How do you roll out an AI agent gradually?

Widen one dimension at a time. A gradual rollout moves along three dials: which slice of work the agent sees, what share of that slice it handles, and what depth it holds. Turn one dial per stage, write down the evidence that lets you move on, and keep the earlier stage ready to return to if the evidence turns.

What are the dials in a gradual rollout?

• Slice: a product, a case type, a region, a team, a customer segment. Start with the slice that is common, low in consequence and well documented.
• Share: the proportion of that slice routed to the agent. A tenth first, then half, then all.
• Depth: read, draft, execute within bounds. The agent earns its depth through certification.

Changing two dials at once is the most common mistake. If you add a new product line and raise depth in the same week, and the error rate rises, you cannot tell which change caused it, and you cannot roll back one without the other.

What does each stage need?

A new slice starts again at shadow, even if the agent holds depth 3 elsewhere. Certification rests on the cases it was run on, and a new product brings cases the agent has not seen.

Who acts at each step?

1. Process owner: chooses the first slice and writes entry and exit criteria for each stage. Evidence: a signed rollout plan.
2. Agent, depth 1: runs in shadow on that slice. Evidence: every proposal filed with its sources.
3. Reviewer: scores shadow output against what people actually did. Evidence: scored cases.
4. Platform owner: sets routing rules for share and slice. Evidence: a versioned routing configuration.
5. Process owner: moves stages only on written exit evidence. Evidence: a filed decision for each move.
6. On-call owner: returns to the previous stage if a trigger fires. Evidence: the trigger, time and person.

What does a staged rollout look like?

Say a bank introduces an agent to refresh customer due diligence files. The process owner picks retail customers rated low risk, with a single account and no change of address, as the first slice. The agent gathers documents and drafts a refresh summary in shadow for four weeks, then drafts for a tenth of the slice while analysts decide.

Once the edit rate settles and the held queue clears daily, the share goes to all low-risk retail files. Only after a further certification does the agent close files at depth 3 where every document is current and nothing has changed, still for low-risk retail alone. Business customers and higher-risk ratings come later, each starting again at shadow. Whether and how an AI agent can carry out KYC checks at all depends on that evidence.

The same pattern works in finance. An agent that flags out-of-policy expenses might start with one department's travel claims before moving to all claims, and only then to approving low-value ones.

Which controls keep each stage reversible?

• Written return triggers for each stage: an edit rate above a set level, any scope breach, a complaint linked to the agent.
• Routing that can be changed in minutes, so share can fall to zero without a release.
• A versioned bundle per stage, so rolling back the agent restores the right brief, sources and bounds. The detailed steps are in rolling back after a bad change.
• Recertification before any depth change, never as a formality after it.

The full sequence is in our guide to deploying an AI agent.

Where Pryme fits

Pryme grants authority per agent and per customer, proven on that customer's own cases. Attest certifies each agent at a depth, and a new slice of work is a new set of cases to prove. Routing share and depth are separate settings, so you can move one dial without the other.