Token Creep: The Variable AI Cost Owners Have Not Budgeted

Most hotel software is a fixed monthly fee. AI is billed by the token, so the same tool costs more when guests write longer messages, when someone adds the full menu to the prompt, or when an agent retries a step. Gartner says AI cost estimates can be off by 500 to 1,000 percent. The fix is a cost control layer: one gateway every AI call passes through, tagged by property and workflow, with a router that sends simple requests to cheap models, prompt caching, hard caps, and a monthly owner report that prices AI per occupied room.

It's November, so somewhere an owner is reading a 2027 budget with a technology line that was copied forward from this year with a 3 percent escalator.

I've built that line more times than I can count. At Starwood, where I ran sales strategy and operations across more than 800 hotels, and later as an owner of ten hotels of my own, the technology section of the budget was the easy part. PMS fee, channel manager, the guest Wi-Fi contract, a handful of seat licenses. Each one a fixed monthly number. The variance column next to them on the monthly report was zero so often that nobody read it.

AI doesn't fit in that column. It's billed by the token, which means by usage, which means the number moves every month for reasons that have nothing to do with the contract you signed.

This piece is about that gap: why AI costs creep, what actually drives them at a hotel, and the specific layer of technology that turns a variable line back into one an owner can budget, cap and read.

The Problem: A Fixed Line That Started Moving

A token is a fragment of text, roughly three quarters of a word. Every time a guest messaging assistant answers a question, the model reads everything it was sent (the instructions, the property facts, the conversation so far, the guest's new message) and writes a reply. You pay for both sides, per million tokens, and the reply side usually costs four or five times more per token than the reading side.

That pricing has three consequences most hotel budgets haven't absorbed.

First, the cost scales with volume you don't control. A wet weekend, a delayed flight bank, a group arriving at once. Guest message volume spikes and so does the bill.

Second, it scales with design decisions nobody prices. When someone on the team pastes the full spa menu, the F&B hours and the cancellation policy into the assistant's instructions "so it always knows," every single message now carries those extra tokens. The tool got more accurate and more expensive in the same afternoon, and nobody signed anything.

Third, the unit price itself moves. Google's published Gemini API price list shows Gemini 3.8 Flash at $0.75 per million input tokens and $3.75 per million output tokens through December 31, 2026, rising to $1.50 and $7.50 from January 1, 2027. Its flagship model doubles a month earlier, on December 1. If your vendor builds on either and passes the cost through, your per-conversation cost doubles on a date you probably haven't heard about.

Gartner saw this coming. At its 2024 Symposium in Australia, distinguished VP analyst Mary Mesaglio told the audience that "500 to 1000 percent errors of AI cost estimates are possible," according to The Register. One of the causes she named was token pricing itself: once people trust the tool, they "start writing longer and more complex queries that cost more."

In June 2026 the firm went further. "Through 2028, at least 50% of GenAI projects will overrun their budgeted costs due to poor architectural choices and lack of operational know-how," its report said, as covered by Campus Technology. It expects inference, the per-use running cost, to account for at least 70 percent of a model's lifetime costs, and it flagged agents directly: "AI agents can trigger multiple model calls, retrieve data, access external tools, and execute multi-step workflows."

That last sentence is the one hotels should read twice.

The tool got more accurate and more expensive in the same afternoon, and nobody signed anything.

The Data: What a Guest Interaction Actually Costs

List prices are public, so the math is checkable. Here are current per-token prices for the model tiers most hotel AI products are built on.

The spread between the cheapest and most expensive row is 40 to 1 on input. That ratio is the whole opportunity, and the whole risk.

To turn prices into something an owner recognizes, I modeled two interactions. A simple guest conversation: four back-and-forth turns, each sending about 6,000 tokens (instructions, property facts, history) and getting about 250 back. And an agentic task, like changing a reservation and confirming it by email: ten model calls, averaging 15,000 tokens in and 400 out, because each step carries the tool results from the steps before it. Then a month at a 150-room hotel: 2,500 guest conversations and 600 agentic tasks.

At one hotel, those dollars are small. I'll say that plainly, because some of the AI cost commentary out there is written for companies spending millions.

The problem is the shape. A $349 line that becomes $1,300 by spring is a 270 percent variance on a line the owner was told was fixed, and across a ten-hotel portfolio that run rate is more than $110,000 a year over budget. When I was signing software orders as founder and CEO of Collective Retreats, ten locations by the time it was acquired in 2025, the conversations I remember weren't about big numbers. They were about numbers nobody could explain.

Here is how that $349 gets to $1,300 without anyone making a bad decision.

Every one of those changes was reasonable. More channels is good. Handing upsells to the agent is the point of buying it. The richer prompt reduced wrong answers. Nobody owned the sum.

The wider data says hotels aren't unusual here. McKinsey's State of AI in 2026, surveying 1,719 people in 97 countries, found one in five respondents say their organization is limiting AI use because of operating costs, a share "broadly consistent across organizations of different sizes and across many industries." It also found 60 percent expect to increase AI investment next year. Those two numbers together describe a budget that's about to grow faster than anyone's ability to read it.

And small firms are arriving in force. The CFO Survey from Duke and the Richmond and Atlanta Feds reported in December 2025 that 48 percent of small companies invested in AI in 2025, and about 80 percent expected to in 2026. Most independent hotels are small companies.

