# Insight that cannot become an action is entertainment. — Radexus

> The output landed in a spreadsheet nobody opened, because letting software write into the ERP was unthinkable. Here is how to make it thinkable.

*Source: https://radexus.com/journal/write-back/*

---

[The Radexus Journal](https://radexus.com/journal/)  /  Governance

# Insight that cannot become an action is entertainment.

- Topic: **Governance**
- Published: **31 July 2026**
- Reading time: **6 min**

In this essay
- [Five properties](#h-0)

The output landed in a spreadsheet nobody opened, because letting software write into the ERP was unthinkable. Here is how to make it thinkable.

The most common terminal state for an enterprise AI deployment is not failure. It is a weekly export that three people look at for a month and then stop opening. The analysis was right. Nothing acted on it, because acting meant retyping it into the system where work actually happens.

Write-back is what closes that gap, and it is refused in most companies for an entirely rational reason: nobody will accept software mutating a system of record on its own judgment. The answer is not to argue them out of that. The answer is to build write-back that a cautious CIO would design themselves.

## Five properties

-   A named approver, per play, per field and per value threshold. Not a role, a person, with a queue.
-   A dry run first, showing exactly what would change, before anything is enabled.
-   A complete audit trail: who approved, when, on what evidence, and what the prior value was.
-   A rollback window, thirty days by default, during which any approved write can be reversed.
-   Builder and approver roles that are strictly separate. Our engineers can change how a play works and can never approve a write.

There is a second-order benefit that we did not anticipate the first time we built it this way. The approvals are themselves evidence. Every rejection teaches the system something about which surfaced items were not worth surfacing, and the approval queue becomes one of the highest-quality labelled datasets in the engagement.

**Want to argue with this?**

We would rather be corrected than agreed with. The first meeting is an hour and produces a real result.

[Start a discovery](https://radexus.com/start/)

[Back

### All the radexus journal

](https://radexus.com/journal/)[Next

### We assert the problem instead of asking what keeps you up at night.

](https://radexus.com/journal/first-meeting/)