Insight Engine — statements from the data

It’s 7:55 a.m. at Hotel Peaqplus City. Daniel still remembers the period — it was not so long ago — when the first half hour of the morning went on opening six to eight reports: the pickup report, the pace view, compset rates, the segment breakdown, last year’s same point, the booking window. Numbers in every one of them, and between the numbers it was he who hunted for the story: where did something change that matters?

The picture is different today. On the opened screen, no raw tables are waiting but an insight list — finished statements, in order of importance. The top one reads:

The leisure segment’s booking window has shrunk from 19 to 12 days over the past 30 days — new bookings land on average a week closer to arrival than in the previous period.

Daniel is not looking at a KPI but at a sentence: subject, predicate, time window, reference base. And precisely this difference is the subject of this lesson: the raw metric (“how much?”) and the statement extracted from the data (“what changed, where, since when, how unusual?”) are two different genres — and between the two sits half of an RM’s time. In the previous lesson (51) we mapped the division of labour between the AI and the RM; now we look up close at how raw data becomes a statement, and a statement becomes a decision — because an insight by itself is not yet a decision, only its raw material.

A KPI measures, an insight states

A KPI (key performance indicator) is the result of a measurement: OTB (on the books — the bookings already held) at 62%, ADR (average daily rate) at 98 EUR, yesterday’s pickup at 24 room nights. These are answers to the question “how much?” — and by themselves they are mute. Is 62% good or bad? It depends on which date, how far before arrival, compared to what. It is you who must put the KPI into context, every single morning, at every single number.

An insight, by contrast, is a statement that already contains the context. A usable insight has four mandatory elements — if any is missing, it is not a statement, only noise:

Put any sentence called an “insight” to this test. “Pickup is weak” — no where, no since-when, no against-what: noise. “August is behind” — no baseline and no magnitude: noise. The four-element test is good not only for filtering machine signals — it works on your own sentences in the meeting too. If Adam asks “how are we doing?” and the answer does not pass the test, no analysis has happened yet, only an exchange of impressions.

How a statement is born — baseline, significance, relevance

At the conceptual level, every machine insight goes through the same three steps — and every one of the three can be gotten wrong, which is why it is worth understanding what happens under the bonnet.

1. Baseline selection. The system measures the current pattern against a reference base. That can be last year’s same point (same point — last year’s state at the same number of days before arrival), last year’s fill pattern (how the month filled up last year), a generic fill curve (when there is not enough own historical data), or simply the previous period of equal length (the past 7 days against the 7 before them). A good system also tells you which baseline it used — Peaqplus’s pace view, for example, states whether the target curve comes from last year’s pattern or from the generic curve. This is no technical footnote: you know from lesson 37 that the reading of a booking curve stands or falls on its base.

2. Significance filtering. The deviation has to stand out from the usual fluctuation. A single day’s −2 room nights of pickup deviation is noise; a booking window shortening consistently across 30 days is a signal. The thresholds are typically banded — on Peaqplus’s pace map, for example, a month gets a green (“On track”), amber (“Watch”) or red (“At risk”) status by its deviation from the target curve, and the booking tempo only flips to “Accelerating” or “Slowing” when the change crosses a minimum threshold. The threshold protects the RM from the most expensive error: reacting to every flutter.

3. Action relevance. Even among the significant deviations, only those deserve a signal that can carry a decision. A small deviation 90 days out is not urgent even if statistically real; a worsening trend within 15 days demands attention even while it is small. Lesson 48’s principle echoes here: the booking window is the currency of action.

In Peaqplus this layer lives in the Insight views — each specialised on one core question, serving the raw data in a near-statement form:

And on top of the views sits the Pulse AI layer: a summary written by a large language model (LLM) and a prioritised list of alerts (graded critical / warning / positive) that translates the tiles’ numbers into sentences. How exactly this language layer works and where its limits are is the subject of the next lesson (53) — for now, let’s stay with the statement itself and work through the morning’s example.

The worked example — the shrinking booking window, end to end

The raw data

Behind the insight sits this: the average booking window of pickup filtered to the leisure segment — that is, of the bookings that actually came in new, not the whole book — (the number of days between making the booking and arrival, weighted by room nights) was 12 days over the past 30 days, against 19 days in the period before. An important subtlety: the old method, which averaged the entire OTB, would have shown a misleadingly long window of 80+ days — filtering to fresh pickup gives the true picture of how the guest books now.

