AI and RM — what the machine sees and what it doesn't

It’s 8 a.m. at Hotel Peaqplus City. The coffee is the same, the routine is the same — the 15-minute morning pickup review that closed the advanced level (lesson 50). Daniel opens the Insights screen, and before starting his own checklist, his eye catches on one sentence of the system’s morning summary: leisure pickup for Tuesday–Wednesday arrival dates has been slowing week over week for three weeks — the shortfall is growing.

The first reaction is suspicion — a healthy reflex; keep it. Daniel opens the pickup view’s segment and day-type breakdown, and the number checks out: the pace of incoming leisure bookings for Tuesday–Wednesday arrival dates has fallen for three consecutive weeks. And here comes the uncomfortable realisation: his own routine would not have caught this — not for weeks. Lesson 50’s morning routine looks at the highlighted dates: weekends, event days, known weak spots. A slide that narrows to one day type and one segment, deepening by a few room nights a week, does not stand out on any highlighted date — at date level it is 1-2 percentage points everywhere, which looks like noise. The machine, however, does not watch highlighted dates; it watches everything, every morning, with the same attention — and it flags the pattern that emerges from many small pieces of noise.

Two days later the other half of the story arrives — not from the database but at the city hoteliers’ breakfast, in a half-sentence: the low-cost carrier has cancelled its Tuesday-to-Thursday rotation from the hotel’s biggest foreign feeder city, effective September. The midweek city-break guest physically cannot arrive on a Tuesday, because there is no plane. This information is in no booking database anywhere — the machine saw the symptom; it could not see the cause.

This pair of scenes opens the expert level, and it is also the level’s whole programme in a nutshell. The machine saw something in the data that Daniel had missed. Daniel knows something about the market that the machine never will. This lesson draws that boundary line: what AI (artificial intelligence) adds to RM work, where its systematic blind spots are, and what the division of labour between the two looks like — the one all the following lessons build on.

What the machine sees in the data

Two machine layers work in the modern RM toolkit. One is ML-based (machine learning) pattern recognition and computation: relationships learned from historical data, forecasting, anomaly detection. The other is the LLM layer (large language model): the readable narrative written from the numbers, the natural-language question and answer. Together they make up what you experience as “the AI” in daily work. What is this pair strong at?

1. Coverage. Size up the monitoring task: the next 365 days × 6–8 segments × at least 3 metrics (OTB, pickup, rate) — close to ten thousand cells, and something can happen in every one of them. The 15-minute morning routine covers the critical 30–40 of them — that is what a human can do, and that is fine. The machine watches all of them. The Tuesday–Wednesday leisure slide lived precisely in the range human attention would never have reached.
2. Tireless consistency. The machine has no bad days, does not get bored by the four-hundredth check, does not relax after a good month, and flags with the same threshold in August as in November. The greatest enemy of human attention — monotony — does not exist for it.
3. The small, cumulative deviation. A dramatic break (a group cancellation, a dead channel) a human notices too. The slide that deepens by a few room nights a week, smeared across many dates, they do not — at date level such a trend is noise; aggregated, it is already a story. The machine looks exactly at the aggregate pattern level.
4. Calculation speed. A displacement calculation (lesson 40), a pace projection, a segment-level comparison — what takes half an hour to assemble by hand, the machine returns on request, in seconds.

In Peaqplus this layer is part of the daily work. The Insight Engine is the dashboard side: pickup, pace, performance and plan-tracking views, with segment breakdowns, automatic flags on outlier deviations and a morning text summary — this is what Daniel’s eye caught on. Pulse Chat is the conversational layer: you ask in natural language (“how did leisure pickup develop over the past 30 days?”, “which segment is falling?”) and the system answers from the real report data, with tables and charts — so the diagnosis tree’s check steps (lesson 50) run in minutes. A proven check question can also be turned into a routine: it runs on a schedule, and the result arrives as a notification.

But notice: every item on the list above lives inside the database. The machine works from the hotel’s booking past and present — from what has ever become a number. And this is where the other half of the lesson begins.

What the human sees in the market

The flight cancellation is not in the booking data. Nor is the construction starting next week on the neighbouring plot, the rumour of the competitor hotel’s renovation, the embassy event the sales colleague heard about, or the fact that the loyal partner of ten years has changed owners. A critical share of RM work’s inputs never becomes data — it lives in conversations, newsletters, the receptionist’s half-sentence.

