An account scoring model assigns every account a number meant to represent how worth pursuing it is. The arithmetic is the easy part. The hard part is the sentence that comes after the number, which has to say what somebody does at 80 that they would not do at 55.
Teams almost always build in the other order. They gather signals, argue about weights, ship a score, and then look for something to do with it. By then the model has no specification, so there is nothing to test it against and no way to say whether it is working.
Start at the threshold, not the inputs
Write down the decision the score is meant to drive, in one sentence, before anyone opens a spreadsheet. There are only four honest endings to that sentence.
- Who gets worked at all. The score is a cutoff and everything below it goes unworked or into nurture. This is the strongest version and the one teams are least willing to commit to, because it means saying out loud that some accounts will not be called.
- Who gets it. High scores go to senior reps or a named account team. Routing reads the score, so the score has to be defensible to the people whose pipeline it reassigns.
- What happens first. A high score triggers research and a person; a low one triggers a sequence. The score buys effort rather than attention.
- Nothing. The score sits in a column and gets sorted occasionally. This is where most models land, and it is not a failure of the maths.
If the answer is the fourth one, the project is a reporting exercise. That can be worth doing, and it should be funded and staffed as one rather than as a prioritisation system.
Fit and intent are two models wearing one number
The single most common structural error is blending how well an account matches your profile with how much interest it is currently showing, then reporting the average. The two answer different questions and expire at different speeds.
| Signal | What it tells you | How fast it goes stale |
|---|---|---|
| Fit | Whether this account could ever reasonably buy: size, sector, structure, stack. | Quarters, sometimes years. |
| Intent | Whether somebody there is thinking about it now: pricing visits, relevant hiring, review-site activity. | Days. |
| One blended score | Neither question, answered confidently. | Unknowable, because the halves decay at different rates. |
Blending destroys the only distinction that changes behaviour: whether to call or to wait. A strong-fit account with no intent is a patience problem and belongs in a long nurture. A weak-fit account with high intent is usually a student, a competitor or somebody researching for a neighbouring problem. Averaged together, both arrive as a 60 and both get the same treatment.
Keep two numbers. The grid of fit against intent is also far easier to explain to the people being routed by it than a weighted sum they cannot reconstruct.
Where the weights should come from
Asking the team produces a flat model. Everyone defends the signal they contributed, nothing gets cut, and you end up with twelve inputs at roughly eight points each, which is a model that cannot separate anything.
The alternative is to derive weights from closed deals: take what you won and what you lost, and keep the attributes that actually separate the two. That requires enough closed business to be more than an anecdote, and the honest answer is often that you do not have it yet.
Rules are not a lesser version of a model. A rule says a thing a person can argue with, which means it can be corrected when it is wrong. A fitted model on forty deals produces a number nobody can argue with and that is mostly noise.
A score needs a date next to it
Scores are usually recomputed on a schedule and read continuously. Between runs the number is drifting away from reality, and it looks exactly as authoritative on day six as it did on day one.
This matters most for the intent half, which can be hours old or three weeks old with no visible difference. Store the timestamp alongside the value and show it wherever a human reads the score. The same logic applies to the inputs underneath: enriched fields decay while continuing to look current, and a score built on a two-year-old headcount is confidently describing a company that no longer exists.
When not to build one
If you can work the entire list, prioritisation is overhead. Under a few hundred target accounts with a team that can cover them, a score ranks things that were all going to be worked anyway.
If nothing downstream will read it. The test from the first section, applied before the work starts rather than after.
If the fields it depends on are not governed. A score computed from a field that four processes write in three formats is an expensive way to formalise bad data. Fix the write path first.
And the score only becomes a decision at the point something acts on it automatically, which is usually where a good model quietly gets discarded in favour of whoever is next in the rotation.
Questions people ask
What is an account scoring model?
A method for assigning each account a number representing how worth pursuing it is. The useful version is defined by the decision it drives, not by the signals it consumes.
Should fit and intent be one score or two?
Two. They answer different questions and decay at different speeds, and averaging them hides the distinction between calling now and waiting.
How many closed deals do I need before fitting a model?
Enough that the pattern is not an anecdote. Under about fifty closed-won, write rules you can defend out loud instead, because a fitted model on that sample mostly describes noise.
How often should scores be recalculated?
Fit can refresh slowly. Intent has to refresh close to continuously or it is misleading, and either way the score should carry the date it was computed.
How do I know the model is working?
Compare win rate above and below the threshold. If they are the same, the score is not separating anything, whatever its distribution looks like.
Or skip the search
Want the person who owns it after it ships?
If the work is a standing obligation rather than a bounded project, tell me which system you want built first. You get example profiles and a call with your pick before anything starts.
Placement runs through AYTalent. Engineers are employed and trained on the bench before they start with you.