Write-back: returning a corrected number to the system it came from

What operational write-back is, how it differs from reverse ETL, and the controls it needs.

Converge Team, 2026-11-12

Write-back is the path by which a change made in a data platform reaches the source system the data came from. Someone sees that an order value is wrong, corrects it where they see it, and the ERP that owns the order ends up with the corrected value.

Most data platforms only read. They copy data out of operational systems, and a person who finds a mistake in a report has to open the source system, or send an email to someone who can. Write-back closes that gap, and because it changes a system of record, it needs more control than anything on the read side.

Write-back and reverse ETL

The two are often confused. Both send data from an analytical platform to an operational one.

Reverse ETLWrite-back
Unit of workA table or a segment, on a scheduleOne change, made by a person or an agent
TriggerA sync jobAn action someone applies
Typical usePush a computed score to a CRMCorrect a record, change a status
Who is accountableThe pipeline's ownerThe person who applied the change, and the approver

Reverse ETL moves results. Write-back carries decisions. A platform can offer both; they need different controls.

The path a change takes

In Converge, for a source added through Connect, a change does not go to the source directly. It passes four steps.

  1. An action. The change is a named action type with declared parameters. Submission criteria decide whether this person may apply it with these values, and postconditions check the result before the edit is committed in Converge.
  2. An approval, where required. An action type can require a second person. Applying it then creates a pending request, and the data does not change until an approver accepts it.
  3. A write grant and a write policy. Being allowed to read a source does not allow writing to it. Writing needs its own grant, and each source has its own write policy.
  4. An audit entry. The write-back is recorded with the actor, their roles, the operation and the resource, in a log where each entry carries a hash of the one before it.

For these sources, a direct push that skips the action or the approved export is refused.

Limits worth knowing

  • An export is approved as an export. It does not pass through an action type.
  • An action can also notify an external system through a webhook. The webhook is sent before postconditions run, and such an action cannot require approval, so a webhook is the wrong tool for a change that needs a second person.
  • Sources set up by hand, outside Connect, do not have the separate write grant and write policy.

Why the extra steps

A wrong value in a dashboard misleads the people who look at it. A wrong value written to an ERP becomes an invoice. Three properties follow from that:

  • Validation before the write. The source system should never receive a change the platform could have rejected.
  • Separation of duties. For changes that matter, the person who proposes is not the person who approves.
  • A record that survives. Months later, "who changed this, and who agreed?" must have an answer that nobody could have edited.

AI agents use the same path. An agent changes data only by applying an action, its action tools wait for a person to review the call, and the action's own approval still applies afterwards.

Questions to ask

  1. Can anything write to a source without going through a declared action?
  2. Is write permission separate from read permission, per source?
  3. Can an action require a second person, and is self-approval blocked?
  4. Where is the write recorded, and can that record be altered?

Read more

Converge