Checklist for calibrating AI lead scoring

Improve AI lead scoring accuracy by refining customer descriptions, adding negative constraints, and adjusting the fit line to filter bad leads.

Get started

Short answer

To improve AI lead scoring, add negative constraints and specific industry markers to your customer profile. If the AI lets in bad leads, the description is too broad; if it filters out good leads, the requirements are too rigid. Review misclassified leads, update the description to exclude those traits, and adjust the fit line to tighten the filter.

Identify the scoring gap

AI lead scoring fails when the instructions are vague. Tell an AI to find marketing professionals and it will score a freelance social media manager and the CMO of a 500 person company the same way, because both match the keyword. The score looks precise, but the list is not.

The gap shows up in two directions. False positives are people who look right on paper and will never buy: consultants, students, agency people selling to the same buyer you sell to. False negatives are real buyers who describe their job in language your description doesn't contain, so the AI scores them low and drops them.

Both errors come from the same place. The description you gave the AI is a rough sketch, and the score is a number attached to that sketch. Until you know which traits separate a buyer from a near miss, the fit line is a guess with a decimal point.

Broad descriptors

  • Experienced sales leader
  • Tech enthusiast
  • Business owner

Precise constraints

  • VP of Sales at a Series B SaaS company
  • Head of Infrastructure running production on AWS
  • Founder of a 10 to 50 person logistics agency

A score only means something against a description you have tested. Move the fit line after you fix the description, never before.

Audit a sample of bad leads

You can't fix a scoring model without seeing where it breaks. Pull the leads that scored high and are obviously wrong. Twenty is enough to see a pattern, and you probably already have them in a lead group or an export.

Read the profiles, not the scores. Look for the word or job title the AI treated as a signal. Common ones: 'consultant' in a search for experts, 'founder' in a search for decision makers at larger companies, 'partner' at a firm that resells your category instead of buying it.

Then check the company. A small agency and a corporate department can look identical in a headline. The AI reads 'Head of Growth at Northwind' and has no idea whether Northwind has nine people or nine hundred.

  1. 1

    Export 20 high scoring bad leads

    Gather the leads the AI scored 80 or above that you would never contact.

  2. 2

    Find the trick keyword

    Look for the word or phrase that appears in most of these profiles and is misleading the score, such as 'Consultant' showing up in a search for 'Expert'.

  3. 3

    Check the company type

    Decide whether the AI is confusing company size or industry, like mistaking a small agency for a corporate department.

  4. 4

    Write down the disqualifier

    Name the one trait that makes these leads unqualified so you can turn it into a negative constraint.

Refine the customer description

Rewrite the description with concrete nouns and numbers. Adjectives like 'successful', 'growing' or 'innovative' give the AI nothing to check. 'VP of Operations at a logistics company with 50 to 200 employees' gives it three things to check.

Negative constraints do most of the work. Tell the AI who to ignore, by trait, not by name. Excluding freelancers, consultants and students removes whole categories of near miss that would otherwise score in the seventies because they share vocabulary with your buyer.

Keep the exclusions short and specific. A list of thirty banned words turns into a puzzle the AI tries to solve in creative ways. Five or six traits that describe the people who waste your time are enough to start.

Vague role

Avoid this
Marketing expert
Use this instead
Director of Demand Gen
Reason
Removes freelancers and generalists

Broad industry

Avoid this
Healthcare
Use this instead
Private outpatient clinics
Reason
Excludes hospitals and insurance

Generic size

Avoid this
Mid sized company
Use this instead
50 to 200 employees
Reason
Gives the AI a hard boundary

Soft signal

Avoid this
Interested in growth
Use this instead
Hired three SDRs in the last year
Reason
Something you can check on the profile

Paste this job

Target profile: VP of Operations at a logistics company with 50 to 200 employees. Must have worked on cold chain shipping or warehouse operations. Exclude: freelancers, consultants, students, and owners of companies with fewer than 10 employees. Exclude anyone whose company sells logistics software or consulting.

Calibrate the fit line

The fit line is the score below which leads are dropped. If the description is right and bad leads still get through, the line is too low. If almost nothing gets through, the line is too high or the description is too narrow.

Change one thing at a time. If you rewrite the description and move the line in the same pass, you won't know which change did the work. Rewrite first, look at the score distribution, then move the line.

Expect the distribution to shift every time you edit the description. A line at 70 against one profile is not the same line at 70 against another.

  1. 1

    Set a starting line at 70

    Run a search and look at how scores spread before you judge any single lead.

  2. 2

    Raise it when bad leads still pass

    Move the line up in steps of five or ten and check whether the leads at the top of the list are still people you would contact.

  3. 3

    Fix the description before lowering the line

    If good buyers are missing, add the language they actually use to the description. Lower the line only after that.

  4. 4

    Recheck the audit sample

    Run the twenty bad leads through the new description and confirm they now score below the line.

Test against leads you already know

The fastest way to trust a score is to run it against people whose value you already know. Export ten customers you would take a call with tomorrow and ten leads you have already disqualified. Score both lists.

If the customers land above the line and the disqualified leads land below it, the description is doing its job. If the two lists overlap, the description is still too loose, and no amount of fit line adjustment will separate them.

Keep the two lists. Every time you change the description, run them again. It takes a few minutes and it tells you whether the change helped before you spend any invites.

Testing by feel

  • Scoring a fresh search and judging the top ten by gut
  • Moving the line whenever a bad lead reaches the inbox
  • Rewriting the description after every campaign

Testing against a fixed list

  • Scoring ten known customers and ten known bad leads
  • Comparing the two score ranges before changing anything
  • Editing the description until the ranges stop overlapping

Doing this with Kinlead

Kinlead finds people on LinkedIn who match the customer you describe, or your website, and scores each lead from 0 to 100 against that profile. Leads below the fit line are not kept, so the calibration above decides what lands in your lead groups.

You can import a CSV of LinkedIn profile URLs, which is how you run the known customer and known bad lead test without spending a search. Leads can be exported to CSV too, so the audit sample can come from an earlier campaign.

Outreach runs from your own LinkedIn account through Unipile, at most one action every five minutes, inside the daily limits and send window you set.

Frequently asked questions

Related

Get started with Kinlead

Prefer markdown? Checklist for calibrating AI lead scoring.md