We scored every target account for AI maturity. Here’s what the scores got wrong.

A score that doesn’t publish its error bars gets treated as a fact.

Sound familiar? There’s a score on every account. A rep skips one because it ranked low. Six months later a competitor closes it.

On a client engagement I helped build an AI-maturity model for a field team. The brief was simple: tell reps which accounts were ready for an advanced conversation, so each rep wasn’t forming a private opinion from a web search. Every account got scored on three dimensions from public sources only: AI in production, data infrastructure, and organizational readiness — whether someone there could own it and buy it.

Each classification mapped to a motion: which conversation to lead with, which people to target, how hard to push. It worked. Reps stopped arguing about which accounts were “real,” and leadership got a defensible view of what the pipeline was made of.

Use the minimum, not the average

The most important decision was how those three combined. We used the minimum, not the average.

Averaging lets two strong dimensions hide a fatal gap. A company with AI in production and no executive owner isn’t ready; your champion has no authority to sign.

Four ways the score was wrong

The rule was right. It was also the source of the model’s most common mistake, along with three others:

• It only saw what was public. Private companies and quiet industries scored lower than they deserved. That’s bias with a direction, not noise.
• It confused talking with doing. A fresh press cycle scored high; a company that went quiet while scaling scored low. It measured publicity about AI, not AI in use.
• The deciding dimension was the hardest to see. Executive hires and governance teams often aren’t announced, so the minimum rule leaned hardest on the weakest evidence.
• The cautious rule cost real accounts. I’d make that trade again — a false negative delays a conversation, a false positive burns a rep’s credibility — but it’s a cost, and pretending otherwise is how a model loses the field.

Ship the caveats with the score

None of those are reasons to skip scoring. A field team without a shared score just runs on gut feel, which has all four of these biases and no way to correct them. They’re reasons to be honest about what the score is: a set of inferences from public information, not a verdict.

What kept it useful was shipping the caveats with the scores, in the same document. Reps were told what the model couldn’t see and to override it when they knew better. Corrections went back into the CRM, so the score got better instead of aging quietly. And it stayed a signal, never a gate. The moment a score blocks someone from working an account, people route around it, and you lose the feedback that keeps it honest.

If you’re choosing a rule, here’s how the three common ones behave:

If you’re building one

• Publish the error bars with the output. Every user should be able to say in one sentence how the score can be wrong.
• Pick the combination rule on purpose. Minimum, average and weighted behave very differently. Write down why.
• Build the override path before launch. A score nobody can correct gets worse every week.
• Name the proxy honestly. Ours measured public discussion of AI. That correlates with maturity. It isn’t maturity.

(The data underneath a score matters as much as the rule on top. That’s the subject of fix your context layer.)

How does your team rank accounts today? If a rep asked how the score could be wrong, would anyone have an answer?