Auditable rules

A score says how likely. A decision says why.

That means the business can review the logic, change the logic, audit a decision, and explain it to an internal stakeholder.

A typical score

Churn risk: 78%.

No visible basis for the number.

A LiftMiner decision

Eligible because this customer has three completed orders, $412 in historical revenue, is 34 days beyond their normal reorder interval, is subscribed to email, and has not received this campaign in the last 21 days.

Illustrative example.

Evidence

Every inclusion and exclusion carries its evidence.

Each customer record shows the conditions they satisfied, or the reason they did not qualify.

Customer AOverdue reorder policy
ELIGIBLE
Completed purchases
3 (needs 2+)
Beyond expected reorder window
34 days
Historical revenue
$412
Email consent
Valid
Campaign contact in past 21 days
None
Exclusion conditions
Clear

Routed to the client's Klaviyo audience. A random share of eligible customers is withheld as the holdout.

Customer BOverdue reorder policy
EXCLUDED
Completed purchases
Meets policy
Beyond expected reorder window
Overdue
Historical revenue
Meets policy
Email consent
Unsubscribed
Campaign contact in past 21 days
None
Exclusion conditions
Clear

Overdue and historically valuable, but unsubscribed. Never enters a contactable audience.

Illustrative example: one policy, two customers, each decision recorded with its evidence.

What is preserved

The record behind each decision.

LiftMiner is designed to keep an auditable record from the data to the outcome.

  • Data freshnessWhen the source data behind the decision was last updated.
  • ConditionsEach condition in the policy and the value the customer had.
  • EligibilityWhether the customer qualified, and why.
  • ExclusionsWhich suppression or exclusion rule removed a customer.
  • Audience assignmentWhether the customer was contacted or held out.
  • Measured outcomeWhat the customer went on to buy, compared against the holdout.

AI evidence

AI extracts a condition. The source text stays attached.

For the planned AI layer

The goal is not “ask AI to guess what a customer feels.” The goal is to extract a usable condition from text while retaining the supporting source evidence.

See how safeguards apply to AI evidence

Start with an audit of the revenue you are already leaving behind.