There’s a New Line Coming to the Hotel P&L. I Call It Cost Per AI Task (CPAT).

Range Anxiety, 2.0. Why thinking just became a variable cost.

It was 8pm on a Friday, and my AI assistant had run out of fuel.

I was eleven slides into a thirty-slide deck for a presentation I needed to give on Monday. The build was actually going well. Then a little message appeared in the corner of the screen, the digital equivalent of an empty fuel light, telling me I’d hit my session limit. Wait four to five hours, or dip into extra credits. There was no halfway option. You can’t tell a half-built deck to come back tomorrow.

So I paid. Of course I paid. Then I sat there for a second, slightly stunned, watching the meter tick down and thinking the same thought a lot of us are quietly thinking right now.

When did thinking become a variable cost?

This is the new anxiety of the AI age. The internet has already named it “token anxiety” (or “credit anxiety”) and the memes are stacking up. It’s the same low hum you get when your phone battery slips under 20% and you’re nowhere near a charger. Every small action suddenly costs more than it should.

We’ve built ourselves a fuel gauge for our brains. And like all fuel gauges, it changes how we drive.

The future: intelligence as a P&L item

For most of the hotel industry, “AI cost” right now is still a tidy subscription line. Microsoft Copilot. An enterprise ChatGPT seat. Maybe a white-labelled Claude or Gemini for the marketing team. Predictable. Manageable. Roughly the same as a Salesforce licence.

That phase is ending faster than people realise.

Goldman Sachs recently forecast that agentic AI could drive a 24-fold increase in token consumption by 2030 as agents do more work in the background for both consumers and enterprises. Microsoft’s own internal cost analyses, picked up by Fortune in May, show that in some workflows the agents are already more expensive to run than the humans they replaced. Uber’s CTO admitted the company burned through its entire 2026 AI coding tools budget in four months. One growth-stage SaaS firm with 35 engineers opened an $87,000 bill from a single April. A healthcare enterprise quietly consumed a trillion tokens over six months, around $6 million of unplanned spend, before the finance team fully understood what was driving the curve.

Most of these are tech-sector stories. But the structure travels. Wherever agents are doing real work, the same dynamic shows up. Tokens compound geometrically because every step of a reasoning loop sends the whole accumulated context back to the model. Chatbots sip. Agents drink. Apparently 50x more, on average.

Translated into hotel-speak: intelligence is about to behave less like a SaaS subscription and more like electricity. A line on the P&L that scales with usage, fluctuates with demand, and demands real management.

The risk: optimising the wrong thing

Two patterns worth watching, both of which the wider economy has already crashed into.

The first is the lesson from the early enthusiasm. Forrester’s 2026 Future of Work report found that around 55% of employers regretted AI-related workforce decisions, and Gartner expects roughly half of those companies to rebuild some of those roles within the year. The pattern, across several sectors, has looked similar. AI handled the easier 80% well. The trickier 20%, the work that requires human-to-human trust, judgement, or accountability, has been harder to automate cleanly. The full cost has tended to show up later, often in places nobody had budgeted for.

The second is what one analyst called “token maxing.” Defaulting to the most expensive, most capable model for every task, with zero routing logic or cost visibility. The AI equivalent of plugging an iPhone into a fast charger when it’s already at 90%. Enterprises that introduced intelligent routing (cheaper models for simpler jobs, premium models only when the work demands it) have been cutting spend by 60-80% without users noticing a difference.

For our industry, the comfort is that we’re still mostly upstream of both problems. We haven’t seen large-scale AI workforce decisions in hospitality. Most of us aren’t running unsupervised agent fleets. The intelligence line on the P&L is small. For now. The risk is reading this as permission to ignore the issue rather than permission to learn from those who moved faster.

Hotels have always been slow on tech. That’s both a curse and a quiet superpower. The curse is the lag. The superpower is that, once in a generation, we get to watch other industries make the expensive mistakes first.

The range anxiety lesson

Here’s the part I find genuinely reassuring.

We have done this before.

Range anxiety nearly killed the electric car. In 2012, the Nissan Leaf had real-world range of about 70 miles, charging networks were patchy, and every long drive was a logistics exercise involving printed-out station maps and a hopeful prayer at the Cracker Barrel parking lot. The internet was full of “EVs will never work” takes.

Range anxiety didn’t kill EVs. It forced three things to happen at once. Batteries got better. Charging networks got denser. And drivers developed a new literacy around when to charge, how to plan, and when an EV was simply the wrong tool for a particular trip.

Token anxiety is on the same arc. Unit economics are collapsing as inference costs fall year over year. Routing infrastructure is emerging. Subscriptions are evolving to reflect real usage rather than $20 flat-rate-for-everyone economics. And slowly, a new literacy is forming around when to spend a token and when to spend a brain cell.

The anxiety phase is not a forever phase. But while we’re in it, we have to actually live through it. Which brings us to the practical part.

CPAT: the new line you’ll need to manage

Hotels have CPOR. Cost per occupied room. Every GM can, hopefully, quote theirs. Numbers vary by region, segment, and currency, but every operator watches theirs like a hawk. CPOR is how we know whether the business is healthy at the unit level.

The next decade will require its sibling. Call it CPAT. Cost per AI task. The all-in cost of AI doing a discrete piece of work on behalf of your hotel, your team, or your guests, normalised against something you actually care about (a booking, a stay, a service interaction, a piece of content created).

Right now CPAT is invisible because most AI cost is buried inside subscription stacks. It will not stay invisible. Marriott has been publicly building an “agentic mesh” across marketing, operations, customer service and revenue (as the company’s CIO described in industry interviews earlier this year), as part of a broader multi-year technology transformation. Hilton, as reported by Skift in late 2025, has been running 41 AI experiments, and by their own honest admission only three paid back inside six months. IDeaS, Cendyn, Revinate and the rest are baking more intelligence into their platforms with every release. The cost of “thinking on behalf of your hotel” will, sooner than most people expect, show up as a line item you can measure.

Three habits worth building before that line gets big.

Build an intelligence budget by department. The same way you build a labour budget. Marketing gets X. Revenue gets Y. Operations gets Z. Without it, the spend goes wherever the loudest user is, not wherever the highest ROI sits.

Match the model to the job. You wouldn’t put your GM on a folio reconciliation. So don’t put a frontier model on a spell-check. The discipline of “what’s the cheapest model that can do this well” is the new “what’s the cheapest channel that can fill this room.”

Set agent boundaries before agent budgets. You wouldn’t let a junior team member work unsupervised for 12 hours. Not because of trust, but because of compounding. Every hour of unchecked work is an hour of decisions you haven’t seen. The same logic applies to an agent burning tokens at 3am.

None of this requires a PhD. It requires the same operational instinct hoteliers have been using for a hundred years. You manage what you measure, and you measure what you spend.

Back to the deck

The deck eventually got finished. The result was good. The extra credits I burned at 8pm probably cost less than a round of drinks at the lobby bar.

But the small jolt of watching the meter tick down has stayed with me. Not because the money mattered, but because the gauge mattered. For the first time in my working life, I could see the cost of thinking in real time.

We’re not going to put that gauge back in the box. We’re going to learn to read it.

We’ve always been in the business of managing scarce things and selling them to people who really want them. Rooms. Time. Attention. Now there’s a new one on the inventory list. Intelligence.

Know your range. Plan your route. Then, deliver the magnificent.