Nobody Wants to Be First

Owners are right to be careful. More than 40% of agentic AI projects will be canceled by the end of 2027, according to Gartner, which estimates only about 130 of the thousands of agentic AI vendors are real. MIT studied 300 deployments and found that 95% of pilots delivered no measurable P&L impact. In our industry, 98 percent of hotel owners and developers have started incorporating AI, only 32 percent report AI embedded across most operational areas, and 73 percent want to do more but are unsure where to start. Over 40% of agentic AI projects will be scrapped by 2027, Gartner says +5

So the hesitation is earned. Here’s what it costs.

Second is not safe. Second is a ten-year lag. When Marriott’s SVP of sales first pitched Bill Marriott on revenue management in 1985, Marriott had to chuckle. He went first anyway. By the mid-’90s, Marriott’s revenue management was adding $150 to $200 million in annual revenue. Every owner who waited for RM to be “proven” waited for Marriott to prove it, on Marriott’s hotels, for a decade. Nobody remembers who was second. ciophocuswire

And the brands aren’t waiting this time either. Marriott confirmed corporate job cuts on October 28. On Hilton’s Q3 call, Nassetta’s first technology priority was system efficiency: using AI, technology and process change to reduce corporate costs and the cost of running the system for owners, with RISE 2.0 coming in the first quarter of 2027. Read that carefully. The brands are going first on AI, for their own P&L. Whether any of it reaches your GOP line depends on what your operator is doing in the middle of the P&L. If your operator is waiting, the savings are flowing upstream without you. Promotion title +2

Candidly, the issue isn’t courage. It’s who holds the risk. In a standard HMA the owner pays a base fee on revenue whether the new thing works or not. The operator gets paid either way. So “nobody wants to be first” really means “nobody wants to pay to be first.” Then don’t. Make the operator go first. If the agents don’t move GOP, the operator doesn’t get paid. We’ve put our fee where our mouth is: we’ve offered owners deals where 100% of our compensation comes from GOP improvement over trailing twelve months. Any operator pitching you AI that won’t make that offer just told you what it thinks of its own agents.

But an operator can only make that bet on something it controls. Which brings me to the question nobody asked me this week.

Making Hospitality Leaders Smarter On: Who Employs Your Agents?

Start with the two numbers that frame everything. Hotels spend about 1.4 percent of total operating revenue on IT. We spend 35% of revenue on labor. Agents are not a software line. They are a labor line. They’re coming for the 35%, not the 1.4%. Which is why “which vendor should I buy” is the wrong question. The right one is the oldest question in our business: who is the employer of record? 10minhotel

Microsoft calls agents digital labor, and 82% of leaders are confident they’ll use digital labor to expand workforce capacity in the next 12-18 months. Its phrase is agent boss: someone who builds, delegates to and manages agents. 28% of managers are considering hiring AI workforce managers to lead hybrid teams of people and agents. We already did. We created an Agent Deployment Manager role in October. Microsoft's Latest Report Says 2025 Is The Year of the “Frontier Firm”—Here's What That ACTUALLY Means… +2

Now apply the rule we apply to every other worker in a hotel. In about 80%-90% of current HMAs, the manager is legally the employer of the hotel’s employees, and under most HMAs the operator has exclusive authority to operate the hotel on behalf of the owner. I ran a management company for eight years. That’s the entire job: hire, train, reorganize, redeploy. Owners pay for that control. jmbmcostar

If your operator’s agents are rented from a software vendor, your operator is not the employer of its digital workforce. It’s a client of a staffing agency. Five things follow.

1. You can’t redesign a process you don’t own. BCG puts AI value at 10 per cent algorithms, 20 per cent tech and data, 70 per cent people and processes. McKinsey’s winners are three times more likely than peers to fundamentally redesign workflows when deploying AI. Everyone else sprinkled AI on old processes, which is why only 39 percent see any EBIT impact, and most of those say it’s under 5 percent. MIT’s diagnosis is blunter: most GenAI systems do not retain feedback, adapt to context, or improve over time. One CIO said the vendor tool felt like a one-size-fits-all box that forced his teams to reshape their workflows to fit it. That’s every hotel SaaS contract you’ve ever signed. A vendor serving 3,000 hotels can’t let your GM change the workflow on Tuesday because of what she found on Monday. Its business model depends on you not changing the product. And MIT found the money was never in the shiny stuff: over half of GenAI budgets go to sales and marketing tools, yet the biggest ROI is in back-office automation. Back office is process. Process is ownership. AI-focused services make up 40 per cent of BCG’s 2025 revenue +5