The arrival-band distribution confirms the same: of the past 30 days’ leisure pickup, a much larger share of room nights falls into the 0–14-day band before arrival than in the previous period, while the 31–60-day band’s share has shrivelled. So the phenomenon is not the work of one or two outlier bookings but a distribution-level shift. And one more check before thinking anything of it: the price quality. The average rate of the new bookings in the 0–14-day band is not lower than the existing book’s average — so the late demand is not “cheap” demand, but the same guest deciding later.

Why it is significant — and what it rewrites on the demand curve

The −7 days (−37%) is far above the usual monthly fluctuation, and leisure is ~65% of the house’s mix — this is no fringe phenomenon. But the real question is not whether it is unusual; it is what it rewrites. The shrinking of the booking window shifts the booking curve built in lesson 37: the same demand arrives later, so a lower OTB is the “normal” at every early checkpoint.

Daniel rebuilds the leisure demand curve for the comparable Saturdays — next to the old pattern, a new one computed from the fresh booking behaviour:

The key row is T-14: it used to be that 20% of leisure demand arrived in the last two weeks; by the new pattern it is 35%. The mass of demand decided in the last-minute window has grown by more than half as much again.

What it means for the T-21 target — the calculation

Take the nearest affected date: Sep 19 (a Saturday), 21 days before arrival. The leisure final expectation for the day is 60 rooms (in the 80-room house, Saturday is leisure-dominated). The current leisure OTB: 36 rooms.

• Measured with the old curve: the T-21 target is 60 × 70% = 42 rooms → the 36 rooms are −6 rooms, −10 percentage points behind (36/60 = 60% instead of the expected 70%). Amber-red zone, suspected intervention case.
• Measured with the new curve: the T-21 target is 60 × 58% = 34.8 ≈ 35 rooms → the 36 rooms are +1 room, on track.

The same 36 rooms, two opposite diagnoses — the difference is entirely in the baseline. Without the insight, Daniel would have measured against the old curve, seen a “shortfall”, and quite likely pressed the price button. Let’s calculate what that would have cost: a −10% rate cut on the remaining 60 − 36 = 24 leisure rooms, at a 98 EUR ADR, is 24 × 9.8 EUR ≈ 235 EUR per date — and since six such Saturdays run on the same pattern this season, ~1,410 EUR of direct ADR loss, for “solving” a problem that does not exist. Plus the harder-to-quantify price: the cheaper level the market memorises (lesson 44’s rate-bridge discipline).

The decision consequences — in three directions

So the insight’s first achievement was a prevented error. But a good statement implies more than “do nothing” — it rewrites something in three areas:

1. The pricing path. If 35% of demand arrives in the last two weeks, the T-14 → T-0 window is more valuable than before: 60 × 35% = 21 rooms are decided there (instead of the old 12). So the rate-increase steps are worth planning later and higher: Daniel records a conditional +6% step for T-14, for the case that pickup runs on or above the curve — with lesson 50’s trigger logic, in advance, in writing. If the trigger is met, the potential is 21 × 98 EUR × 6% ≈ 123 EUR per date, ~740 EUR across the six Saturdays — a rate increase instead of a rate cut, from the same data.
2. Forecast interpretation. An early pace shortfall now means something different: a larger part of a T-21 or T-28 gap closes on its own than before — at the same time certainty also arrives later, and the variance of early projections grows. A good machine forecast learns this shift over time (lesson 55), but in the transition weeks the RM has to know: the definition of “behind” has changed. This is a direct update to lesson 37’s curve and lesson 50’s diagnosis tree.
3. Promo timing. The early-bird type of campaign that rewards booking early (lesson 46) loses power in this segment — the guest’s decision moment has shifted. The centre of gravity of the visibility tools moves to the last-minute window. One warning here: with last-minute discounting it is easy to teach the market to book late — the shrinking window can be partly cause and partly effect, which is why the closed-group, non-public forms of promotion (lesson 46) are especially warranted here.

And one open question the insight does not answer: why is the window shrinking? A market trend, a side effect of your own earlier promotions, the compset’s behaviour (lesson 44), a change in the composition of demand — or a cause outside the database, like lesson 51’s flight cancellation? That is already diagnosis — and the diagnosis is yours.

