CRM Enrichment With Change Review
Works with
Implementation kit
- The spend gates, with the caps worksheet
- The research prompt, citations required
- The overwrite rules, freshness and survivorship
- The change-review checklist
- The golden-records eval guide
An incomplete record becomes a preview of proposed updates, every claim carrying its source, with a person on the approve step and an undo behind every change.
The full method is below. The pattern behind it turned an estimated twelve to twenty hours of account research into a thirty-minute review in our own client work.
An estimated twelve to twenty hours of account research became a thirty-minute review of finished research.
From our signal-detection case study
How it works
01Gate the spend
Enrich only records something downstream will use, capped per day. An enrichment nobody reads is a bill with no customer.
Records that miss the gate queue rather than burn. The queue drains inside the cap, and queued records expire and re-qualify, because the reason a record qualified goes stale too.
02Research with citations
The agent works public sources and every claim carries where it came from. Unknown stays unknown, never guessed.
A cited claim can be audited in one click. An uncited claim is a rumor in a database.
03Date every fact
A title verified last week beats one scraped last year. Every fact carries its date, stale evidence gets suppressed visibly instead of masquerading as current, and a pre-existing value with no date is treated as challengeable, because undated data lies.
Watch for the field that is confidently wrong. Undated data does not age, it lies.
04Propose, never overwrite
Silent overwrites are how trust dies. The agent produces proposed changes. Old value, new value, source, side by side, with conflicts and unknowns rendered on the same screen and suppressed proposals shown collapsed, because a rule that says never silently resolved needs a place on the screen. The record itself has not moved yet.
05Review the changes
A named person drains the review on a cadence, and approval takes seconds because the preview did the work. Approve-all covers high confidence only. A field class graduates to auto-apply only when the approval streak AND a passing golden-records score both hold, and consequential fields never do.
06Keep it reversible
Every applied change is logged with its prior state. Undo restores only when the current value still equals what the change wrote; a later human edit turns the undo into a conflict for a person, never a clobber. An enrichment pipeline you cannot undo is a pipeline you will eventually fear.
07Evaluate on golden records
Pick five accounts you know cold and write down the truth about them. Score the agent with the 5x arithmetic (correct minus five times wrong minus five times hallucinated, over findable fields) against a pass bar you wrote down first, re-score whenever the prompt, model, or sources change, and grow the set from your rejections, because failures point at the fixtures you need.
The eval is the difference between an agent you tested once and an agent you can keep trusting.
An enrichment is a claim. The record should hold evidence.
The implementation kit above packages the gates, the prompt, the rules, and the golden-records eval.
