Read the Meter

In Budget the Meter, Not the Machine, I argued that AI doesn’t come with a subscription fee, it comes with a meter. In The Meter Starts Before Hello, I showed that the meter now runs before the guest arrives, and that someone will be invoiced for every look an AI agent takes.

Both pieces assumed something most hotels can’t yet do: read the meter.

McKinsey has just made that easier. In a new explainer, What does AI really cost?, three of its experts, Lari Hämäläinen, Kaavini Takkar and Wasim Lala, take the AI bill apart line by line. Takkar’s verdict sums up the problem: “AI cost is perhaps not hard to decipher, but it is very complicated.”

Lala’s sums up the answer: “It’s not just about cost per token. It’s about cost per task.”

In a hotel, the task is a guest served or a booking that becomes a stay. That is TCPG, Token Cost Per Guest, and its front end, TCPB, Token Cost Per Booking. But before anyone can count cost per guest, they need to know what they are counting.

Token literacy: every word runs the meter

Hotels taught a generation of managers to read a utility bill: kilowatt-hours, peak tariffs, the standing charge. AI needs the same literacy, and most teams don’t have it yet. Call it token literacy.

The McKinsey explainer is the clearest primer I have seen, so here it is in its experts’ own words, with what each idea means at a hotel.

Quotations from McKinsey, “What does AI really cost?” (14 October 2026). Hotel column: Pertlink.

The memory line deserves a worked example. Take a guest who chats with a hotel’s AI concierge across a five-night stay, adding roughly 2,000 tokens of conversation a day. Because the model is stateless, each day’s reply means re-reading the whole thread: 2,000 tokens on day one, 4,000 on day two, rising to 10,000 on day five. By checkout, the model has read 30,000 tokens to hold a 10,000-token conversation. (The figures are Pertlink’s illustration, and caching can soften them.)

That is token literacy in one line: every word runs the meter, including the ones nobody reads.

What’s actually on the bill

Takkar names four places AI spend hides in a company. Each has a hotel address.

• Direct spend with AI providers. The hotel’s own accounts with the model makers, and the staff subscriptions somebody put on a corporate card.

• Cloud compute behind every model call. Often billed to IT, or to a technology partner, rather than to the department using the AI.

• AI built into software the hotel already buys. The guest messaging platform, the review tool, the revenue system. For many hotels this is the largest AI line, and the least visible, because it sits inside a vendor fee.

• Experimentation. The pilot at one property, the trial a department head signed up for, the tool nobody switched off.

An AI agent adds its own line items. Takkar’s starting point is simple: “Every single time an agent receives a prompt, it is going to make model calls. You multiply your model calls by the per-token price.” Then come the costs on top: human validation, monitoring, storage of the hotel’s knowledge for the AI to search, the plumbing that connects the agent to other systems, and the cost of failure.

Failure is the line most business cases leave out. Takkar’s example is an airline rebooking agent that offers a flight, waits for the traveler to decide, then has to say: “This flight is actually already gone. Let’s go through this process again.” Every turn of that loop runs the meter again.

The hotel version is easy to picture: the upsell agent that sells a room type already sold out, or the concierge that confirms a spa slot someone else has taken. The guest gets an apology. The hotel gets the bill for both conversations.

Why the bill jumps

One of the explainer’s section titles asks the question every owner will ask: “If the price per token keeps falling, why are the bills going up?” The answer is that price per token was never the number that mattered. Volume, design, and discipline are.

The McKinsey experts give five ways a bill jumps. Each has a hotel equivalent.

Only one of these is a change in price. The other four are management decisions that nobody knew they were making.

Cost per successful output

So how do you know if an agent is worth it? Takkar’s answer: “One good metric to look at would be the cost of the agent per successful output.” Then compare that, she says, with “the fully loaded cost of a human or of another tool that had been doing the job.”

That is the TCPG principle, put in McKinsey’s language. Two parts of it matter most for hotels.

Failures count. The upsell that sold a room already gone isn’t a successful output. Neither is the review reply a manager had to rewrite. Their tokens go into the numerator, but they add nothing to the denominator. That is why TCPG is divided by guests served, and TCPB by bookings that become stays, never by conversations held.

Frequency decides. Takkar contrasts the thousands of rebooking requests an airline receives every day with an agent that audits employees’ tax filings “perhaps once a year.” The first earns its build cost back quickly. The second may never. Hotels have both kinds of work. Guest messaging runs all day, every day. The annual budget pack runs once. An agent may suit the first long before it suits the second, however impressive the demo.

