Keeping AI employees honest: grounding every claim in real data
An AI that writes well can also write numbers nobody counted. How to build AI employees whose figures come from data blocks and whose gaps are said out loud.
Part of the guide AI employees
An AI employee stays honest when the counting is done by data blocks and the AI only reads and words what they found: every figure in its message should trace back to a field a block returned, every gap (no data here, sample figures, a late day) should be said rather than smoothed over, and anything that leaves your company should wait for a person until you trust it. Language models are good at writing; they are not a source of facts about your market, so design the flow so they never have to be.
This is for anyone who has set up, or is about to set up, an AI employee in xMap Flows: owners, analysts, sales and operations leads. It sets out the rules that keep reports truthful, shows what a dishonest report looks like next to an honest one, and lists the settings and habits that make the difference.
Rule one: data blocks count, the AI words
In a well-built flow, the numbers are produced before the AI sees anything. Nearby places counts competitors, Demographics counts residents, Team findings counts what your other AI employees found this week against last. Then xMap LLM turns those fields into sentences.
Say so in the instructions. A line such as "Use only the figures in each item. Never add a number that is not in the data. If a figure is missing, say it is missing." changes what comes back. For scores, ask for a reason that names the fields it used, so a reader can check it.
Rule two: say what the data cannot see
Every data source has limits, and an honest report repeats them where they matter:
- Footfall is people-hours of presence, not visitors, and each day arrives about 10 days late.
- Traffic Flow gives live speeds and how jammed roads are, never car counts.
- Population figures exist in about 100 countries, not the United Kingdom or the United States.
- Newly found is not newly opened. A place appearing for the first time may have traded for years; read the opening date where there is one.
xMap Flows already marks some of this for you. A block that runs on sample data says "sample" in its step, and emails carry a note that the figures are examples. A travel area or a route worked out as an estimate is said to be one. A place with no footfall gets a note instead of a zero. Do not hide those notes when you rewrite a message; keep them in the template of the email.
What a dishonest report looks like
A made-up example. A restaurant group in Singapore asks its weekly analyst for a brief. Two versions of the same paragraph:
- Not grounded: "Competition is intensifying across the city, with openings up sharply and diners shifting to casual formats." Nothing here was counted; the trend about diners was invented.
- Grounded: "Your flows found 7 new restaurants this week against 4 last week, 5 of them within 1 km of Tanjong Pagar. Two of the 7 belong to a chain. No figures on diner preferences were collected."
The second is less dramatic and far more useful: each number can be checked in the results, and it says plainly what it does not know. The figures in both are invented for the example.
Rule three: check the work before trusting it
- Run a test first. A test run sends nothing, and the run panel shows what each block passed on: click a block to see its fields and first items.
- Read the AI's reasons on a few results. Every result has Ask why, which explains a score from the item's own fields and the flow's rules, and says when the data does not show something.
- Run it once for real and compare the email with the results. If a sentence has a number you cannot find in the results, tighten the instructions.
- Look at the step that failed, if any. A block that fails says why; it does not fill the gap with guesses.
Rule four: let sends wait until you trust it
Each AI employee can do more or less on its own. Set it to Ask me first and an approval step stops anything going out (emails, messages, reports) until a person approves; you can even edit a drafted message before it leaves. Move it to Act and tell me once a few weeks of output have been right, and keep Approval on anything that writes to customers.
An honest weekly brief, block by block
- Schedule: every Monday at 07:00.
- Team findings: what the other AI employees found this week against last, each finding with its own figures.
- xMap LLM on the whole list, told to use only the figures given, to name what is missing, and to keep it under 300 words.
- Approval: a manager reads it first.
- Make a report or Email to the leadership team.
What lands in the inbox is a short brief where every number matches the results page, sample or estimated figures are marked as such, and the closing line says what the data did not cover. Templates built this way: the weekly market brief, Weekly competitive intelligence for my cafes and New places every morning, only new.
Questions
Can xMap LLM make up numbers?
Any language model can write a number it was not given, which is why the flow should count with data blocks and tell the AI to use only the figures in each item. Its reasons are meant to cite the item's own fields.
How do I know if a result came from sample data?
Blocks that ran on sample data say so in their step, and emails and messages carry a note that the figures are examples.
What does Ask why do?
It explains a result's score from the item's own fields, the flow's rules and the AI's reason, and says when the data does not show something.
Should every AI employee need approval?
Start with Ask me first for anything that writes to people outside your company. Loosen it once the output has been right for a few weeks.
Templates that do this
Read next
More on ai employees and automation · Every article · Templates