Draft-only mode is a setting in which an AI agent may prepare work but never complete it. Every capability is held at depth 2: the agent drafts the email, the refund or the case note, and a named person reviews and sends it. Nothing reaches a customer, ledger or system of record without a human decision.
Why use draft-only mode?
There are two common reasons. The first is at the start of a deployment. Before an agent acts on its own, the business wants to see real drafts on real cases, and reviewers need practice deciding them. Draft-only mode gives both, with no action the business cannot catch.
The second is during an incident. If something has gone wrong, such as a source turning out to be stale, a pattern of bad drafts, or a suspected prompt injection through an inbound document, dropping the agent to draft-only mode keeps the work moving while removing its ability to act. That is often better than switching the agent off, which leaves the queue to people who may no longer know the routine. Training a team for that moment is covered in how to train a team to use draft-only mode during an incident.
How does it work in practice?
The agent's depth is capped at 2 across every capability, whatever its certificate would otherwise allow. Each output arrives as a held item for a named reviewer, who may send it as drafted, correct it, or decline it with a reason. Every one of those decisions is filed in the append-only record, so the business can later see how often the drafts were right.
Two things change for the team. Volume through the review queue rises sharply, so rotas must be planned for it. And reviewer attention becomes the bottleneck and the control at once. A draft that looks fine but rests on a weak source is the classic trap; what a reviewer should do when an agent's draft looks right but feels wrong covers it.
An example
Take a hypothetical insurer's claims team. An agent has been settling small glass claims at depth 3. A supplier changes its invoice format, and the agent starts reading the VAT line as the repair total. The claims manager drops the agent to draft-only mode the same morning. Handlers review each settlement draft, correct the amounts, and send. Volume is manageable because glass claims are simple. Once the source is fixed and the agent passes a fresh certification run, routine settlements return to depth 3.
How is it different from shadow mode?
In draft-only mode, the agent's drafts are used: a person reviews them and sends them, so they reach the customer. In shadow mode, the agent's output is not used at all. People do the work as before, the agent works the same cases alongside them, and the two are compared afterwards. Shadow mode tests the agent without affecting anyone; draft-only mode puts the agent's work into production with a person between it and every outcome.
Draft-only mode is also different from the drop to depth 1 that follows a lapse or a certification run that clears no depth. At depth 1 the agent may only read; it does not draft at all. The practical steps for a shadow run are in how to run an AI agent in shadow mode.
Related terms
Held actions, review queues, depth and hypercare all connect to draft-only mode. The guide to AI agent skills by role shows which roles review drafts and what they need to know.
Questions people ask
Who can switch an agent into draft-only mode?
Usually the agent owner, and in an incident the incident commander. The authority should be set in advance and the switch filed in the record.
How long should an agent stay in draft-only mode?
Until there is evidence it can act safely: enough reviewed drafts to judge, and a passing certification run at the depth it will return to.