Define what an account expansion or lead scoring pilot has to prove, and how the sales team will actually adopt it, before the pilot starts.
Usually adoption rather than accuracy. The score lands in a dashboard reps do not open, or it arrives without a reason so the first time it contradicts a rep's instinct it gets ignored. Deciding where the score appears and what explanation goes with it does more for the outcome than another round of model tuning.
Rules are often better, and a firm that never says so is selling you something. If you have few expansion events per quarter, or the drivers are well understood, a transparent rule set is cheaper, easier to adopt, and easier to argue with. Models earn their cost when the pattern is not obvious and the volume supports learning it.
At least one full sales cycle and preferably two. Anything shorter measures noise and whoever is sceptical will say so correctly. Set the end date and the decision that happens on it at the start.
Yes, or you cannot separate the model from a good quarter. A holdout of comparable accounts receiving no scores is the cheapest way to make the result defensible, and its absence is the first thing a sceptical CFO will find.
Then that is the project, and the scoring comes after. A model built on three disagreeing versions of the same account inherits the disagreement. This is the most common reason these pilots produce a result nobody trusts.
Expansion analytics fails on adoption far more often than on accuracy. The model is usually fine. Reps keep working their own list because the scores arrive somewhere they do not look, or contradict what they already believe and nobody explains why. Agree the success criteria and the adoption plan before the pilot, because afterwards is too late to find out nobody opened it.
Define what an account expansion or lead scoring pilot has to prove, and how the sales team will actually adopt it, before the pilot starts.
An account growth pilot needs two things agreed before it starts: what it has to prove numerically, and how reps will use the output in the workflow they already have. Most of these pilots produce a defensible model that nobody adopts, because the adoption half was never specified and no one owned it.
How will we know this worked, and will the team actually use it?
Pick a small number of measures and commit to them before you see any results.
What proportion of accounts expand today, over what period, without any model. Everything is measured against this. If nobody knows it, that is the first finding.
State it as a number now. A model that improves targeting by two percent may be real and still not worth the operating cost. Deciding the threshold afterwards invites motivated reasoning.
Which accounts are in scope and which comparable group is not receiving scores. Without a holdout you cannot separate the model from a good quarter.
At least one full sales cycle, preferably two. Shorter and you are measuring noise. State the end date and the decision that happens on it.
The reason these pilots stall is usually underneath the model rather than in it.
If three systems hold three versions of the same customer, the model inherits that. Check before building, because this is the common blocker and it is a different project.
Prediction needs enough prior expansions to learn from. A business with a handful of expansion events per quarter should use rules rather than a model, and that is a legitimate answer.
If expansion is logged differently by different reps, the target variable is noise. Fix the recording before modelling anything.
A model that uses data only available after the expansion will look excellent and predict nothing. Confirm every input is available when the rep needs the score.
This is the half that gets skipped, and it is the half that decides the outcome.
In the CRM view the rep already works in, on the object they already open. A separate dashboard is a model nobody sees. This is the single biggest adoption factor.
A score with no reason gets ignored the first time it disagrees with a rep's judgement. Two or three drivers in plain language is the difference between a tool and a curiosity.
Name the behaviour. Work the top twenty accounts first, or log a reason when skipping a high score. A pilot without a required behaviour cannot be evaluated.
A named sales leader, not the data team. Track whether reps are opening and acting on scores separately from whether the model is accurate, because those fail independently.
Agree now what each outcome triggers, so the result is acted on rather than discussed.
The common result. The fix is workflow and enablement, not a better model. Agree that in advance or the instinct will be to rebuild the model.
Stop, and redirect the budget to the account view underneath. Pretending otherwise produces a second failed pilot.
Compare against the operating cost. Some models are correct and not worth running, and saying so protects your credibility for the next one.
A cruise line found expansion opportunity once customer activity was joined across systems. The analysis was possible because the account view came first.
The adoption answers matter more than the modelling answers.
What is our current expansion rate, and how will you establish it before we start?
Where exactly will the score appear in the rep workflow, and what explanation appears with it?
What happens if the model is accurate and adoption is poor? Who owns that?
Usually adoption rather than accuracy. The score lands in a dashboard reps do not open, or it arrives without a reason so the first time it contradicts a rep's instinct it gets ignored. Deciding where the score appears and what explanation goes with it does more for the outcome than another round of model tuning.
Rules are often better, and a firm that never says so is selling you something. If you have few expansion events per quarter, or the drivers are well understood, a transparent rule set is cheaper, easier to adopt, and easier to argue with. Models earn their cost when the pattern is not obvious and the volume supports learning it.
At least one full sales cycle and preferably two. Anything shorter measures noise and whoever is sceptical will say so correctly. Set the end date and the decision that happens on it at the start.
Yes, or you cannot separate the model from a good quarter. A holdout of comparable accounts receiving no scores is the cheapest way to make the result defensible, and its absence is the first thing a sceptical CFO will find.
Then that is the project, and the scoring comes after. A model built on three disagreeing versions of the same account inherits the disagreement. This is the most common reason these pilots produce a result nobody trusts.