From idea to product4 min read

Choose What AI Should Do

How to choose an AI role in a workflow, set its authority and design for useful human control.

A person directs an inspectable AI work process using visible controls and a manual override.
On this page
  1. 1.Compare the complete workflow
  2. 2.Set authority deliberately
  3. 3.Design the failure path
  4. 4.Test before expanding

A team wants to reduce the time customers wait for a support response. It could use a model to draft a reply, give an agent permission to retrieve account details, or let the system send certain replies without a person checking each one.

Those choices look similar in a demo. They create very different products. Each changes what the system needs to know, which mistakes matter and what the team must be able to detect and repair.

Start by describing the outcome and the current work. Then decide which part of that work is suitable for AI, and which responsibilities should stay with people or conventional software.

Compare the complete workflow

A model can help interpret varied language, generate a draft or work through an unfamiliar request. A deterministic rule can enforce a known threshold or prevent an action outside a permission boundary. A person can resolve ambiguity, apply domain judgement and accept responsibility for the decision.

Consider the full path rather than the generated output alone. For the hypothetical support workflow, the system may need to find the current policy, access the right account, recognise exceptions and hand an uncertain case to a person. A fluent reply is not useful if it applies the wrong policy.

Compare the proposed approach with the current process and simpler alternatives. Include customer effort, accuracy, review, latency, integration, operating cost and recovery. A small improvement inside an existing workflow can be worthwhile; a redesigned workflow may deliver more value, but also introduces more to operate.

Set authority deliberately

Authority is the set of actions a system is allowed to take. It is distinct from capability: a model might be able to draft a message but should not necessarily send it.

ActionA possible boundary
Draft a responseKeep it editable and make the source material available to the reviewer.
Retrieve customer informationRestrict access to the data needed for the task and the current user.
Send a responseRequire approval or use a narrower automatic path supported by evidence and controls.
Change an accountApply independent permission checks, keep an audit record and provide a recovery path.

These are examples, not universal approval rules. Choose controls based on consequence, reversibility, quality evidence and the ability to detect failure. An approval button only helps when the person has enough context, time and authority to make a meaningful decision.

Design the failure path

Ask what happens when a policy is missing, a tool is unavailable, context is stale or a request contains instructions that conflict with the user's task. Decide when the system should stop, ask for clarification or hand work to someone else.

Place consequential limits outside the model: permissions, spending or transaction limits, and checks before an external action should not depend solely on an instruction in a prompt. Preserve enough history for an owner to understand and recover from a failure.

The security playbook covers those boundaries in more detail. AI UX addresses how people see progress, intervene and recover.

Test before expanding

Use representative cases to see whether the system performs the task under the conditions that matter. Include ordinary requests, exceptions and cases where it should decline or escalate. Assess both the result and, when the system can act, the path it took.

A pilot can answer questions that an offline test cannot: do people use the output, can they review it, what happens to support load and does the cost make sense? Start with a boundary that makes failures observable. Expand only when the evidence and the team's ability to operate the workflow support the change.

The evaluation chapter explains how to build the test set. Business viability covers the cost of running the whole workflow, including review and recovery.

v3.1 · Updated Oct 2026