What Are AI Summaries?

An AI Summary is a chart option in Insights that reads the results table on that chart and writes a short, plain-language interpretation of it — the headline trend, the main driver, and anything that stands out. It’s there to speed up how quickly you can read a table, not to replace your own analysis.

If you’re new to Insights’ AI features, start with How AI in Insights Helps You Work Smarter and Faster for a complete overview.

What a Summary Gives You

A summary describes what is in the table in front of it. A typical one covers:

The headline — the main change or pattern
A short explanation of what’s behind it, based on the columns you’ve included
Any anomaly that’s visible in the data

It will only suggest next steps or actions if you ask for them and the table gives it enough to say something specific. On a single table it usually can’t, so a forced “recommendations” section tends to come back generic. If you want real recommendations, take the question to the full AI chat.

💡 Tip: Use AI Summaries to read faster, then apply your own judgement. The summary is a starting point, not a decision.

Before You Write a Prompt: How the AI Summary Actually Works

The AI Summary does one thing, once: it reads the table on this chart and writes a single interpretation of it. Four things follow from that, and knowing them up front will save you a lot of back-and-forth.

1. It’s a single pass, not a multi-step analysis

The AI reads your table one time and produces one answer. It does not:

Reach a conclusion and then go back to dig deeper into it
Run follow-up queries or pull in other charts, tables, or time periods
Work through a numbered list of analytical tasks in order

A prompt like “analyse X, then based on what you find investigate Y, then diagnose the cause of any Z below 100” is a multi-step request. The summary will attempt a shallow version of the first part and drop the rest.

2. Shorter prompts give more accurate summaries

There is a direct trade-off: the more sections, rules, and conditions you add, the more the AI flattens or half-answers them. A prompt that names one or two things to look at will almost always beat a prompt with six required sections.

3. It can’t match a fixed layout

You can steer the shape lightly — “a short paragraph,” “bullet points,” “add a one-line note on anomalies” — and it will broadly follow that. It won’t reproduce a precise template: exact headings, blank rows between sections, a table with these columns in this order, an “Index Growth/Decline” line under every metric. If a report needs to look a specific way, build that shape into the table and the dashboard tiles, and let the summary describe what’s there.

4. It needs real numbers to work with

Summaries are strongest on tables with populated values and a clear pattern. Sparse data, lots of zero rows, or a brand-new property with no prior-year comparison will produce thin, awkward, or “blank %” output — there simply isn’t enough for it to describe.

If you need a full multi-section report

A single AI Summary tile isn’t the tool for a monthly one-pager that covers overall performance, a channel and segment breakdown, and comp-set / STR positioning — or an executive review with a KPI table, a weekday/weekend split, and a wins-and-opportunities section. For that, either:

Build it as several tiles on one dashboard — one query and one short summary per section, so each summary has a small, clear job, or
Use the full AI chat (the Agent), which can work through a multi-part request across several turns.

The standard Daily Report with Smart Summary looks consistent every day because it runs a FLYR-tuned prompt against one fixed table shape. That consistency comes from the pairing of prompt and table, not from a layout instruction — it isn’t something you reproduce by pasting a format spec into a different chart.

Getting Good Results

1. Structure the Table First

The summary can only describe what’s in the table, so the table matters more than the prompt. A few things that consistently help:

Turn on column totals so the AI can see overall context and proportions.
Add the Currency Code dimension whenever the table has revenue metrics (ADR, Revenue, RevPAR). This is what tells the AI which currency symbol to use.
Put the comparison next to the metric — e.g. ADR and ADR STLP side by side — if you want the summary to talk about the change.
Sort the table the way you’d read it — chronological for time, largest-first for performance.
Only include columns that carry a signal. Extra metrics the summary is “supposed” to comment on but that don’t move are a common cause of padded, repetitive sentences.
Add subtotals only for genuine groupings (e.g. region → property). Skip them otherwise.
Keep the row count sensible. A very large or truncated table gives the AI an incomplete picture.

2. Write a Short Context Prompt

Use the Additional Context field to point the AI at what matters. A good prompt has up to three parts, and every part is optional:

Focus — the one or two things to look at. “Focus on weekday vs. weekend ADR.”
Context the data can’t show — an event, a rate change, a strategy. “We raised rates ~5% in March.”
Format — a light steer, not a template. “One short paragraph plus a one-line anomaly note.”

Prompts that work well:

“Summarise this month vs. STLP. Name the biggest driver of the RevPAR change and any single day that looks unusual.”
“Which segments gained or lost the most units year-on-year? We launched a corporate rate initiative in February.”
“Are we ahead or behind last year on pace, and is rate holding up as occupancy builds?”

Prompts that will disappoint you — too many parts, or multi-step logic:

“Provide: 1. a TY vs LY table with absolute and % change for OCC, ADR, RevPAR and their indexes; 2. a weekday vs weekend split stating if we over- or under-index; 3. a day-of-week deep dive diagnosing the cause for any day where RGI is below 100; 4. top 3 wins, top 3 opportunities, and a trend vs last year.”

