AI narrative and human-readable reports

Sunday evening, Hotel Peaqplus City. Daniel sits at his laptop assembling the Monday-morning report pack: weekly performance by segment, pickup tables (pickup is the net balance of new bookings received), occupancy curves, the ADR trend (ADR — average daily rate), the budget comparison. Forty pages. He is proud of it — everything is in there.

On Monday morning Adam, the GM, calls him in for a coffee. “Daniel, I have a confession to make. Of your weekly pack, I read the first page — plus whatever you add in person. The ownership group asked about June on Friday, and I realised that out of the forty pages I could not have put three sentences together for them. I don’t need less data. I need someone to tell me what it means.”

That sentence is one of the industry’s most honest moments. Report packs grow to forty pages because they want to answer every question — and that is exactly why they answer none. The reader — the GM, the owner, often the RM themselves — does not want data; they want meaning: what happened, why, and where a decision is needed. In lesson 52 we saw how the machine layer finds the outlier deviations; this lesson is about how numbers become five sentences that say what a hundred charts mean — what makes such five sentences good, and what makes a bad set dangerous.

Why more data is not the answer

The classic report pack’s logic: “give them everything, the reader will pick what they need.” It fails on three points:

1. The selection is the expertise. Deciding which three numbers out of forty pages matter this week takes revenue knowledge — precisely what the reader turns to the RM for. If we leave the filtering to them, we have shifted the work, not the information.
2. An unread report is false reassurance. Adam “received” the weekly picture every Monday — in reality he had not known for weeks that the corporate segment was slipping. Data that is sent out but never processed is worse than nothing, because everyone believes they are informed.
3. A number has no direction on its own. A 68% occupancy is neither good nor bad — it becomes one against the plan, against last year, against the market. The raw report leaves the comparison to the reader; the narrative does it for them.

So the direction of the solution is not a prettier chart but the narrative summary: a short text built from claims, which turns numbers into meaning. A machine layer can write it now — in Peaqplus, the AI Report Narrative available on every report with one click is exactly this —, but the fact that a large language model (LLM) does the writing changes none of the quality criteria: they are the same as if you wrote it. More than that: you need to know what makes a data narrative good precisely so you can judge the machine’s.

The anatomy of a good data narrative

A good narrative summary is built from four elements — and each of the four has its own discipline.

Element	What it does	Discipline rule
1. Claim with a number	Every sentence is a verifiable claim with a concrete number: “weekend RevPAR is +12% year-on-year”	A claim without a number (“it was a weaker week”) is not a claim but a mood — it cannot be checked, cannot be debated, cannot be decided on
2. Cause attribution	The “why” — but only where the data supports it	Correlation is not causation: a causal link between two co-moving numbers needs separate proof. Without it, the correct form is “at the same time”, not “because of”
3. Materiality filtering	What makes it into the five sentences — and what stays out	Editing is itself the decision: what goes in gets attention; what we stay silent about, we implicitly declare fine. Omission is a responsibility, not a convenience
4. Action direction	”Worth looking at X” — the narrative ends by pointing at the next step	A direction, not an instruction: the narrative does not decide for the reader; it says where a decision is needed

(RevPAR — revenue per available room.) The order of the four elements is no accident either: claim → cause → weighting → direction. It is the same arc as the revenue meeting’s decision frame from lesson 47 (data → diagnosis → options → decision) — only here it runs in five sentences, not in a meeting.

A good and a bad narrative — for the same week

Before we get into the numbers, look at the style of the two extremes. About the same week (we will derive the detailed numbers in a moment), the bad version:

“Last week showed a mixed picture overall. The city break segment underperformed, probably due to market uncertainty. Corporate is also facing challenges, which can be explained by industry trends. At the same time, there were positives. The trends are worth monitoring.”

Look at what it does: there is not a single number in it (nothing is verifiable), the cause attribution is guesswork (“probably due to market uncertainty” — supported by what?), the materiality filtering is missing (“there were positives” means everything and nothing), and the action direction is empty (“worth monitoring” — what, where, until when?). Its most dangerous property is that it sounds good: fluent, balanced, professional. A busy GM reads it, nods — and knows nothing more than before. Less, in fact: now they believe they know something.

The good version — that will be the output of our worked example: five sentences, each claim backed by a concrete cell in the table. Let’s build it.

The worked example: one week, one table, five sentences

Hotel Peaqplus City (80 rooms), the closed actuals of the week of Jun 15–21, room revenue by segment — plan (budget), actual and last year’s actual side by side:

Segment	Budget (EUR)	Actual (EUR)	vs budget	Last year (EUR)	YoY
City break (leisure transient)	22,000	18,040	−18.0%	17,500	+3.1%
Corporate	9,500	7,600	−20.0%	9,200	−17.4%
Group	6,000	6,900	+15.0%	5,800	+19.0%
Other (direct, walk-in)	2,500	2,460	−1.6%	2,400	+2.5%
Total	40,000	35,000	−12.5%	34,900	+0.3%

A quick check — exactly the move this lesson teaches: the segment rows sum to 18,040 + 7,600 + 6,900 + 2,460 = 35,000, matching the Total row; and the lagging segments’ gross shortfall (3,960 + 1,900 + 40 = 5,900 EUR), offset by group’s +900 EUR surplus, gives the 5,000 EUR house-level gap (40,000 − 35,000).

