1 hour ago

Put AI on the P&L

Hours saved are a feeling. Money is a line item. Hospitality has been spending on the first and reporting as if it were the second.

A framework for hotel operators to move AI spend from "hours saved" to named P&L lines, using Wyndham's voice AI and industry survey data as anchors.

By Terence Ronson, Founder & Managing Director, Pertlink Limited

Photo by Pertlink Limited

“Where does that live on a P&L sheet?”

That was Hyatt’s Pat Nestor, speaking in the run-up to Skift’s 1 October story, “The P&L Doesn’t Lie.” His full line is worth reading slowly: “We all know what the costs are right now. Certainly in the early days, everything was measured in hours saved. OK, great. That’s helpful in terms of your own personal efficiency. But where does that live on a P&L sheet?”

He then predicted more scrutiny next year, as the conversation moves from adoption to absorption.

He is right, and the rest of us should say it out loud before an owner, a lender, or an investor says it for us. The industry has spent two years asking, “Are we using AI?” The question that will decide budgets in 2027 is shorter and ruder: “Where is it in the numbers?”

What the evidence says — and what it doesn’t

Start with the discomfort. Skift cites a Canary Technologies survey of 404 hotel IT decision-makers across North America, EMEA and Asia-Pacific: among AI users, 37% reported more room revenue and 20% reported lower operating costs. Read that the other way. Roughly six in ten users did not report extra revenue; four in five did not report lower costs.

At the same time, budgets are rising. Skift reports most respondents expect IT budgets to rise at least 10% next year, with more than 5% going directly to AI. Separately, Hotel Dive’s coverage of the same Canary research put the share planning to spend more than 5% of IT budget on AI at 85%.

A separate survey of event attendees, Josiah Mackenzie’s State Of Hotel AI, found marketing was the only function where a majority reported results — and only 12% called those results “strong.” Aperture Hotels CEO Charles Oswald, in a LinkedIn comment Skift quotes: “EBITDA has eroded as the tech stack and its cost have grown.”

Three honest caveats, because this paper should hold itself to the standard it is about to ask of you:

Canary sells hotel technology. The survey is a vendor’s read of its own market, taken in early 2026. The direction is credible. The decimal points are not gospel.
The 12% is a different survey from a different audience. It is not a Canary number and should not be quoted as one.
The survey reports no regional split. It includes Asia-Pacific respondents but publishes nothing that tells an APAC owner how APAC hotels are doing. That absence is itself a finding.

There is also a counternarrative, and a fair paper includes it. J.P. Morgan analysts argued in March that 2026 would be the inflection point where chain AI investment starts to show in earnings, citing roughly 20% productivity gains in Hyatt’s group sales teams and lower franchisee labor costs from Wyndham’s AI call centers. A BCG report, as covered by PhocusWire, put RevPAR gains from AI dynamic pricing at up to 15%, cited Ritz-Carlton San Francisco cleaning rooms 20% faster with AI scheduling and Four Seasons Peninsula Papagayo cutting food waste by 50% in eight months — and found fewer than 10% of hospitality companies getting significant results from advanced AI.

So both things are true. AI can move the P&L. Most hotels have not made it.

The case that did land: Wyndham’s phone

The cleanest proof point from Skift Global Forum is not glamorous. Wyndham CEO Geoff Ballotti described an AI voice concierge, built on Salesforce’s Agentforce, answering calls at economy and midscale hotels. The reported figures:

Roughly 25% of inbound hotel calls were being dropped before AI.
85% of callers stay with the AI agent without asking for a human.
15% higher booking conversion on AI-handled calls.
16% higher ADR when calls transfer to a live agent after the AI interaction.
An average Wyndham property takes about 4,000 calls a year — so, by simple arithmetic on those two reported figures, around 1,000 unanswered calls per property per year.

Notice what makes this work. It sits at a revenue leakage point you can count: a call is either answered, or it isn’t. It has a baseline (the dropped-call rate). It has an owner (the franchisee’s revenue line). And Wyndham credits a half-billion-dollar 2016 program that consolidated dozens of property systems into two, making the integration possible. The plumbing came first.

Two cautions. These are company-reported figures. And the results are specific to economy and midscale; luxury remains untested.

Why “hours saved” never reaches the P&L by itself

An hour saved becomes money in only three ways:

You buy fewer hours. Vacancies left unfilled, overtime cut, agency labor dropped, attrition not replaced.
You sell more. Calls answered, upsells converted, rate improved.
You avoid a cost you would have incurred. Waste, rework, a penalty, a claim.

If an AI project can’t name which of the three it is, the hours are just redistributed — to email, to meetings, to the next tool. This is where the labor line bites. Labor is over 40% of the cost base at a typical hotel (CBRE, via Skift), so any honest cost case runs through it.

The CEO positions at Skift Global Forum showed how unsettled this is. Hilton’s Chris Nassetta: “You’re never taking all the people out.” Accor’s Sébastien Bazin, as Skift reports it, projected that roughly a third of Accor’s network workforce faces displacement and said he is already talking to unions about retraining. Marriott and Cloudbeds have publicly rejected job cuts as the goal. Those are different bets on the same line item, and a board should know which one its management is making.

For owner-operators in Asia-Pacific, the practical version is blunter: if your plan for the saved hours is “the team will do higher-value work,” write down what that work is, who will do it, and what it earns.

The ledger: Pertlink’s P&L Landing Test

Before an AI line item is approved or renewed, it passes five questions. Put them in a one-page ledger, one row per tool.