That asks the summary to build a table it can’t format, run a per-row diagnosis it can’t do in one pass, and produce four separate sections. Split it across tiles, or take it to the AI chat.

3. Add Context Only Where It Changes the Read

The AI doesn’t know your market, your calendar, or your strategy unless you tell it — but it only needs the details that would change how a number should be read. Useful additions are things like:

A rate change, promotion, renovation, or closure and roughly when it happened
A local event or demand driver in the period
A target or threshold that matters (“we need 70% occupancy to break even”)
Who the summary is for, if that changes the language (“for our ownership group — keep it plain and RevPAR-focused”)

Two or three lines is plenty. A long background paragraph competes with your focus points and dilutes the result.

Starter Prompts

Copy one and adjust it. Each is deliberately short — that’s the point. Add one line of your own context if the data doesn’t explain something.

Performance vs. previous period

Summarise this period vs. the previous period. Give the headline change, the single biggest driver, and any day or line that looks anomalous. Short paragraph, plain language.

Pick-up

Summarise pick-up over this period: net units and revenue, the pace vs. last year, and any day with an unusual swing. Keep it to a short paragraph.

Segment mix

Which segments gained or lost the most this period, by units and by revenue? Name the one segment that most drove the overall result. One short paragraph.

Forward pace (on-the-books)

Are we ahead or behind last year on pace for this window? Say whether it’s a volume or a rate story, and flag any month or week that stands out. Short paragraph.

Need something broader — performance and channel and segment and STR in one place? That’s a multi-tile dashboard or an AI chat question, not a single summary. See “If you need a full multi-section report” above.

Best Practices

Fix the table before the prompt. Totals on, currency code in, comparison columns next to their metric, and nothing that doesn’t carry a signal.
Name one or two focus points, then stop. If your prompt has more than about three requirements, trim it or split the work across tiles.
Add only the context that changes the read — an event, a rate change, a target. Two or three lines.
Steer the format lightly. “Short paragraph,” “bullets,” “add an anomaly note” — not a fixed template.
If it’s not right, take something out. The fix is almost always to remove a requirement, not add one.
For a full report, use several tiles or the AI chat. Don’t ask one summary to do the job of a report.

Troubleshooting

If…	Why it happens	What to do
The summary is too general	The prompt didn’t point at anything specific	Name one metric or comparison to focus on. e.g. “Focus on the ADR change vs. last month and what’s behind it.”
It misses a trend you can see in the table	It wasn’t told that trend mattered	Call it out directly. e.g. “Focus on how occupancy and ADR moved vs. the previous month.”
The comparison is wrong or vague	The comparison period wasn’t named	Name it. e.g. “Compare this month to STLP and to Previous actuals.”
The output is short or stops halfway	The prompt is asking for too much at once, or for multi-step analysis	Cut it to one or two focus points, or move the work to separate tiles / the AI chat. e.g. “Just the RevPAR change vs. last year and its main driver.”
A sentence just repeats the one before it	It’s being asked to elaborate beyond what the data supports, so it pads	Ask for fewer points, and remove any table columns that don’t carry a signal. e.g. “One short paragraph: headline change and the single biggest driver.”
Blank “%” values, or it reads oddly for a new / low-volume property	There’s no prior-year data (or too many zero rows) for a comparison-based summary to describe	Point it at absolute levels, not year-on-year. e.g. “Summarise this month’s occupancy and ADR levels and the day-to-day pattern. Do not compare to last year — we have no prior-year data.”
The language is too technical for the audience	It wasn’t told who’s reading it	Say so. e.g. “Plain language for a non-revenue audience.”
It ignores something happening at the property	It can’t see events, promotions, or changes	Add one line of context. e.g. “We ran a flash sale 3–7 March and raised BAR ~5% from mid-month.”
You want it to recommend actions	On one table it usually can’t say anything specific, so it gives generic advice	Point it at the biggest gap instead of asking for a strategy section. e.g. “Name the segment furthest behind pace and by how much.” For real recommendations, use the AI chat.
You need a full multi-section report (performance + channel + segment + STR)	One tile summarises one table	Build it as several tiles — one query and one short summary each — on a single one-pager dashboard. Per tile: “Summarise the channel mix and the biggest year-on-year shift.”

The Short Version

Structure the table well, point the summary at one or two things, add only the context that changes the read, and keep the format steer light. If you find yourself writing a long, numbered prompt, that’s the signal to split the work across tiles or take it to the AI chat instead. The best results come from a clear table, a short prompt, and your own judgement on top.

📌 Need a hand? Reach out to the Advisory team via chat — and share any prompts that have worked especially well for you.

Getting Useful Output from AI Summaries on Dashboards