AI governance and release gate
Each AI workflow in use is entered with its mandate, a named responsible lawyer and an evaluation set drawn from your own cases; a change goes live only after it is evaluated and the release is recorded.
A register of every AI workflow, and a gate on every change.
Who it is for
The team that has been asked to say yes to AI across a company and has nowhere to write down what it already said yes to. The first deliverable is usually the inventory, and it is usually longer than anyone expected.
The register is a necessary control of this workflow design for departments running more than one tool or workflow, and it can also be established for a single tool. It can start now from an inventory of existing workflows; an inventory created later does not substitute for approvals and run records kept at the time.
The trigger
A new AI use, or a change to a model, a prompt, a retrieval source or a connector in one that is already live.
An incident also triggers the workflow.
The stages
Register
The workflow gets one entry, with its purpose, owner, the data it touches and the business unit it serves named.
Record the mandate
What the software may read, keep and do is written down and versioned, with its sources listed by name.
Name the responsible lawyer
The lawyer who signs the workflow’s outputs is named, by role and by name. If that name is uncertain, finding it is the first deliverable.
Attach the evaluation set
The evaluation set for that workflow is attached to the entry and carries a version number.
Evaluate
The set is run and its result is recorded as pass or fail. There is no in-between result.
Gate and record the release
The set is re-run before a change reaches live use. A change whose set has not passed is blocked. A named person approves the release, and the decision joins the entry with its evidence.
Review on a cycle
Each entry is revisited on a fixed cycle, and a stale entry is treated as a finding. Every incident closes with a register entry and, where possible, a new example in the evaluation set.
Stage 06 in one picture. This is an illustrative gate, not a result from a live workflow or pilot. Every item is checked before a change reaches live use, and a failing item stops the change.
- Evaluation set re-run (met)
- Citation check (met)
- Mandate unchanged or re-approved (met)
- Responsible lawyer named (met)
- Incident plan current (met)
- Rollback tested (open)
These are the checks of the gate, not a result. No results are published on this site.
What the agent may read, keep and do
May read
- The register and its entries
- granted
- The registered workflows’ mandates and version histories
- granted
- The evaluation results
- granted
- The change request
- granted
- The matter content a registered workflow handles
- not granted
May keep
- Versions and dates
- granted
- Evaluation results against a version
- granted
- Overwrite an existing record
- not granted
May do
- Run the evaluation set and report the result
- granted
- Block a release whose set has not passed
- granted
- Approve a release
- not granted
- Change an entry without a named author
- not granted
- Delete a register entry
- not granted
These are the fields of a mandate, not a real matter. No client work appears on this site.
The agent may read the register, evaluation results, change requests, and registered workflows’ mandates and version histories. It keeps versions, dates and results. It may run the set, report the result, flag missing fields and block a release whose set has not passed. It cannot approve a workflow, release a change, delete a register entry or overwrite a record. A correction is a new entry with a named author. The agent runs the gate; only a named person opens it.
Where a person signs
Stage 06, by the person whose approval a change has to have. If that name is uncertain today, finding it is the first deliverable of the engagement. The agent may block a release and may never approve one.
The approver sees the change itself, the re-run evaluation result and whether the mandate changed. All three must be present before the gate opens.
What the record contains
The register carries one entry per live workflow, and the entry is the artefact a board or a regulator asks for.
These are the fields, not a real matter. No client work appears on this site.
| Field | What goes in it |
|---|---|
| Purpose | What the workflow is for, in a sentence a non-specialist can read |
| Owner | The person accountable for it |
| Mandate | The mandate version in force, its approval date and its permitted sources by name |
| Reviewer | The lawyer who signs its outputs, by role and by name |
| Evaluation | The latest evaluation’s date, set version and result, and the date the set last passed |
| Change | The last change, its reason, author and named approver |
| Incidents | What went wrong, the incident’s closing entry, what was preserved and what was added to the set |
| Workflow | The name, the business unit it serves and the data it touches |
| Transparency | Where AI use is disclosed and the documents that record it |
Entries are written as the change happens and are append-only. A correction is a new entry naming the one it corrects. Section 5 of the operating standard says what that means in practice.
The measure a pilot uses
One measure, chosen by you. The one I would usually propose is the share of live AI workflows with a complete, current entry and a passing result in their latest evaluation. The baseline is whatever the true number turns out to be on day one. That first number is usually uncomfortable. It is still the baseline.
The critical error is a workflow running live without appearing in the register. Tolerance in that class is zero.
The systems it usually sits on
Wherever your policies live now. This is a register before it is software, and a register in a spreadsheet that is kept beats a platform that is not. On the European side, the transparency duties in Article 50 of the EU AI Act have applied since 2 August 2026, and the documents they require are linked to the register. LEXSAS resells none of these and claims no certified connector to any of them.
In Türkiye, the register is linked to the personal data inventory and, where applicable, the VERBİS registration.
The first week
- Day one
- The scope call: who owns the register, which workflows it covers, who has been saying yes, to what, who has to approve a change, and the single measure.
- Days two and three
- The inventory of what is already running, which is usually longer than expected.
- Day four
- The register format is written down, you correct it and the first three workflows are entered.
- Day five
- The gate and the first evaluation set are agreed, the gaps are listed and the pilot dates are fixed.
The engagement brief determines the sequence for each assignment. This is the usual order of work, not a timetable commitment.
Exit terms
If the work ends, the company keeps the following, in an exportable format:
- The register itself and every entry in it
- The register and mandate templates
- The gate definition and thresholds
- The incident response plan
- The gold sets and their evaluation results
The register is the asset, and it is the artefact a board, an auditor or a regulator asks for. It is yours, in an exportable format, from the first day.
Three questions
It is the normal starting position, and the inventory is the first deliverable. It is more useful than any policy written before it.
No. It runs the evaluation set, reports the result and blocks a release whose set has not passed. Approval belongs to a named person, every time.
It is a register before it is software. It can start wherever your policies live now, and the format matters more than the tool.
The controls behind this page are in the LEXSAS operating standard, sections 6, 8 and 10. The reasoning is in AI governance for in-house legal teams: five minimum controls and Evaluating a live AI workflow: the release gate. The other three in-house workflows: legal intake and triage, contract first-pass review and regulatory and compliance operations. All eight workflows are on the workflows page. Get in touch