be priced.

The tools to price it are arriving. On 6 October, Sierra and Meta announced the Personal Agent Protocol, known as Poppy, an open standard for personal agents to identify themselves to businesses. Instinct, Shopify, Stripe and Walmart were among the first partners. Three days later, Sierra added 35 design partners, including United Airlines and Hertz. Once an agent identifies itself, Dunaeva notes, a supplier can price the look.

Airlines and car rental are at the table. No hotel group is on either list.

The hotels that are cheap to sell and know who is looking will negotiate. The ones that are expensive to sell and can’t see who is looking will be invoiced.

Bumps in the road

No argument survives without its objections. Here are the ones I would raise if I sat on the other side of the table.

Back to tokenomics

In June I wrote about tokenmaxxing: rewarding people for burning tokens, as though consumption were progress. The Not Boring essay that named ROT calls tokenmaxxing a mass delusion. ROT is the correction, and TCPB is what the correction looks like at the point of sale.

McKinsey reaches the same verdict. Many organizations, its value curve report says, still measure AI adoption instead of value, tracking users, copilots deployed and tokens consumed. Its advice is to manage AI like any other factor cost, down to the “cost per useful outcome”, and it calls the AI FinOps discipline “a requirement, not an option.” In a hotel, the useful outcome is a guest served or a booking that becomes a stay. That is TCPG and TCPB.

The correction now has an institution. In June, the Linux Foundation announced plans to launch the Tokenomics Foundation, working with the FinOps Foundation to set open standards for measuring and benchmarking AI token spend. Its initial supporters include Booking.com, alongside Google Cloud, Microsoft, IBM, JPMorgan Chase, and Salesforce. Booking.com’s IT finance lead gave the reason: at its volume, small differences in token efficiency compound into very large numbers.

Let that sit. One of the world’s largest booking platforms is helping write the rules for how tokens are counted. Hotels aren’t in the room.

Standards bodies will define the unit: the token, its price, the benchmark. They won’t define a hotel’s denominator. That is our job. If hospitality doesn’t decide what tokens should cost per guest and per booking, the platforms between the guest and us will decide for us, and invoice accordingly.

Five questions for the board

Dunaeva closes with questions for travel boards. Here are mine, for hotel owners and operators.

1. What does an agent booking cost the platform to make, and what will it ask us to pay? Its TCPB is the floor. Its asking price will be higher.

2. What is our own TCPB on direct bookings, and what happens to it if AI costs rise 30 percent? Without it, we can’t say whether a commission is fair. McKinsey’s advice is to stress-test every AI business case both ways.

3. How much of our booking-engine traffic is machines, and who decided how we treat it: us, or our bot-management vendor?

4. Are we cheap or expensive for an agent to sell? Clear rates, clear policies, and content machines can read all lower the agent’s cost, and strengthen our hand.

5. If a personal agent identified itself tomorrow, what would we let it do, on what cancellation terms, and at what price? That is a policy, not a firewall setting.

The look has a price

Dunaeva ends on a lesson travel learned slowly and at great cost: “a search is not a booking.” Agents, she warns, will teach it again, and this time each look comes with a bill.

I’d finish the thought for hotels. For twenty years the battle was the booking. Then it became the context before the booking. Now that context carries a meter of its own, and whoever owns the meter will want to be paid.

The meter used to start at hello. It now starts at the look. Know your TCPG. Learn your TCPB. And decide who pays for the looking before someone decides for you.

The intelligence may be artificial. But the experience is human.

Made with the help of various SI tools, but always with a HITL

Sources

1. Dunaeva, E. (20 October 2026). Return on tokens: Who pays for the free travel agent? PhocusWire.

2. Van der Veken, L., Durth, S., Catlin, T., Alatovic, T., and Isherwood, T., with Thirumoorthy, D., Kazantsev, D., and Nanabhay, S. (October 2026). The AI value curve: The economics of applying AI to existing work. QuantumBlack, AI by McKinsey.

3. McCormick, P., with Wagner, M. (10 June 2026). Return on Tokens (ROT). Not Boring.

4. Chan, W. (19 October 2026). Budget the Meter, Not the Machine: Planning for 2027 When AI Runs on Usage. Shiji Insights.

5. PhocusWire (24 September 2026). Report on Expedia hotel booking inside Meta’s Muse.

6. PhocusWire (27 July 2026). Report on the Trip.com Group antitrust penalty.

7. The Linux Foundation (3 June 2026). Linux Foundation Announces the Intent to Launch the Tokenomics Foundation.

8. Taylor, B., and Bavor, C. (9 October 2026). Sharing a draft of Personal Agent Protocol. Sierra.

9. Pertlink (18–19 October 2026). Are Hotels Closing Their Doors to AI Agents? Parts One and Two.

The worked example of 1,000 requests and 100 bookings is Dunaeva’s. The AI value curve, its industry rankings, and its cost scenarios are McKinsey’s. TCPB, the two scenarios that change Dunaeva’s example, the hotel-side argument, and the Asia-Pacific reading of McKinsey’s findings are Pertlink’s.