Three kinds of AI, three different sums

Hämäläinen makes a point that would save many hotel budget meetings: “Today, we’re using AI very much as a blanket term. If you simplify it, there are three different types of AI.” Each one is judged differently.

1. Personal tooling, the AI “that affects your work as a person.” At a hotel: the sales manager drafting proposals, the GM summarizing the month’s reports. Its value is real but indirect, so treat it as a cost of doing business. Cap it, govern it, and don’t expect a return on every prompt.

2. Automation and efficiency: “We take what exists today and rebuild it for higher productivity.” At a hotel: review replies, invoice matching, the guest questions asked a hundred times a week. This is where cost per successful output, TCPG and TCPB, earn their keep, because the work is countable.

3. Rebuilding for AI, “running the business differently.” At a hotel: redesigning the front office so AI carries the routine and people own the moments that matter. That is the thinking behind Pertlink’s Human Experience Orchestrator (HXO). It is judged over a longer horizon, on guest outcomes as well as cost.

Put all three in one budget line and a staff subscription, a guest-facing agent and a transformation program end up judged by the same yardstick, and all three look wrong.

Who reads the meter

The explainer’s experts agree on one thing: AI cost can be controlled. As Takkar puts it, “with the right visibility, it’s solvable.” Their advice translates to a hotel with very little editing.

• Put every source of AI spend in one place. Takkar’s first step is a single AI control plane: “one place where I can bring all these different sources of my AI spend together.” I have argued for the same in Who Controls the Controls? It is also the answer to a question a reader asked me on LinkedIn this week: how do you measure token burn when vendors don’t show it? You bring every bill into one view, and you make usage reporting a condition of the contract.

• Set the guardrails before handing out access. The bank’s $25 million projection came from access first, rules later.

• Send each task to the right model. Breakfast hours don’t need the most powerful model, and some questions don’t need a model at all. Takkar’s question to companies is “why didn’t you route this particular request to an open-weight model?” Hämäläinen goes further, favoring open-source models “where you can control what the model does and where it runs.”

• Teach token literacy. Lala’s test for any company: “Am I teaching my employees how to write the prompts so I’m not overloading the context window?” It belongs in every AI literacy program; HFTP’s AI in Hospitality: Essentials included.

• Redesign the work before automating it. Takkar’s first step is to ask what workflows the organization has and how to reimagine them. “Then I should think about the AI aspect of it all.”

• Stay in control. “Companies need to be in control of their own AI,” says Hämäläinen.

• Act on what you find. “You can always find the savings,” says Lala. “Executing on them is the hard part.”

And measure everything. Lala’s line could hang in every hotel finance office:

AI will be worth it as long as you can measure it. Without measurement, there can be no success.

Wasim Lala, McKinsey

A meter-reading checklist

Six questions for owners and general managers, before the 2027 budget is signed.

1. Can we list every place we pay for AI? That means provider accounts, cloud, pilots, and the AI inside every vendor fee.

2. Do we know our TCPG and TCPB? Count only successful outcomes: guests served, bookings that become stays.

3. Is each AI use judged by the right sum? Personal tooling, automation, and rebuilding each need their own measure.

4. Which model answers our simplest guest question, and does it need to?

5. What happens to the bill when a pilot becomes a portfolio rollout, or a vendor moves to usage pricing? McKinsey’s research says spend nearly quadruples from pilot to production. Budget for it.

6. Who on the team can explain what a token is? If nobody can, start there.

Budget it. Know where it starts. Read it.

Three pieces, one meter. Part One said put AI on a meter, not a license, and budget for it. Part Two showed that the meter starts at the look, and that an agent’s looking costs money that goes back to the hotel. Part Three is the simplest and the most overdue: learn to read the meter you already have.

Token literacy isn’t a technical nicety. It is the difference between a hotel that can say what its AI costs per guest and one that finds out when the invoice arrives.

Every word runs the meter. Make sure someone in the building can read it.

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. Hämäläinen, L., Takkar, K., and Lala, W. (14 October 2026). What does AI really cost? McKinsey Explainers, McKinsey & Company.

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

3. Ronson, T. (21 October 2026). The Meter Starts Before Hello. Pertlink Viewpoint.

4. Ronson, T. (2026). Who Controls the Controls? Parts 1 and 2. Pertlink Viewpoint.

All quotations and figures in this paper are from the McKinsey explainer, attributed to the expert who gave them. The hotel examples, the five-night worked example, the checklist, and TCPG, TCPB, and HXO are Pertlink’s.