At first glance the table suggests a simple story: “everything is slipping, only group is fine.” But look at the two reference points together — and the picture splits. City break is 18% below plan, yet growing against last year (+3.1%): here demand did not fall; the plan was ambitious. Corporate, by contrast, is behind both reference points (−20% and −17.4%): that is genuine weakening — market-level or account-level. Shortfalls of the same magnitude, two radically different diagnoses — and two different actions. This separation is the heart of narrative writing.

The five sentences, as Daniel (or the machine summary) writes them:

(1) In the week of Jun 15–21, room revenue was 35,000 EUR — 12.5% below budget, but essentially level with last year (+0.3%). (2) The city break segment came in 18% below budget while growing 3% year-on-year — the shortfall here comes from the plan’s ambition, not from falling demand. (3) The corporate softening, by contrast, is real: −20% to budget and −17% to last year’s actual — the only segment behind both reference points. (4) Group delivered 15% (900 EUR) above budget, offsetting 900 of the other segments’ 5,900 EUR shortfall. (5) Worth investigating: the account-level breakdown of the corporate drop — did a key account fall away, or is it spread wide — and the city break summer budget curve, because if the coming weeks’ plan is just as tight, the gap will reproduce itself week after week.

Now let’s trace, sentence by sentence, which number holds which claim:

Sentence	Anatomy element	Supporting number from the table
(1) House-level picture: −12.5% to plan, +0.3% YoY	Claim with a number + double comparison	Total row: 35,000 vs 40,000 vs 34,900
(2) City break −18% to budget, +3.1% YoY → the plan is ambitious	Claim + grounded causal reading	18,040 vs 22,000 (−18%) AND 18,040 vs 17,500 (+3.1%) — the causal claim is licensed by the two numbers together
(3) Corporate −20% / −17.4% → genuine weakening	Claim + contrast with (2)	7,600 vs 9,500 and 7,600 vs 9,200
(4) Group’s +900 EUR offsets part of the 5,900 EUR shortfall	Materiality filtering: the positive is sized, not merely mentioned	6,900 − 6,000 = +900; 3,960 + 1,900 + 40 = 5,900
(5) Two investigation directions: account breakdown + budget curve	Action direction	Follows from the diagnoses in (2) and (3) — not a new number but the consequence of the existing ones

And notice what stayed out: the “Other” segment (−1.6% — noise level; mentioning it would only dilute the point), the daily breakdown, the ADR details. Materiality filtering is at work here: the five sentences are not a summary of the table but the table’s decision-relevant reading. Whoever wants the detail has the table — the narrative is an entry point, not a replacement.

One important contrast for sentence (2): if someone wrote “city break came in 18% below budget, mainly because of the corporate slowdown” — that would be a category error. Corporate is a different segment; its softening worsens the house-level picture but explains nothing about city break’s own shortfall. The two claims deserve separate sentences with separate causes — exactly as the narrative above did it. The discipline of cause attribution turns on details like this: a single badly attached “because” derails the whole summary.

How the machine narrative is born — and what it means for you

Conceptually, the machine narrative has two layers, and separating them is the key to using it correctly:

1. The facts layer — from the system. The report data (structures like the table above) and the insight flags we met in lesson 52 are deterministic calculations: from the same data, the system always returns the same number. The segment rows, the deviations, the reference points all come from here.
2. The language layer — from the model. The large language model writes from these structured facts: it selects, weighs, connects, forms sentences. The numbers come from the system; the language comes from the model.

In Peaqplus this division of labour is tangible. The AI Report Narrative is available on every report: one click, and the system hands the report’s cleaned numbers — with the hotel and property names removed — to the model, which writes a 2-3 paragraph summary: the trends, the outliers, and what is worth looking into. It responds in your language, and the analysis is bound to the view and the data state it was written on: the same report with the same filters carries the same summary — but after a fresh data upload it is rewritten, because the facts layer has changed.

From this two-layer structure follows the basic rule of user discipline: the number behind the narrative is always checkable — and before a critical decision, it must be checked. In practice this is convenient: the summary appears in the same place as the report itself — the cell a claim is built on is one scroll away. If the analysis says “corporate is −20%”, the right row of the table confirms or refutes it in two seconds. And if the summary leaves a question open, you can ask the next “why?” in Pulse Chat, precisely aimed — that is what lesson 54 is about.