The human remains unbeatable at four things:

• Context and causality. The machine sees correlation (“Tuesday–Wednesday pickup is falling”); the cause (“there is no flight”) is found by the human — typically from a source outside the data. And the right action always follows from the cause, not the symptom: a rate cut would be a meaningless answer to a cancelled flight route.
• The one-off and the new. Where there is no historical example, the machine has no pattern — the human, however, has analogy, industry experience, judgement.
• Relationships. The loyal account, the group organiser, the GM of the hotel next door — the value of relationships with market players does not fit into the revenue optimisation of a single date.
• Responsibility. A human answers for the decision — to the owner, the team, the guest. This is not a sentimental point but a practical one: whoever cannot justify a decision cannot defend it, correct it, or learn from it. “The machine said so” is not a justification.

The AI’s three typical failure modes

The “machine in the data, human in the market” boundary becomes sharp in three typical situations. All three are worth knowing by name, because at the AI tools of the coming lessons, these are what mark out, again and again, where you must take the wheel.

1. Rare and unprecedented events

The essence of machine learning is that it generalises from the patterns of the past. Where there is no past example, there is no pattern — the model still outputs a number, just by the logic of the old world. The pandemic is the extreme example: in the spring of 2020, every history-based forecast was systematically wrong for months, because the world it had learned from had ceased to exist. But the small, local variant is more common: a new 200-room competitor opens in the district — the hotel’s historical data contains zero information about how demand will redistribute. The model confidently extrapolates last year, and it is weakest exactly when it would be needed most. This is why modern forecasting is built on hybrid logic (lesson 38, and lesson 55 at expert level): a human correction layer next to the machine projection, precisely for such breaks.

2. Chain reactions and regime change

A relative of the first failure mode, but more insidious: here there is historical data — it is just that the relationships have lost their validity. The airline pullout is like this: for years the model learned how Tuesday–Wednesday demand behaves, what its price sensitivity is, how it responds to a promotion. From September 1 those learned relationships are partly dead — but the model does not know that, and keeps calculating by the old regime. That breakpoint — “from now on the world works differently” — only the human can call, because only the human knows about the cause. The practical consequence: when a regime change is suspected, machine recommendations must temporarily go under stricter human control, until the new pattern builds up in the data.

3. Contextual decisions

The third failure mode is not missing data but an objective-function limit: the machine optimises what we have quantified for it — every other value does not exist for it. The displacement calculation (lesson 40) is the textbook example. Suppose the tour-operator partner of ten years asks for 30 rooms for 2 nights on a strong October weekend, at a contracted 78 EUR. The machine calculates: the group’s revenue is 30 × 2 × 78 = 4,680 EUR; the displaced transient demand is 30 × 2 × 95 = 5,700 EUR; the difference is 1,020 EUR in favour of declining. Mathematically flawless. Except the objective function does not contain the fact that this partner brings ~400 room nights a year of predictable base load, from December to March, when transient demand displaces nothing; that it also generates banquet revenue; and that after a rejection, next year’s allocation negotiation starts from a very different place. The strategic relationship, the brand position, the ethical line (“do we sell at five times the rate in a force-majeure situation?”) — these are contextual decisions, and they always will be. The machine outputs the number; whether the number is the decision is decided by the human.

The worked example — what the machine’s signal is worth, in numbers

Back to the Tuesdays and Wednesdays — let’s now calculate what the machine found and what the early signal is worth. The pickup view’s segment and day-type breakdown showed this (weekly leisure pickup for Tuesday–Wednesday arrival dates, with the previous 8 weeks’ average as the baseline):

(The deviation is the ratio to the baseline: 36/42 ≈ 86%, i.e. −14%; 31/42 ≈ 74%, i.e. −26%; 26/42 ≈ 62%, i.e. −38%.)

Three numbers are worth extracting:

1. The shortfall already accumulated: (42 − 36) + (42 − 31) + (42 − 26) = 6 + 11 + 16 = 33 room nights in three weeks. At the leisure ADR (average daily rate, here 98 EUR), 33 × 98 = 3,234 EUR of lost pickup — this much had already happened by the time the signal was born.
2. The forward-looking risk: if the trend stabilises at last week’s level, the weekly gap is 42 − 26 = 16 room nights. Over the 8 affected weeks of the autumn booking window: 16 × 8 = 128 room nights × 98 EUR = 12,544 EUR at risk if nothing happens.
3. The value of the early signal: as a date-level same-point shortfall, this slide would realistically have become visible about 4 weeks later in the manual routine — by then a further ~4 × 16 × 98 = 6,272 EUR of shortfall accumulates, and the reaction window is that much shorter. That is the machine layer’s added value in this one case — before any decision was made at all.

