Before AI Counts Your Customers, Define “Active”

A customer success lead asks an AI assistant how many active customers the business has. The assistant counts accounts with a current subscription. A delivery manager asks the same question and gets the number of accounts with open work.

In this hypothetical business, both calculations could be correct. They answer different questions. Choosing a more capable model would not resolve the disagreement about what active means.

On October 15, AtScale announced AtScale for AI, a suite connecting business definitions, question clarification, computation, and evaluation. AtScale says Ava and its AI Computation Engine are available now, while CoModel and Canary are coming soon. The company describes clarifying a user's intended metric before calculating an answer. These are vendor statements, not capabilities we have independently tested.

The business decision is broader than this product: who owns the meaning of the numbers an AI workflow produces? Our recommendation is to agree on a small set of decision-specific metrics before making natural-language analytics part of daily operations.

Start with the decision behind the count

The useful definition of active depends on what someone intends to do with the answer.

For planning delivery capacity, accounts with work scheduled during the coming month may be relevant. For a customer engagement review, the team may need accounts that used a service during a defined period. A subscription status can be useful for a different purpose without settling either of those questions.

Ask the process owner what decision follows the count. Then write the population and time period in plain language. Define whether the unit is a customer organization, an account, or an individual user. Those units can produce very different totals even when they come from the same underlying records.

Avoid forcing every department to use one vague label. Where legitimate differences exist, give the metrics distinct names. Accounts with scheduled delivery and accounts with recent service activity are easier to interpret than two competing definitions of active customers.

Make the definition reviewable

A metric needs enough detail that another qualified person can reproduce its meaning without reading the analyst's mind.

Record what qualifies an account, which records supply the evidence, and the relevant time boundary. Specify how the count treats paused accounts and internal test records. If several records represent one organization, decide which identifier establishes the counting unit.

Assign an owner who can resolve disagreement about that meaning. Engineering can implement the calculation, but it should not have to decide whether a paused engagement still belongs in a customer success review. The person accountable for the business decision should approve that rule.

Keep the definition close to the calculation. Where an existing analytics model already expresses the approved rule, connect the AI workflow to that calculation rather than recreating it separately in each assistant. Verify that the connected tool actually applies the rule. A link to a definition document does not establish that the query followed it.

This is a different problem from cross-system source coverage. A workflow can check every required system and still calculate the wrong population because its interpretation of the question differs from the user's.

Clarify when the ambiguity changes the decision

An assistant that asks a question after every request can become tedious. An assistant that silently chooses between materially different metrics can become misleading.

Use approved defaults for a defined workflow, and make those defaults visible in the answer. A delivery planning assistant might report accounts with work scheduled next month because that is the agreed measure for its role. An assistant serving several departments may need to ask which population the user intends.

Keep the clarification concrete. Ask whether the user means accounts with current subscriptions or accounts with scheduled work. Asking the user to clarify active without offering the relevant distinction pushes the same problem back to them.

If the requested meaning has no approved calculation, identify that gap. The assistant can offer to help prepare a proposed definition for review. It should not quietly turn a convenient interpretation into the company's operating standard.

Test membership as well as the total

A plausible total can hide the wrong set of accounts. Two calculations can return the same number while including different customers.

For an initial check, have the metric owner review a small set of records whose expected treatment is known. Include ordinary qualifying accounts and nearby cases that should be excluded. In the hypothetical delivery workflow, test a paused engagement, an internal test account, and two records belonging to the same customer organization.

Check the time boundary explicitly. Work scheduled just outside the reporting period should not slip into the count because the query used a broader date filter. Decide how the workflow handles missing schedule information instead of assuming that missing means inactive.

Compare the included accounts with the expected set, then examine the total. This is a focused business check, distinct from asking another model whether the answer sounds reasonable. As we discussed in validating AI quality scores, an evaluation needs an operating standard that a qualified reviewer can apply.

Keep a definition change from becoming a growth story

Changing a metric's meaning can change its trend even if the business itself has not changed.

Record when a revised definition takes effect and identify the version used in recurring reports. Decide whether to recalculate earlier periods on the new basis or show a clear break in the series. The choice should be explicit so readers can compare periods appropriately.

Keep answers compact but interpretable. Include the metric's descriptive name, period, and a route to inspect the calculation or supporting records. Surface any limitation that could change the decision. Staff should be able to tell what was counted without opening a long technical log.

For a first implementation, choose one recurring question that currently produces disagreement. Name the definition owner, settle the counting unit, and test the boundary cases before expanding the question catalog.

If your team wants AI to answer operational questions with numbers people can use, talk with Revival Group about connecting one business decision to an agreed definition and a verified calculation.