Lead scoring with AI, compared with a points sheet
When lead scoring with AI beats adding up points, when it does not, and how to check a score before your reps rely on it, with Ask why on every result.
Part of the guide AI employee for sales prospecting
Lead scoring with AI means giving a language model each lead's facts and a plain description of your best customer, and asking for a score from 0 to 100 and a one-sentence reason. It beats a points sheet when the facts are messy and the judgment is a matter of reading ("looks like a busy family-run place that would need weekly deliveries"). A points sheet is still better for hard rules ("must be within 10 km", "must not be a chain"). The strongest setup uses both: rules first to remove what can never fit, then the AI to rank what is left, then a check of the reasons before anyone trusts the order.
This is for sales managers and founders deciding how to rank a lead list. It compares the two methods, shows a made-up side-by-side on six leads, lists the mistakes that make AI scores misleading, and explains how to set it up and check it in xMap Flows.
How a points sheet works, and where it breaks
A points sheet adds a fixed number for each fact: plus 10 for a rating above 4.5, plus 15 for being independent, minus 20 for being more than 10 km away. It is transparent, cheap and repeatable. It breaks in three places. It cannot read text, so a description or a category with an unusual name scores nothing. It treats every fact as independent, so a busy chain and a busy independent both get the busy points. And somebody has to keep tuning the weights, which rarely happens.
What an AI score adds
An AI score reads the whole record at once: the name, the category in whatever words it came in, the rating and reviews, the address, the website text if it was brought in, any notes you supply. It can weigh facts against each other the way a person would. The price is that it is less predictable: the same lead can score 72 one day and 68 the next if its facts change slightly, and a vague description of your customer gives vague scores.
| Points sheet | AI score | |
|---|---|---|
| Reads free text | No | Yes |
| Handles odd category names | Only if listed | Yes |
| Explains itself | The sum of its rules | A sentence per lead |
| Same answer every time | Yes | Close, not exact |
| Best for | Hard rules and cut-offs | Ranking what passes the rules |
A comparison of the two methods, not of any measured results.
A made-up side-by-side
This example is made up. A coffee equipment dealer in Cape Town scores six cafes. Its points sheet gives points for rating, reviews and being independent. Its AI instruction says: "Our best customers are independent cafes that roast or brew specialty coffee and are busy enough to wear out a machine in three years. Chains buy centrally. Score 0 to 100 and give one reason."
| Cafe | Points | AI | AI reason |
|---|---|---|---|
| Harbour Roasters | 40 | 91 | Independent, roasts on site, very busy |
| Bean There Express | 45 | 22 | Part of a chain: equipment bought centrally |
| Kloof Street Kitchen | 35 | 58 | Independent, but mostly a restaurant; coffee is a side line |
| Morning Pour | 30 | 77 | Independent specialty bar, opened recently, smaller volume |
| Station Kiosk | 20 | 15 | Takeaway kiosk with a capsule machine |
| Ten Cups | 45 | 80 | Independent, busy, brews specialty coffee |
Made-up cafes and scores.
The points sheet liked the chain because it was busy, and could not tell a restaurant from a cafe. The AI ranked by what the dealer actually sells into. Neither is right by default; the reasons are what let you decide.
Checking a score before you trust it
- Run the scoring on a sample of about twenty leads you already know well: some won customers, some losses, some you would never call.
- Read every reason. A reason that cites a fact the lead does not have, or misses an obvious one, means the instruction or the input needs work.
- Check the order: your won customers should sit near the top, the never-calls near the bottom.
- Change one thing at a time (the description of your best customer, a note, a field you bring in) and run the sample again.
- Only then switch it on for the full list, and look at the reasons on the first real batch too.
Mistakes that make AI scores misleading
- Describing the customer in adjectives. "Premium, high-potential" gives the AI nothing to see. Describe facts it can read.
- Hiding the rules inside the AI. If "no chains" is a rule, filter chains out first. Do not hope the AI remembers.
- Scoring with too little input. An AI can only read what earlier steps brought in. If the website text or the chain flag is not there, the score cannot use it.
- Trusting the number without the reason. Always keep the sentence; it is how you catch a wrong score.
- Never looking again. Your best customer changes as you grow. Re-read the instruction each quarter.
Setting it up in xMap Flows
In xMap Flows, AI scoring is the xMap LLM block with the task Score & rank, which writes a fit score from 0 to 100 and a reason on each lead. The order is:
- A block that brings the leads in: Nearby places, New rows from a sheet, or My data.
- Filter: the hard rules, such as independents only or a minimum rating.
- xMap LLM, Score & rank: your description of the best customer, with any notes or reference files.
- Sort and limit: highest fit score first, keep the top 20.
- Email or a sheet.
In Results, every scored lead has Ask why: it shows the facts the score was built from, the instruction it followed and which rules the lead passed, and answers questions such as "why was this one ranked below that one?". For a strict yes or no per lead instead of a score, Jev decision gives a typed answer. The template Score new leads scores rows as they arrive in a Google Sheets spreadsheet, and Morning call list with decision makers and call notes scores and ranks a daily list. The guide to AI employees for sales prospecting shows where scoring fits in the rest of the work.
Questions
Is AI lead scoring accurate?
It is as good as what it is given: the facts brought in about each lead and a concrete description of your best customer. Check it on leads you already know before relying on it.
Should I replace my points model with AI?
Usually not entirely. Keep hard rules as filters, and use the AI to rank what passes them.
Why did the same lead get a slightly different score?
AI scores are close but not exactly repeatable, and they move when a lead's facts change. Look at the order and the reason rather than a single point of difference.
Can I see why a lead got its score?
Yes. Each score comes with a one-sentence reason, and Ask why on a result shows the facts and rules behind it.