Short answer
A staged rollout for AI agents releases an agent in steps, from watching only to test senders, a share of conversations and then everyone, with results checked before each step widens. In AI, rollout simply means putting a model or agent into real use.- Also called
- Staggered rollout, phased rollout, gradual rollout, canary release
- Stage by
- Share of traffic, channel, request type or customer group
- First stage
- Watching only: the agent decides and drafts but sends nothing
- Rolls back by
- Turning the stage off, or returning to drafts
An AI agent that does well in testing still meets new wording, new edge cases and real customers on its first day live. A staged rollout makes that first day small. Problems show up while they affect a handful of conversations, and the team widens only what has earned it.
What rollout means in AI
In software, a rollout is the release of a new version to the people who will use it. In AI, rollout usually means the same thing: putting a model or agent into real use, with real customers.
A staggered rollout releases to groups one after another, a region or a share of users at a time. In reinforcement learning the word has a second meaning: a rollout is one simulated run of an agent through its environment, used to estimate how good its choices are.
The stages of an AI agent rollout
Most teams move through five stages, and can stop or step back at any one.
- Watch only. The agent reads live conversations and decides what it would do, but sends nothing. The team compares its drafts with what people actually sent.
- Drafts. The agent writes replies for teammates to review, edit and send.
- Test senders. The agent answers live, but only for internal testers or a short list of customers.
- A share of conversations. It answers a set percentage of real traffic.
- Everyone. It answers all conversations in scope.
In Aide, the drafts stage happens inside the helpdesk the team already uses: drafts land as internal notes in Zendesk, as Front drafts, or on the Aide card in Gorgias with an insert-draft macro, and they learn from the team's macros and past replies.
Staging by traffic or by request type
Most software rolls out by traffic share. AI agents roll out better by request type, because an agent's performance differs sharply between requests.
A Shopify store's agent might be ready to answer order status on its own in week one, while refunds and damaged items stay at drafts for a month. Rolling out 10% of all traffic would mix both, and hide the difference. Rolling out by request type keeps each decision separate: order status goes to everyone, refunds stay with the team until their drafts are reliably sent unchanged.
The approach this page rejects is the all-at-once launch: switching an agent on for every request on day one and judging it by the first month's averages.
What to check before widening
Agree the exit criteria for each stage before it starts, so widening is a decision, not a feeling.
- Drafts sent unchanged. How often teammates sent the agent's draft as written.
- Handoffs. Why the agent passed conversations on, and whether those reasons were right.
- Customer outcomes. Satisfaction and repeat contacts on the conversations it handled.
- Actions. Every order change or refund it made or proposed, checked by a person.
With Aide, Ask Aide reads the conversations where the team rewrote or replaced Aide's draft, grades the draft against what was actually sent, and proposes the smallest fix, such as an article, an instruction or a procedure, then retests it and shows the new answer.
How long each stage should last
Long enough to see the request types that matter, which depends on volume more than days. A busy store may see enough refunds in a week; a quiet one may need a month. Peak periods such as Black Friday are a poor time to widen, because volume rises and the mix of requests changes at once.
Where Aide fits
Every helpdesk inbox connected to Aide moves through Watch only, where Aide decides and drafts in its Activity log without writing anything to the helpdesk, then test senders, a share of conversations and everyone. Separately, each of the team's topics has an AI level, AI Agent, Drafts or None, and new topics start at Drafts, so order status can run on its own while refunds stay with the team. See how to add AI to your helpdesk.