2. Temp agencies mark up labor 25% to 40%. Rented agents mark it up 100x. Staffing agencies charge a 25% to 40% markup on contract labor, up to 75%. Now price digital labor. Salesforce Agentforce costs $2 per conversation. Intercom’s Fin charges $0.99 per resolution. The tokens underneath run about $0.002 for a six-turn support conversation on a small model and about $0.21 for a long sales conversation on a frontier model. Pennies to a quarter, sold for a dollar or two. Where does the spread go? Median SaaS gross margin is 77% and public SaaS companies spend a median 39% of revenue on sales and marketing. You’re funding your vendor’s sales team to sell to your comp set. Our sales agent runs on about $1,000 a month in tokens. The owner pays the tokens at cost. We get paid on GOP. How Much Do Staffing Agencies Charge for Fees in 2026? +6

3. Tokens deflate. Seat prices don’t. For an LLM of equivalent performance, cost is falling 10x every year. Stanford measured it: GPT-3.5-level inference fell 280 times between November 2022 and October 2024. Who captures that? Whoever owns the agent. An owned agent gets re-pointed at a cheaper model every quarter and the savings land in the hotel’s P&L. A rented agent gets a renewal notice. No per-seat price in the history of hotel tech has fallen 10x. a16ztomshardware

4. The Uber question. The sharpest thing an owner asked me this week: what if cheap tokens are a subsidy and the bill comes due? He’s right to ask. Rideshare prices rose 92% between January 2018 and July 2021 once the venture money stopped covering the fare. It’s starting in AI. AI vendors are pushing through price increases, and switching model vendors now means moving context, workflows and institutional memory. Google’s AI triples in price each year. xAI raised top-tier pricing 140% in August. And Gartner predicts inference cost per agentic workflow will rise more than fivefold through 2028 despite falling token prices, because agents burn 5 to 30 times the tokens per task of a chatbot. Why Uber And Lyft Rides Got So Expensive +5

So yes, the token bill can go up. Ownership is the hedge, not the exposure:

• You can route. An owned agent runs on whichever capable model is cheapest this quarter. OpenAI, Anthropic and Google all discount cached input tokens 90 percent; 9 of 15 providers take 50% off batch jobs. Open-weight is the floor: DeepSeek at $0.27 per million input tokens, and self-hosting decouples your per-token cost from third-party API rates. Why Falling Prices Have Not Reduced Bills +3

• You can govern tokens per task. Gartner’s own fix is inference tiering, token limits and monitoring. You can only tier, cap and monitor an agent whose trace logs you own. A rented agent’s token burn is invisible until the invoice. techstrong

• The arbitrage survives a 5x, and renting compounds the shock. A back-office task that costs a nickel in tokens replaces minutes of a $25-an-hour human. Multiply the nickel by five. Still over 90% cheaper. But if tokens triple and you rent, the increase arrives as a vendor price letter you can’t audit, stacked on the 77% margin and the 39% sales budget you were already funding. Every layer you don’t own marks up the risk.

5. What happens to the agents when you fire the operator? You’ve lived this with people: when the operator is the employer, the owner has more restrictions on what it can do with employees on termination of the operator or transfer of the hotel. Agents, prompts, learned workflows and the data lake are the next version of that fight, and nobody has written the clause. Write it now. The hotel’s data lives in a canonical data lake the owner can take with it. Trace logs are the owner’s. Attribution telemetry is the owner’s. Agent portability gets negotiated with key money and termination fees, not after. And fix the plumbing while you’re at it: only 11% of hotel operators have a fully integrated technology stack and 91% still rely on manual reporting. An agent can’t learn from data nobody can find. Buying a hotel - the Hotel Purchase Agreement documentation and process +2

Six questions for anyone pitching you agents

1. Who owns the agent, the prompts and the tools? Can you change its workflow without a vendor ticket?

2. What does it cost per task in tokens, and what am I charged per task? Show me both.

3. Which models does it run on today, and how long does it take to switch?

4. Can I see the trace logs? If the answer is a dashboard, the answer is no.

5. Who, by name, is accountable for this agent’s performance at my hotel? Tom says it tomorrow: he’s watched plenty of launches end with someone asking where the numbers went.

6. If we part ways, what leaves with you and what stays with the hotel?

Here’s my prediction. Within three years, agent ownership and data portability will be standard clauses in every institutional HMA, the way key money and termination fees are today. The owners who ask first get the best terms. The owners who wait will inherit whatever the first ones negotiated.

Hotel management was always a people business. The operator’s job was never the software. It was to employ, train and redeploy the people who do the work. The workers are changing. The job isn’t. An operator that rents its agents has outsourced the one thing you were paying it to do.

Cheers,
Sloan