What AI scores can and cannot tell you about a business

An AI fit score tells you how well a business matches the description you gave, judged only on the facts your flow found about it. It does not tell you the chance that the business will buy, it does not know anything the earlier blocks did not find, and it is only comparable with other scores written by the same instructions. Read it as a sorted reading list with reasons, and check the reasons, not just the number.

This is for people using an AI employee that scores places or companies and wondering how much to trust the numbers. It covers what goes into a score, how to read one, a made-up example of a score that misled, the most common misreadings and how to set up a flow so scores stay honest.

What a score is made of

The xMap LLM block, set to score and rank, writes two fields on each item: a fit score from 0 to 100 and a one-sentence reason. To write them it reads three things:

• the item's own fields, as earlier blocks found them: name, category, rating, number of reviews, chain or independent, address, website, opening date when known, counts of places nearby, a job title, anything else added;
• your instructions: what a good fit is for you;
• any notes or reference files you attached.

That is all. It does not look the business up anywhere else during the run. If the flow did not find the business's size, the score cannot be based on size, however the instructions are worded.

What a score can tell you

• Order. Within one run, a business at 85 matches your description better than one at 40, on the facts available.
• Why. The reason names the facts that pushed the score up or down. It is the most useful part.
• Where to start. Results scored 75 or more are counted as strong fits in run summaries and the morning briefing, which is a sensible first slice for a busy team.

What a score cannot tell you

How the score is produced, not a measured error rate.

The last row surprises people. The AI employee does not watch which results your team called or ignored and adjust. If the scores are wrong in a consistent way, the fix is in the instructions or in the facts the flow gathers.

A worked example

A made-up case: a company in Singapore sells a staff rostering app to restaurants and scores new restaurants every morning. The first week, the highest scores go to large, well-reviewed places, and the sales team finds most of them already use a rostering tool.

1. The manager reads the reasons for the top ten. Nine say some version of "large, busy, many reviews, so many staff to schedule".
2. That is a fair reading of the instructions, which asked for "restaurants with many staff". Many reviews was the only fact in the item that hinted at staff numbers.
3. She changes two things. The instructions now say that young independents with a growing number of reviews are the best fit, and established chains are a poor one. And she adds Job postings before the scoring, so restaurants advertising for floor staff carry that fact.
4. The next week's top scores are mostly independents under two years old, several of them hiring. The reasons now mention hiring by name.

The score was never wrong about its own question; the question was wrong. Reading the reasons is what showed it.

Misreadings to avoid

• Treating 70 and 72 as different. Small gaps mean little. Think in bands: strong, possible, unlikely.
• Comparing this month's 80 with last month's 80 after the instructions changed.
• Trusting a score whose reason cites a fact you cannot see on the item. The reason should name only the item's own fields; if it does not, tighten the instructions.
• Filtering at a high line before checking. A filter at 90 on the first day may hide good businesses the instructions underrate.

How to keep scores honest in an AI employee

1. Put every fact you want judged before the xMap LLM block: Nearby places for the basics, counts around each place, Job postings, your own table through Match with my table.
2. Write instructions that say what high and low mean, and that the reason may cite only the item's fields.
3. Use Sort and limit to keep the top results instead of a hard filter, at least at first.
4. On each result, use Ask why to see in plain words which facts and rules led to its score.
5. If you need a strict yes or no, or one label from a fixed list, a typed question with Jev decision or the Classify task fits better than a score.

What lands in the inbox is the list sorted by score, each with its reason beside it. Templates that work this way include New places every morning, only new and Score new leads from a sheet.