Write-back safety
How to let software write into a system of record without losing sleep.
Approver models, dry runs, thresholds, audit trails and rollback windows, written as a specification you can hand to a vendor.

Radexus. Institutional memory for revenue. radexus.com


You are usually the person asked to make somebody else's idea safe, on top of a job you already have. This is the specification we would want if we were in your chair, and you are welcome to hand it to any vendor including us.


## The specification

  - Approver granularity. Per play, per field, and per value threshold. A change to a delivery date and a change to a price are not the same risk and should not share an approval path.
  - Dry run mandatory. Every write path runs in shadow first, producing the exact diff it would have applied, for a period you specify. Nothing is enabled until you have read one.
  - Prior value retention. Thirty days minimum. Without the prior value, rollback is a rumour.
  - Rollback window. Any approved write reversible within the retention window, in one action, by your team, without the vendor.
  - Role separation. Builder and approver roles strictly distinct. Vendor engineers hold builder only. Enforce it in the identity provider, not in policy.
  - Rate limits. A cap on writes per play per day, so a misconfiguration is a small incident rather than a large one.
  - Complete audit. Who approved, when, on what evidence, what changed, and what it was before. Exportable to your SIEM.


## The argument for the CFO

Insight that cannot become an action is entertainment. A deployment that produces a weekly export is a deployment that will be abandoned by month six, because the cost of acting on it falls entirely on people who did not ask for it. Write-back is what converts analysis into outcome, and the controls above are what make it acceptable.

[An unexpected benefit] Approvals are evidence. Every rejection teaches the system which surfaced items were not worth surfacing. In our engagements the approval queue is consistently one of the highest-quality labelled datasets produced, and it costs nobody any additional effort.


(c) 2026 Radexus, a Global AI Forum company. San Francisco and Chennai.