Three readers, three narratives

The same June week of data — but three readers, with three questions:

Reader	Their question	What their narrative contains
RM (Daniel)	“Where do I need to intervene?”	Operational detail: segment breakdown, booking window context, compset flag (lesson 44), specific dates — the five sentences above + the investigation list
GM (Adam)	“Do I need to decide anything?”	Decision points: what is on plan, what is not, what the RM proposes, what needs GM approval — no jargon
Owner	”Do I need to worry?”	Trend + money: the month-level picture, plan attainment, the look ahead — no terminology, with a status signal

The owner layer has the strictest language discipline: not “pickup is accelerating” but “bookings are coming in at a good pace”; not “the ADR index is 106” but “our room rates are above the market average”. This is not dumbing down — it is the same truth in the reader’s vocabulary. In Peaqplus, two built-in surfaces serve this layer. The Executive Summary shows nine key KPIs with green–amber–red status — each with its own threshold, measured against budget, last year’s same point or the market —, no terminology on the tiles, and it also arrives by email on Monday morning, readable without logging in. The morning Daily Briefing carries the same discipline into a daily rhythm: short written blocks about yesterday and the days ahead, including a few AI-written flags — anomaly warnings and the day’s top actions. In both, materiality filtering is a built-in rule: few claims, with numbers — not forty pages.

At the end of the coffee, Daniel shows Adam the summary that sits on top of the weekly report. Adam reads the five sentences and stops at the third: “That’s it. Corporate is what I want to know about. The city break plan story is interesting, but that’s your desk — what I need is which account we are slipping on, and whether I should be calling one of their managing directors.” Notice what happened: Adam did not ask for the data; he was looking for his own decision point in it — and he found it, because the narrative was built from claims, not moods. The forty pages never achieved that in thirty weeks.

Traps — what makes a narrative dangerous

The narrative summary’s power is also its risk: whoever writes well gets believed. Four traps, each with a defensive routine:

1. Hallucination risk. A language model writes confidently and fluently even when a claim has no data (or misread data) behind it — linguistic quality is no proof of factual accuracy. Defence: the number-checking routine. At low stakes, a spot check is enough (tracing one claim of the five back to the table); before a pricing decision, an owner report or partner communication, check every numerical claim against the source report. The narrative is an entry point, not a final truth.
2. Narrative laziness. If the summary is accurate for weeks, the team unlearns looking behind the numbers — and the one time it errs (or merely highlights something other than what is currently critical), nobody notices. Defence: at the weekly revenue meeting (lessons 28 and 47) the narrative is a starting point, not an agenda replacement — every week, someone traces at least one of its claims back to the source number, out loud. It takes two minutes and keeps the culture of verification alive.
3. Over-attribution of causes. The language tempts: “while” slides easily into “because”, and machine text is also prone to attaching a cause where there is only co-movement. Defence: at every “because”, ask — what data rules out the other explanations? In the example above, the city break causal reading was licensed by two reference points pointing the same way together; a single number is never enough for causality.
4. The “says everything, highlights nothing” report. The ten-to-fifteen-sentence “narrative” listing every segment and every metric reproduces the forty-page problem in prose. If everything is important, nothing is. Defence: the five-sentence ceiling as an editing constraint — and a conscious list of what stays out (“what we are deliberately staying silent about right now, because it is fine”).

Notice the common pattern: all four defences are human routines. The machine narrative takes the mechanics of report writing off your shoulders — not the judgement. The division of labour laid down in lesson 51 holds here too: the machine scales the work; you carry the responsibility.

Back to Sunday evening

Three weeks later, Daniel’s Sunday evening looks different: instead of editing forty pages, he spends fifteen minutes checking and sharpening the weekly narrative — re-verifying the critical numbers, striking an over-written cause, adding an action direction only he can know (the status of Wednesday’s group offer). The ownership group actually read the monthly report for the first time — because it was five sentences, with numbers, ending with what to look into. And Adam, based on the corporate account breakdown, called the managing directors of the two largest accounts — it turned out one has a quarterly travel freeze, the other merely rescheduled.

The forty pages did not disappear — they sit behind the narrative, and when a question comes up, that is where everyone reaches. Only now someone says what it means — and someone checks that it is telling the truth.

Key takeaways

• A report pack is not communication — forty unread pages are false reassurance. The narrative summary turns numbers into meaning: five sentences that say what a hundred charts mean.
• The four elements of a good data narrative: a claim with a number (verifiable), cause attribution only with data support (correlation is not causation), materiality filtering (omission is also a decision) and an action direction (where to decide, not what).
• The double comparison is the key to diagnosis: in the example, city break is −18% to plan but +3.1% to last year — an ambitious plan; corporate is −20% / −17.4% — genuine weakening. The same size of “shortfall”, two different stories, two different actions.
• The machine narrative has two layers: the numbers come from the system (deterministic facts), the language from the model. From this follows the discipline: the narrative is an entry point — the number behind it is always checkable, and before a critical decision it must be checked.
• The same data, three narratives: the RM gets operational detail, the GM decision points, the owner trend and money — in the reader’s vocabulary. Against the four main traps (hallucination, narrative laziness, cause over-attribution, “says everything, highlights nothing”) the defence is human routine, not a technical setting.