And now the other side: this is as far as the machine got. Managing the 12,544 EUR risk takes a diagnosis, and the diagnosis — there is no Tuesday-to-Thursday flight — came from outside the database. Without it, the obvious reflex (a rate cut on the weak days) would have burned money: the demand did not stay away out of price sensitivity, but because it physically cannot arrive. Offering an unreachable product at a discount — that is the textbook case of treating the symptom.

The division-of-labour model

The case generalises, and this is the backbone of the lesson — and of the whole expert level. Let’s state it explicitly:

The failures of the two columns are complementary — that is the point. The machine is weak where the human is strong, and vice versa. Two practical rules follow. One: the AI signal is the input of the routine, not its output — the diagnosis tree (lesson 50) still runs unchanged; you just get the “where to look” step ready-made, earlier and with fuller coverage. Two: never accept a machine recommendation without a justification, and never discard one without a justification — in both directions, the “why” is your work. Whoever follows the machine blindly pays at the first and the third failure mode; whoever ignores it blindly throws away the coverage advantage the tool was brought in for.

Returning to the Tuesdays and Wednesdays

Daniel’s decision sequence was built like this. Rate cut: no — the demand is not price-sensitive; it is physically constrained. Instead, redirecting the demand: reweighting the midweek packages to Monday and Thursday arrivals (when there is a flight), with 2-3-night structures, so the Tuesday and Wednesday nights fill from the neighbouring arrival days — classic day-by-day thinking (lesson 48). In parallel, the midweek campaign turns from the lost feeder city towards markets within rail and driving distance. Daniel manually corrects the forecast’s Tuesday–Wednesday expectations downwards (lesson 38 — the human of the hybrid logic exists precisely for such breaks), so the coming months’ plan numbers do not measure against a dead regime.

And one step that is the new routine of the AI era: Daniel teaches back to the machine what he saw in the market — he records the flight cancellation in the hotel’s AI context (the special circumstances). From then on, the system’s analyses and summaries read the data together with this knowledge, and do not flag the same thing again and again as an “inexplicable weakness”. The boundary between the machine’s terrain and the human’s is not a static wall: what is only in your head today is — if you record it — the machine’s working context tomorrow. An AI is not only to be used; it must also be fed.

The late-October balance: Tuesday–Wednesday occupancy recovered to ~85% of its earlier level from the redirected arrivals — the larger part of the 12,544 EUR risk never materialised, without a rate cut. Neither the machine nor Daniel would have solved it alone: without the machine’s signal, Daniel starts weeks later with a smaller window; without Daniel’s market knowledge, the machine’s signal would have run into a rate cut.

The map of the expert level

The level builds on from here. The next five lessons take the pieces of the AI toolkit one by one — each with the same pair of questions: what do you entrust to it, and where do you take over? Lesson 52 starts with the Insight Engine (statements from the data — how to read, verify and hold machine signals to account), lesson 53 continues with the AI narrative (the power and traps of text written from numbers), lesson 54 with Pulse Chat’s questioning method (asking as an analytical tool), lesson 55 with Smart Forecast Enhanced’s hybrid model, and lesson 56 with the Pricing Engine’s ML-based rate recommendations. Then come the questions of operating the system: data quality (lesson 57 — GIGO: the machine is only as good as the data it works from), experimentation and A/B testing (lesson 61), the life story of decisions (lesson 64) — and at the end of the level, an outlook on the future of RM (lesson 66 — AI agents). The division-of-labour model — filter-signal-calculator versus diagnostician-decision-maker-responsible — will echo through all of them.

Key takeaways

• The machine sees in the data, the human in the market. The machine’s strengths are coverage, tireless consistency, catching small cumulative deviations, and calculation speed — but all of it lives inside the database. What never became a number (causes, news, relationships) remains the human’s terrain.
• The AI’s three systematic failure modes: (1) rare, unprecedented events — no pattern, the model extrapolates the old world; (2) chain reactions and regime change — there is data, but the learned relationships have lost their validity; (3) contextual decisions — values outside the objective function (relationships, brand, ethics) do not exist for the machine.
• The division of labour: the AI is filter + signal + calculator (“where to look”, “how much”), the RM is diagnostician + decision-maker + responsible owner (“why”, “what move”). The machine’s signal is the input of the routine, not its output — the diagnosis tree remains your work, unchanged.
• Neither follow the machine’s recommendation blindly nor ignore it blindly — both directions require a justification. Blind following pays at the failure modes; blind ignoring throws away the coverage advantage.
• An AI must be fed, not just used: record the knowledge you gain in the market (regime change, one-off events, context) in the system’s hotel context — that way the machine’s analyses read the data together with your market knowledge. In the worked example, the early signal alone prevented a further ~6,272 EUR of shortfall, and finding the cause prevented a money-burning rate cut.