Skip to Content
The Adoption PathReview and Records

Review and Keep Records

This is the eighth and final step in The Adoption Path. It has two separate jobs that are easy to conflate: keeping the records the first seven steps already generate, and periodically checking whether the whole system built by those steps still reflects reality. A firm that only does the first has files but no idea if anything has gone stale; a firm that only does the second has no evidence to check against.

Why it matters

Every step before this one produced something — a list, a map, a matrix, approvals, a policy, workflows, review sign-offs. None of that work protects the firm if it sits unreviewed and ages out of date, and none of it can be reviewed meaningfully if it was never kept. This step is not a new layer added on top of the first seven; it is what makes the first seven still true six months from now instead of a snapshot of what was true on the day they were done.

What actually happens

Keep what the earlier steps already produced

This step does not create new records; it names which of the ones the first seven steps already produce need to be kept, and points back to where each one came from rather than inventing a separate recordkeeping scheme. That means the use-case list from Decide What AI Is For, the duty map from Map the Rules That Apply, the data matrix from Sort Your Data, tool approvals and the terms in effect at the time of each approval, the policy from Write the AI Policy and its acknowledgments, the workflows from Build the Workflows, and the client-facing review sign-offs from Handle Client-Facing Use. How long each of these needs to be kept is a figure — see this year’s figures rather than a period stated here.

Set the review cadence and what it checks

Set a regular cadence, and be specific about what each review actually checks, rather than a general “make sure everything’s fine.” Has a use case changed or expanded beyond what was approved. Has a tool’s terms changed since approval. Is the duty map still current against the firm’s own regulator’s published guidance. Are workflows actually being followed, or quietly drifting from what was written down. Is the client-facing review line from step seven holding under volume, or eroding the way that step warned it could.

Name who reviews, and write down what they found

Name who does the review, and where possible, make it someone other than the person whose own work is being reviewed — a reviewer checking their own approvals is checking less than it looks like. Require the review itself to produce a dated record, the same way the duty map in step two did; a review that happened but left nothing written down is functionally indistinguishable from a review that never happened.

Treat an exam or a complaint as the same review, not a new one

An examination or a client complaint is an unscheduled trigger for this same review, not a separate process invented under pressure. If the firm’s answer to a regulator’s question is being worked out for the first time during the exam itself, the review has already failed at its actual job, which is having the answer ready before anyone asks.

Close the loop back to the step that owns each finding

A finding from the review routes back to the specific earlier step it affects, rather than being patched in place at the review step itself. A tool whose terms changed goes back to Vet and Choose Tools; a use case nobody anticipated goes back to Decide What AI Is For. Fixing a finding at the review step instead of routing it back skips the very check — the matrix, the map, the approval — that made the original decision trustworthy in the first place.

Questions to ask yourself

  • Could you produce, today, the use-case list, duty map, data matrix, tool approvals, policy, and workflows the firm is actually supposed to have?
  • When did the last review happen, and is there a dated record of what it found?
  • Is the same person checking their own work with no second reviewer, anywhere in this process?
  • If an examiner asked tomorrow whether the system was still current, would the honest answer be yes?
  • When a review finds something wrong, does it get routed back to the step that owns it, or patched on the spot?

Common mistakes

  • Keeping records but never reviewing whether the system still matches reality.
  • Reviewing informally, with nothing written down to show for it.
  • Having the same person review their own work with no second check.
  • Treating an exam as the first time anyone looks at whether the system holds up.
  • Fixing a finding at the review step instead of routing it back to the step that owns it.

Which Firm Functions apply

At a larger firm, this review usually already happens as part of a broader compliance testing or internal audit cycle that a compliance function runs on its own schedule. The job is not to build a separate AI-specific review; it is to make sure the records from steps one through seven are the kind that existing cycle can actually pull in and check, rather than sitting somewhere it doesn’t look.

Back to the start

Completing all eight steps once does not mean the firm is finished. The system runs on the cadence set in this step, not on a one-time pass through The Adoption Path. A use case added six months from now re-enters at step one, Decide What AI Is For — not at whatever step feels closest to where the new idea came up.