The methodology of insight consumption — triage, diagnosis, action

The value of the insight list depends on how you process it. Three steps, in the same order every morning.

1. Triage. Not every statement demands action. Three baskets:

The machine prioritisation (critical / warning / positive) is a good starting point for this, but no substitute: the system does not know that a 30-room group offer came in this very morning for the date flagged “critical”.

2. Diagnosis. An insight is a symptom, not a diagnosis — this is the core of lesson 51’s division of labour, and here it becomes daily practice. The system tells you what changed; whether why, and whether anything should be done about it, is decided by lesson 50’s diagnosis tree: distorted base? segment shortfall? general demand weakness? technical cause? The booking window example shows how easy this is to skip — and how expensive.

3. Action + follow-up. If the diagnosis justifies action: a decision in writing, with the expected effect and a review date (lesson 48’s sixth step). And — this is the expert level’s extra loop — you follow up on the insight itself: if you responded to the shrinking booking window, two weeks later you look for the confirmation in the insight list. Has the picture closed? The action is working. Still deteriorating? The diagnosis was wrong, not the execution.

The three traps

• Insight overload. If you turn every morning signal into a task, the list does not focus you — it scatters you, killing exactly the 15 minutes it was supposed to save. The triage’s first basket rarely holds more than two or three items; if yours holds six every day, your triage is too soft.
• Automatic action without diagnosis. The quality of the statement seduces: if the sentence is finished, the action feels finished too. It is not. From the “leisure window 19 → 12” insight, both a rate cut and a rate increase can follow — in our example it was the latter. The insight is the input of the decision, not the decision.
• Baseline blindness. An insight is exactly as good as its baseline. If last year’s pattern is distorted by a one-off event (lesson 50’s food festival), this year’s “shortfall alarm” is false. If the input data is wrong — a misrecorded segment, a missed data refresh — the statement will be confidently wrong. The “garbage in, garbage out” (GIGO) problem gets to the bottom of things in lesson 57; until then the rule: the first check on a surprising insight is always its own baseline.

The morning 15 minutes — reloaded

Of the five steps of the morning routine built in lesson 50, the insight list condenses the first three — where to look: yesterday’s pickup picture, the critical dates’ baseline deviation and the outlier days arrive as ready signals rather than manual searches. The structure of the routine does not change; its centre of gravity does: less of the 15 minutes goes on searching, more on triage and on starting the diagnosis. What does not change: the slope mindset (the trend says more than the snapshot), the diagnosis tree and the scenario table — those remain your work. The system has taken over the “where to look”; the “what move” stayed with you.

Daniel’s morning closes like this: the booking window insight went into the first basket, the diagnosis ran the same day, and the result was not a rate cut but a refreshed demand curve, a conditional rate-increase trigger at T-14 and a rewritten promo calendar. At the weekly meeting Adam asks about the September Saturdays’ “shortfall” — Daniel’s answer is a single sentence: “The yardstick changed, not the demand: more than a third of leisure now arrives in the last two weeks.” The mark of a good statement is that it can be told to the GM in one breath.

Key takeaways

• A KPI measures, an insight states. The four mandatory elements of a usable insight: what (phenomenon) + where (segment/date range) + since when (time window) + against what (baseline and magnitude). Whatever lacks one of these is not a statement but noise — a test valid for machine and human sentences alike.
• A statement is born in three steps: baseline selection → significance filtering (does it stand out from the usual fluctuation) → action relevance (can a decision attach to it). If you understand the three steps, you also see where the system can go wrong.
• The shrinking booking window rewrites the baseline: if 35% of leisure demand arrives after T-14 instead of the former 20%, the T-21 target drops from 70% of the final to 58% — the same 36-room OTB is a −10 pp alarm by one yardstick and on track by the other. The wrong yardstick would have cost ~1,410 EUR in a needless rate cut in the example; the right reading opened ~740 EUR of rate-increase potential.
• An insight is a symptom, not a diagnosis and not a decision. The processing order: triage (action-demanding / to watch / to note) → diagnosis (lesson 50’s tree) → action with a review date — and a follow-up on the insight itself.
• Three traps: insight overload (a task out of every signal), automatic action without diagnosis, and baseline blindness — the first check on a surprising insight is always its own reference base.