The Mechanism: A Cost Control Layer Between the Hotel and the Models

The fix isn't to avoid usage pricing. Usage pricing is fair: you pay for what you use, and the cheapest models are astonishingly cheap. The fix is to put one piece of software between every AI feature at the property and every model it calls, so usage becomes visible, routed and capped.

Engineers call it an AI gateway. Open-source options like LiteLLM and hosted ones like Cloudflare AI Gateway already do the plumbing: one endpoint, logging of every request, spend tracking, rate limits, caching. What a hotel needs is that plumbing configured around its own P&L. Six parts.

1. One door. Every AI call from every tool the hotel builds or controls goes through the gateway with a single set of keys. No staff member, vendor integration or side project holds its own model API key. This is the step that makes everything else possible, and it's mostly a policy decision.

2. Tags on every call. Each request carries a property code, a workflow name (guest messaging, review replies, the Monday revenue report) and, where relevant, a reservation or ticket ID. The gateway logs input tokens, output tokens, cached tokens, model and cost. Now the bill has rows.

3. A router. A small, cheap classifier reads each incoming request first and decides which model should answer it. "What time is checkout?" goes to the smallest tier. A multi-step change to a group block goes to the large one. If the small model's answer fails a confidence check, the request escalates. The research behind this is solid. Stanford's FrugalGPT showed a cascade of models "can match the performance of the best individual LLM (e.g. GPT-4) with up to 98% cost reduction." The LMSYS team's RouteLLM reported cost reductions of over 85 percent on one benchmark "while still achieving 95% of GPT-4's performance." Hotel guest traffic is mostly simple questions, which is exactly the traffic routing is good at.

4. Caching and context discipline. The instructions and property facts at the top of every request are identical thousands of times a day. Model providers discount repeated content heavily. On Anthropic's price list, a cache hit on Sonnet 5.5 costs $0.10 per million tokens against $2.00 for fresh input, and Google prices context caching at roughly a tenth of input. The gateway structures prompts so the stable part sits first and gets cached. Separately, it retrieves only the facts a question needs (the spa menu when someone asks about the spa) instead of pasting everything into every call.

5. Batch for anything that can wait. Review responses, the overnight report, tagging last week's guest feedback. None of it needs an answer in two seconds. Anthropic and Google both list batch processing at half the standard price, and OpenAI offers a Batch API for the same kind of deferred work.

6. Caps that act, and a report that prices AI per occupied room. Each workflow gets a daily soft limit that alerts and a monthly hard limit that downgrades to the small model or hands off to a person. Then the gateway's log rolls up into one monthly page: cost per occupied room, cost per guest conversation, cost per workflow, and the hours or revenue each workflow returned.

That last report matters more than any single saving. An owner can approve a variable line. What an owner can't approve is a variable line with no unit.

An owner can approve a variable line. What an owner can't approve is a variable line with no unit.

Run the month-6 volumes from the creep table through this layer, with 80 percent of conversations routed to the small tier, the rest and every agentic task on the large tier with cached instructions, and the model cost drops from $1,293 to about $252. That's below month 1, at more than twice the traffic.

The Framework: Guardrails an Owner Can Read

Most of what hotels buy isn't a raw model. It's a vendor product with AI inside, priced per conversation, per credit or per "AI action." The same creep happens there, one layer removed, and the gateway can't see inside it. So the guardrails have two halves: the ones the hotel runs itself, and the ones it writes into vendor contracts.

Three notes on using it.

Budget AI as its own line. If it's buried inside "software subscriptions," a doubling disappears into the rounding. Under the USALI structure most owners already read, it belongs with information and telecommunications systems, as its own row.

Budget it in units, not dollars. Cost per occupied room and cost per guest conversation survive occupancy swings; a flat monthly dollar figure doesn't. If occupancy runs 8 points ahead of budget in July, AI cost should too, and nobody should get a variance memo for it.

Ask every vendor for three things. A monthly usage export at the conversation or action level. A contractual cap on overage charges. And written notice before any price change in the underlying model is passed through. The vendors I'd trust will say yes to all three. If the answer is no, you've learned how they plan to grow revenue. Our pieces on negotiating the hotel AI contract and reading software fine print go further on the contract side.

Implementation: Six Weeks From an Unread Invoice to a Line With a Unit

None of this needs a data team. It needs one person who owns the outcome, a few days of engineering or a capable partner, and three months of patience for the numbers to settle.

Two cautions from doing this kind of thing.

Shadow the router before it goes live. A week of logging what it would have picked, compared with what the large model actually said, tells you where the small tier gets things wrong. At a hotel those are usually questions that look simple and aren't: "can I bring my dog" at a property with a weight limit and a fee that changes by season.

Watch the agents hardest. Gartner's warning about multiple model calls per task is the one I'd put on the wall. An agent stuck in a retry loop against a PMS that's timing out can burn a week's budget overnight, and it'll look like a busy night in the logs. The hard cap is the backstop for that, which is why it downgrades or stops rather than just sending an email.

If you'd like an outside read on where your AI spend sits today and what a gateway would change, a structured technology and AI audit is a sensible first step. It starts with the register in week one, which most hotels find useful on its own. The total cost of ownership piece covers the costs around the tokens: integration, training and the hours staff spend checking the output.

The models will keep getting cheaper per token. Hotels will keep finding more for them to do. Which of those two curves wins at your property is decided by whether anyone is reading the meter.

Frequently Asked Questions