An AI report generator that sticks to the figures it was given
Turn each run into a branded PDF or slides with a summary xMap AI writes. How the figures are counted and how to keep the AI from inventing numbers.
Part of the guide AI employees
An AI report generator is only as trustworthy as the numbers in it. The safe design splits the work: software counts the figures, draws the charts and lays out the list, and the AI writes only the sentences around them, from the data it was handed. In xMap Flows the Make a report block does the counting and the layout, as a PDF with your logo and colour or as 16:9 slides; an xMap LLM block set to The whole list writes the summary at the top; and Check the work can reject any summary that quotes a number the data does not hold.
This article is for anyone who has to produce the same document again and again: a monthly market note for a board, a competitor pack for a client, a site summary for a committee. It explains what goes in, where the AI stops, and how to set it up.
What the document holds
Make a report builds the pages from the items that reach it. Each part appears only when you include it and the items have it:
| Part | Who makes it | What it shows |
|---|---|---|
| Summary | xMap LLM (the whole list) | A few paragraphs on what stands out |
| Figures | Make a report, by counting | How many results, their average rating, how many strong fits |
| Charts | Make a report, by counting | Bars by kind of place and by city, the top 8 of each |
| Map | Make a report | Every item that has a position |
| The list | Make a report | The first 100 items with their common columns |
The figures and charts count every item, up to 2,000; the list shows the first 100.
Slides are shorter: a title page, the figures, the summary, the top eight of the list and the leading kind or city, with a band of your colour. You can make both from the same run; each is its own PDF.
Keeping the AI to the facts
A language model will happily write numbers rose sharply when nothing in front of it says so. Three habits stop that:
- Ask the xMap LLM block for words, not arithmetic. A line such as Describe what stands out in this list. Do not calculate totals or percentages; the report adds them. Use only names and figures that appear in the items. keeps it in its lane.
- Give it the context it would otherwise guess. Short notes you add to the block (your areas, what counts as a good result) are read with every run.
- Put Check the work between the AI and the report, set to check the run's report, with the ready-made check no numbers that are not in the item's own data and a rule such as no claims about the future. A summary that fails is sent back with what was wrong and written again.
Where it goes
Make a report emails the PDF as an attachment or saves it to a Google Drive folder you name, or one named after the flow. Emails follow the same rules as the Email block: on the Free plan only people at your own company's domains, a daily limit for the organisation, and anyone who unsubscribed is left out. Saved to Drive, each item gets a link to the file for any block after it, such as a Slack message saying the pack is ready.
A worked example (made up)
A coffee chain in Mexico City has twelve cafes and a board that meets on the first Tuesday of each month. The operations lead wants a short pack on the coffee shops opening and being rated around the cafes, without spending a day on it.
She builds: Schedule (the first of the month, 7:00 Mexico City time), My data (the twelve cafes), Nearby places (coffee shops within 1 km of each), Remove duplicates, xMap LLM on the whole list (three paragraphs for the board: which cafes face the most competition, which rivals are best rated, anything new; no percentages), Check the work on the report, then Make a report set to Both, sent by email to the board's three members at the company's domain.
The first test run shows the AI writing competition increased. Nothing in a single month's list can show an increase, so she adds do not compare with earlier months to the instructions. Check the work would have caught it too, as a claim with nothing in the data behind it.
Mistakes to avoid
- Asking the AI to work out averages. The report counts them already, and counts them correctly.
- One report for every audience. A board wants slides; an analyst wants the full list. Make both, or two flows.
- A cover title that never changes. Left empty, the title is the flow's name and the date, which helps when files pile up in a Drive folder.
- Forgetting your logo and colour. Set them once under Settings, Logo and colour, and every report and email uses them.
Where this fits among other automations is covered in the workflow automation guide. For a quick start, the monthly brief on your store network template already ends in a written summary.
Questions
Can the report compare this month with last month?
Only if both months are in the items it receives. A single run's list cannot show change on its own, so tell the AI not to claim any.
Can I edit the report before it goes out?
Put an Approval before Make a report: a teammate looks at the items and approves before the document is made and sent.
Does it work without a summary?
Yes. Leave the xMap LLM block out and the report holds the figures, charts, map and list only.
What sizes are the pages?
The report is on A4 pages and the slides are 16:9, each as a PDF.
Templates that do this
Read next
More on ai employees and automation · Every article · Templates