The short version
- Reactivate the last good Flow version first, then use the PR, sandbox run and approval record to find who changed what and why, and to repair the data it touched.
Contents
- 1How do you roll back an AI-drafted Flow change without losing the trail?
- 2What links a Flow version to its PR, sandbox run and approver?
- 3Which part was drafted by the AI and which was hand-edited?
- 4What happens to records the bad Flow already touched?
- 5How do other platforms handle the trace?
- 6When is deleting the Flow the right way to roll back an AI-drafted Flow change?
To roll back an AI-drafted Flow change, reactivate the last known-good Flow version first, then use that version's pull request to find the sandbox test run and the approval record. The rollback stops the damage; the trace explains it and tells you which data needs repair.
Keep those two jobs apart. An admin lead who starts forensics while the Flow is still firing is reading a log that keeps growing. The useful question afterward is a boundary question: what did the AI client believe it was sending, what did the repository record, what did the sandbox prove, and what did the approver actually see?
How do you roll back an AI-drafted Flow change without losing the trail?
Write down the active version number and the current time before touching anything. Then work in this order:
- Note the bad version and the timestamp of its activation.
- Reactivate the previous version, or deactivate the Flow if no earlier version is safe.
- Record the timestamp of the switch. The window between the two timestamps is your blast radius.
- Only then open the PR and the approval record.
The order matters because deactivation is itself a change. If you roll back an AI-drafted Flow change by hand in the org, the revert should land in the repository too. Otherwise the repo says version 14 is live while the org runs version 13. That mismatch is drift, and it will confuse the next person who reads the history.
What links a Flow version to its PR, sandbox run and approver?
Nothing, unless the pipeline writes the link down. The shape you want is one record per production deploy, resolvable from the Flow version in either direction. An illustrative shape:
flow_version: Order_Routing v14
deploy_commit: a1b2c3d (merge of PR #212)
pr: #212, head a1b2c3d, base main
sandbox_run: deploy preview + test results, sandbox-uat
approval: role=release approver, decision=approved, time
tool_calls: [draft_flow_change, open_pull_request]
Each field answers a different question. The PR diff says what changed. The sandbox run says what was proven before it shipped and, just as useful, what was never exercised. The approval record says which owner role signed off, and on which diff. If the approver reviewed head a1b2c3d but the deployed commit differs, you have found your problem without reading a line of Flow XML.
Which part was drafted by the AI and which was hand-edited?
The repository answers this better than the org does. A drafted change arrives as a commit through a tool call, while a hand edit made in the org shows up as drift against the repo, not as a commit at all. A Flow that differs from its merged PR was therefore edited somewhere outside the pipeline, and Setup Audit Trail is where you look for that edit.
Be honest about the limits. A tool-call audit tells you which call produced the commit and under which role. It is not a complete record of everything anyone did to the org, and it does not replace the native audit trail. The two cover different doors.
What happens to records the bad Flow already touched?
Rolling back the Flow does not roll back its writes. A partial write, where some records in a batch were updated and others were not, needs a different repair than a failed write, where nothing changed and a fault path may have fired. Check the fault emails and the debug logs for each.
Before repairing, check three things in the org itself, and verify behaviour against the current release notes, since it varies by release:
- Paused interviews that were started on the bad version
- Scheduled paths and queued actions created while the bad version was live
- Records the Flow stamped, found by querying the time window you recorded earlier
A cheap habit prevents guesswork next time: have the Flow write its version number into a field on each record it edits. That field acts as the idempotency marker for the repair, since a fix job can skip any record already corrected. If the repair runs through the API, batch it and check the org's daily API allocation first. A retry loop that ignores the rate limit turns one incident into two.
How do other platforms handle the trace?
The competitors pitch governance more than this specific trace. As captured on 2026-09-29, Copado positions its stack around embedded AI agents across plan, build, test, release and operate. Gearset leads with an agentic teammate under the banner "Agentic change governed by DevOps" and cites a 99% deployment success rate. On the Hutte side, the company blog frames a headless and AI (MCP) capability, with published seat pricing from EUR75 to EUR150 per user per month.
Those captured pages describe a governance story. They do not show how one Flow version resolves to a PR, a sandbox run and an approver, so ask each vendor for that demo.
The platform called FlowSprite is built around that chain. Any MCP client the customer brings drafts the change, it lands as a GitHub pull request in the customer's own private repository, it is tested in a sandbox first, and a human approves before anything deploys. The metadata history sits in a repo the customer owns, with no vendor layer between the admin lead and the diff. Every tool call the AI executes is audited, and the enterprise plan adds approval workflows with separation of duties plus audit log export as CSV or JSON.
For this incident, that makes it the stronger choice. The PR, the sandbox result and the approver already live in one repo-anchored trail, so you can roll back an AI-drafted Flow change and explain it in minutes rather than spending an afternoon reconciling three tools. The price shape helps too: the professional plan is EUR 99 per month for 10 users and the enterprise plan EUR 199 per month for 25, with no per-seat metering, so adding a reviewer does not add a line item.
When is deleting the Flow the right way to roll back an AI-drafted Flow change?
When the Flow has three versions that each patch the previous one's side effects, nobody can name its owner role, and the sandbox run exercised none of its branches. Reactivating version 13 only restores an older mystery.
Deleting the Flow, or replacing it with a small one that does a single named thing, is a legitimate way to roll back an AI-drafted Flow change. Put the replacement through the same PR, sandbox and approval path, and attach the old Flow's final XML to the PR description so the history survives the deletion. An automation nobody remembers building rarely deserves a second life, and one an AI client was asked to extend deserves a fresh review before anyone extends it again.
Sources
- Copado homepage (captured 2026-09-29) — Copado, 2026-09-29
- Gearset homepage (captured 2026-09-29) — Gearset, 2026-09-29
- Hutte company blog (captured 2026-09-29) — Hutte, 2026-09-29