Predictive account scoring
Rank the companies, not just the people. ax1om trains on your own account history and shows which of your CRM fields drove every score, with a stability measure on the explanation itself.
Account scoring ranks organizations
The distinction everyone agrees on, stated once so we can get to the part nobody writes about.
Lead scoring
Unit · A person
Asks · Is this individual likely to convert?
Used for · Routing and prioritizing individual follow-up.
Account scoring
Unit · An organization
Asks · Is this company likely to buy, renew, expand, or leave?
Used for · Choosing which companies get the team, the campaign, or the save.
They are complementary, not substitutable. Account scoring decides where the effort goes; lead scoring decides who inside that account gets worked first.
Account models fail on the population, not the model
Lead scoring usually has a clean population, because a lead is a lead. Account scoring almost never does, and that is where the trust goes.
The population is not the question
The common version: the records are every account in the CRM, prospects and dead accounts and partners and duplicates included, and the target is every lost deal. That model answers "which deals lose" correctly. It gets reported as though it answered "which customers leave". Those are different questions with different populations, and no amount of model tuning reconciles them.
Record-keeping artifacts count as losses
A duplicate, a merge, a deal re-opened on the same account a week later. None of them is a departure, and all of them look like one to a model. A churn model trained on those learns your CRM hygiene alongside your customers, and the resulting score is partly a measure of how tidy your data entry is.
Every account is measured from a different moment
Measure each account from its own outcome date and you describe them at different distances from the decision. A renewal-shaped question needs a fixed anchor: look back a set window before the renewal decision, so every account is described at the same distance from the moment that decides the answer.
None of these is a modelling error. Each one produces a technically correct model of the wrong question, which is why tuning does not fix it and why the resulting score quietly stops getting used.
What ax1om does about each one
The population is on the card
Two of the four starting points on Create a Score rank accounts. Both carry a population line before you click: trains on accounts that are current customers, read from the account type field on Salesforce or the lifecycle stage on HubSpot. Picking the goal produces the goal, rather than producing a filter you have to remember to apply.
Losses that are not departures leave
Losses identified as record-keeping artifacts, from your own loss reasons or from a new deal opened on the same account shortly after, are removed from the population instead of counted. The result carries a record-keeping block showing how many records left and the base rate before and after the cut. When it cannot run on your data, it says so rather than quietly counting everything.
The lookback anchors on the decision
Account-shaped models swap their cutoff discipline: rather than measuring each account from its own outcome date, they anchor on the renewal decision and look back a fixed window before it. An account with no resolvable renewal date is counted as unresolved rather than filled in with a guess.
One person rolls into one account
Person-to-account resolution runs before anything is aggregated. On Salesforce the contact’s account reference wins. Otherwise the email domain is matched against account website domains, with free-mail providers excluded so a personal address does not invent a match. On HubSpot the primary association wins, then the first listed. A person contributes to exactly one account, never several, so nothing is double counted.
Fields can come from more than one object. Account-level fields and the person-level fields that roll up to them merge into one training set, and the field panel shows how populated each one is before you spend a run on it.
Reading an account score
The score
A 0-100 ranking of the account against the outcome you defined. It ranks. It is not a percentage chance, and it is not a grade.
The factors
Per-record SHAP factors, rolled back to your own CRM field names, showing which fields pushed this account up and which pushed it down. This is the part the category argues for and does not implement.
The stability
A Feature Stability Score on the explanation itself, measuring whether those factors hold up across retraining. An explanation that changes every retrain is not an explanation you should route a save play on.
Scores land in a field on the account record. The same predictions are available through the scoring API and to agents over MCP, so routing, campaign selection, and agent workflows read one number rather than each re-implementing their own.
The same account scores, callable by your agents
Scoring is live on ax1om's MCP surface today. An agent working an account list calls the same trained model your dashboards and API read, and every score comes back with the fields that drove it, so the agent can show its math instead of writing a plausible number.
ax1om for AI agentsCommon questions
What is account scoring?
Ranking organizations rather than individuals by how likely they are to reach an outcome you care about: buying, renewing, expanding, or leaving. Lead scoring evaluates a person and answers whether to work them. Account scoring evaluates a company and answers where to put the team, the campaign, or the save. Most revenue teams need both, and they are complementary rather than substitutable.
How is account scoring different from lead scoring?
The unit changes, and so does what can go wrong. Lead scoring usually has a clean population, since a lead is a lead. Account scoring almost always has a population problem: which accounts belong in the model at all, and what counts as the outcome. A churn question asked of every account in the CRM, against every lost deal, produces a model that correctly answers a question nobody asked.
Why do account scoring models produce numbers nobody trusts?
Usually because the population and the target disagree. The most common failure is a model whose records are every account in the CRM, including prospects, dead accounts, partners, and duplicates, and whose target is every lost deal. A second common failure is counting record-keeping artifacts as losses: duplicates, merges, and deals re-opened on the same account. Both produce a score that looks like churn and is not.
Can you explain why an account got its score?
Yes, per record. Each account score carries SHAP factors rolled back to your own CRM field names, showing which fields drove it up and which drove it down, plus a Feature Stability Score measuring whether that explanation holds across retraining. The factors do not sum to the score; they show direction and relative weight.
Which record gets scored when a person belongs to several companies?
Exactly one. On Salesforce the contact account reference wins. Otherwise the email domain is matched against account website domains, with free-mail providers excluded. On HubSpot the primary association wins, then the first listed association. A person never contributes to more than one account, which is what prevents double counting in account aggregates.
What data does account scoring need?
Your own account history and a definable outcome. Fields can come from more than one object: account-level fields and the person-level fields that roll up to them are merged into one training set. A first model trains on as few as 50 conversions, and ax1om reports what it found in your data before you spend a training run on it.
Where do account scores end up?
In a field on the account record, so the team reads them where they already work. The same predictions are also available through the scoring API, and to AI agents over MCP, so routing rules, campaign selection, and agent workflows can use the score without anyone re-implementing it.
Score your accounts on your own history
Pick the goal, and the population comes with it. Train on your own account history and read the factors before you route anything.