Question	What a good answer looks like
1. Which P&L line moves?	A named line: rooms revenue, F&B cost, payroll, agency fees, direct-booking share. Not “efficiency.”
2. Which of the three conversions?	Fewer hours bought, more sold, or cost avoided.
3. What is the baseline?	Measured before go-live, over enough weeks to survive seasonality.
4. Who owns the number?	A department head with the line in their budget, not the vendor and not IT.
5. What is the full cost, including oversight?	Licence plus tokens (TCPG) plus the human review time the tool needs.

Then add two columns that most pilots omit: check date and decision — scale, hold, or stop.

Two measurement disciplines, borrowed from the sound parts of vendor-published ROI guidance (treat that source as a vendor view): report lift net of market movement against your comp set rather than gross, and where you can, run the tool on half your volume and keep the other half as a control. A seasonal rate rise is not an AI win.

The cost side has a third column nobody budgets: oversight

Agoda’s AI Developer Report, published 28 September across seven Southeast Asian and Indian markets, is the best APAC data point in this conversation precisely because it is about the cost of trusting AI. 55% of developers now save at least seven hours a week, up from 18% in 2025. 53% run AI agents in production. And 86% review AI output always or most of the time; 79% require human approval before anything reaches production. 28% name cost as a barrier to going further. Agoda’s CTO, Idan Zalzberg: “The real cost is not simply the model call. It is the work required to make AI trustworthy in production.”

That is a technology company that has been at this longer than most hotels, telling you the oversight is the bill. Your ledger needs it as a line. The human approval step is not a failure of automation; it is the Decision Envelope, and the rate at which humans overrule the machine — the AI Override Rate — is a cost and a quality signal at once. A tool whose override rate never falls isn’t saving what the vendor said it would.

The Asia-Pacific lens

Subsidy changes the cost column. From 1 November 2026, Singapore hotels gain a hotel-specific Digital Concierge category under Enterprise Singapore’s EDGE for Productivity scheme, announced at the Singapore Hotel Association’s Hospitality Exchange. Hilton Asia Pacific took the Workforce Transformation Award at the same event. Funding plus workforce redesign is the pairing the Skift survey says hotels are missing. (Confirm eligibility terms with Enterprise Singapore; the detail here comes from trade-press coverage of the event.)

The guest’s wallet is tightening. Per Hotel Online, mainland Chinese travelers on the 13-day Golden Week window spent about ¥911 ($135.70) per trip, a three-year low, even as stays of seven nights or more rose 123%. Only 5% of respondents chose luxury; 33% chose mid-range. Revenue-side AI that depends on a guest trading up will struggle here. Revenue-side AI that fixes leakage — the unanswered inquiry, the unmapped room type — works at any price point.

Your people are watching. In Agoda’s survey, 49% of junior developers feel less secure about their career prospects. Hotel teams will ask the same question about the front desk and the reservations office. The P&L conversation and the people conversation are one conversation; have it once, honestly.

What to do this week

In the next 30 days

List every AI tool you pay for, including the ones buried in existing PMS, RMS and CRM contracts. Put each in the ledger.
For each, answer the five questions. Anything that can’t name a P&L line goes on a watch list.
Find your equivalent of Wyndham’s unanswered call: a place where demand reaches you and you currently lose it. Count it.

In 60 days

Set baselines for the two or three tools that survive. Agree on the owner and the check date.
Add the oversight cost: how many human-hours does each tool require for supervision?

In 90 days

Hold the first scale/hold/stop review. Stop at least one thing. If nothing is ever stopped, the process isn’t working.
Report the ledger to the owner or board quarterly, in the same format as other capital and operating items. Net of market movement.

Five questions a board should ask management this quarter

Which three AI tools have changed a P&L line, and by how much, net of market?
What are we paying in tokens, licenses, and human review for each?
Which tool have we stopped, and why?
What is the plan for saved labor hours — and has anyone told the team?
Where do we have measurable revenue leakage that we haven’t tried to fix with AI?

The bottom line

None of this is an argument against AI. Wyndham’s phones, BCG’s RevPAR range, and Agoda’s engineering data all say it works when it is pointed at something specific and measured honestly.

It is an argument against unaccounted AI: spending justified by enthusiasm, renewed by habit, and defended by anecdote.

Put it on the P&L. Name the line, set the baseline, name the owner, count the oversight, and set the date you’ll check. The tools that survive that will be the ones worth scaling. The ones that don’t were going to show up in the numbers anyway.

Related Pertlink Viewpoint papers

The TCPG (Token Cost Per Guest) series — the cost side of this ledger. TCPG answers “what does AI cost per guest? The P&L Landing Test asks, “What did that cost buy?”
“Spend Wisely” — the Benefit Capture Map and the Efficiency / Cost Saving / Value Creation distinction; the three conversions above are its operational form.
“Flat Is the New Front Desk” and “The Camry Guest Experience” — the AI economics papers on cost structure.
“Forget Me Not” — adoption versus absorption, the same distinction Pat Nestor draws.
“Engineered In, Not Assumed” and “Who Controls the Controls?” — why oversight is a designed cost, not a surprise one; Decision Envelope and AI Override Rate.
“The 7% Problem” and “The Meeting With No Memory” — the team-level and daily-meeting mechanics for running a ledger review.
“The Human Touch Has No Audit Trail” — instinct needs an operating model; so does ROI.

Sources

Made with the help of various SI [AI] tools, but always with a HITL.

Finance Artificial Intelligence Revenue Management AI ROI Labor Costs