The Friction Finder

Somewhere in a corporate office this week, a leadership team is debating how to deploy their AI platform. The platform decision is already made: Microsoft or Google. Except there is now also DeepSeek, and Qwen, and Kimi K3: a frontier-class Chinese model that arrived last month and, according to the industry press, rivals every US model at a fraction of the cost. Grok is said to offer context windows running to millions of tokens. GPT-5.6 leads on reasoning. Muse Spark, built by Meta Superintelligence Labs, is natively multimodal. A new model drops every week and each one is described as the one that changes everything. The question of which LLM implementation partner to bring in is equally unsettled: one of the big global consultancies running an enterprise AI practice, or a boutique specialist who actually understands the industry? And should the company hire a Head of AI at the corporate office? What AI capabilities should sit on property versus at the centre? These are all real questions. They are also, at this moment, the wrong questions. What is often still missing is the answer to an earlier one: which work is actually worth improving?

Organisations are increasingly beginning their AI journey by putting enterprise licences in front of their teams. It is a reasonable first step. The licence, however, is only the beginning. Adoption comes from solving meaningful operational problems, not simply from making powerful tools available. Many quite rightly invest in helping their people become familiar with new AI tools. Familiarity, though, is only part of the equation. Knowing how a tool works is different from knowing which problems are worth solving, and how to apply the answers in a way that creates measurable value.

I learnt this long before AI. Early in my career in luxury hospitality, we ran a guest comment card programme. Cards were collected from the rooms, from the reception desk, and from the daily post delivered to the GM's office, opened by hand, read, and typed into a system, one by one. It was a cottage industry. One day we noticed several cards from different rooms written in exactly the same handwriting. A colleague in concierge had been helpfully filling them in. We were staffing an entire workflow to report on Internal Departmental Guest Satisfaction Scores that were at times fiction.

The instinct back then was to buy a feedback management system. The real questions were more basic. Why are we collecting this? What decision does it feed? Which steps add nothing? Twenty years later the technology has changed beyond recognition but the instinct to delete or modify the email IDs of problematic guests from the survey database has not. Most hotels do not have an AI tool problem. They have an application problem.

Go and watch the work

We called them Gemba days, borrowed from Lean. You go to where the work happens, you observe, and you do not interrupt. You follow one enquiry, one invoice, one room turnaround, from beginning to end. It is the least glamorous process improvement technique in existence and it is still the most valuable thing I do with any hotel.

What you find is rarely a technology gap. Across the hotels I have worked with, the same patterns appear consistently. A pre-arrival team assembles VIP profiles every morning by pulling information from three separate systems into a briefing document that nobody has formally requested in over a year, but which everyone assumes someone else is acting on. A finance team spends the last four working days of every month in a close process where perhaps one day is genuine reconciliation and three are spent chasing numbers from operators who submit them in different formats each time. A banquet team builds proposals by copying previous proposals, then manually stripping and replacing every client-specific detail: four hours with no creative input at all. Watching someone work at their desk for five minutes often tells you more than a month of process interviews. A screen crowded with contract folders named 'Final', 'Final v2' and 'Final SEND THIS ONE', with no version control and no shared structure, is a reliable signal of how a department actually operates. Engineering teams log work orders on paper, photograph them, and email the photographs to a coordinator who retypes them into a system. The system has a mobile interface that nobody was trained on. In almost every case, the people doing the work have known about these patterns for years and have a clear view of what they would change. The question had never been asked.

Some of this does not need AI. It needs someone to ask why the step exists at all. Automating a process nobody has questioned simply produces the same confusion at a higher speed.

There is a layer of operational chaos that rarely appears in any process map: the hotel WhatsApp group. Most properties now run dozens of them: one for each department, several more for each shift, and a growing cluster of subgroups for specific issues, specific guests, specific weeks. Information moves fast and vanishes just as quickly, buried under three hundred messages before anyone has had the chance to act. Decisions made at 11pm exist only in a thread nobody can find the next morning. Colleagues on leave are pinged without apology because the group has no off switch. The volume generates the illusion of coordination while quietly replacing the structures that make coordination real. No AI tool fixes this. What fixes it is a clear decision about what information belongs where, and the discipline to hold that line.

Five questions before any technology conversation

When we run this exercise inside hotels, five questions do most of the work.

1. Where does leadership time actually go? Not where people say it goes. Follow a leader through one real day and write down every task that required genuine judgment and every task that did not. The ratio is almost always a surprise. Management time consumed by administration (chasing approvals, consolidating reports, answering messages that should not have reached a senior leader at all) is time that cannot be spent coaching, deciding and improving. The map will tell you more than any survey.

2. What decision does your data actually feed? Not what it describes: what it feeds. Most hotel operations generate considerable data and act on very little of it. A daily report that takes an hour to compile and is read by nobody is not an information system. It is a ritual. Before automating any reporting, identify the specific decision each number is meant to inform, who makes that decision, and how often. If no decision changes as a result of a report, the report should not exist.

3. Where is your organisation's AI energy going? Energy and curiosity about AI are not the same as shared direction or practical application. Many hotel teams have individuals experimenting with AI tools on their own time, producing results that impress in a demonstration and then disappear. Scattered curiosity without a shared framework produces noise, not capability. The question is not whether your team is interested. It is whether that interest is being channelled toward problems that matter.

4. Which improvement have you made more than once? If a process was redesigned and then drifted back, the problem was not the process: it was the system around it. Execution consistency fails when an improvement is introduced without redesigning the incentives, handovers and accountabilities that surround it. The initiative works while attention is on it and stops working when attention moves on. Before adding any new technology, ask which improvements are currently holding and which have already been made twice.

5. How far is it from a good idea to a working process? Most hotel leadership teams can generate strong ideas in a workshop. Far fewer can point to a clear, documented path from that idea to a changed way of working, held consistently across shifts. The gap between insight and embedded operational reality is where most improvement effort is lost. Closing it is not a technology problem. It is a system design problem, and it must be solved before any AI tool is introduced into the workflow.

Only after these five questions does the tool conversation become useful, and by then it is usually a short conversation. Most hotels already pay for an enterprise-grade assistant their teams have barely opened. Start there, inside your approved environment, on work that touches no sensitive data.

What is coming makes this more urgent, not less

Earlier this year we ran an experiment to see how far an AI agent could get booking a hotel room on its own. It searched, compared and navigated impressively, and then hit a wall: not at payment, but at the point where systems would not expose rates to a machine. That wall is already coming down. In August, Google confirmed that AI-agent hotel booking is in live testing in the United States, with Amadeus as technical partner and Booking Holdings, Expedia, Marriott, Wyndham and IHG involved. Sabre, PayPal and MindTrip are building an end-to-end agentic booking pipeline of their own. The operational questions it raises, about approval limits, data quality and audit trails, will land on hotels that have not tidied the basics.

This is why the friction finding matters. If your processes are undocumented and your data disagrees with itself, autonomy will amplify the mess. If the foundations are clean, you can extend real authority to these systems, carefully and within limits, and the gains compound.

AI will not transform hospitality. The leaders who know how to use it will.

Every hotel has one process that quietly consumes hundreds of management hours each year.