AI agents
How to choose the right workflow for an AI agent
· 6 min read
The first question in an AI-agent project should not be “which model should we use?” It should be “which piece of work is valuable, repeatable, understandable, and safe enough to improve?”
Ambition is not a use case
Many AI initiatives begin with a broad direction: automate support, improve sales, build an internal assistant. These may well be worth pursuing, but none of them is yet implementable.
A production agent needs five concrete things — a defined trigger, trusted context, permitted actions, a completion condition, and a clear point at which a person takes over. Until a proposal can name all five, it is a theme rather than a project.
1. Is the workflow repeated often enough?
Agents create the clearest value where a task occurs frequently. Repetition is what makes it possible to observe the current process, evaluate quality against real examples, and measure afterwards whether anything actually improved. A rare, high-stakes decision is usually the wrong place to start.
2. Can the work be described as steps?
A promising workflow has a recognisable path: information arrives, context is gathered, a judgment or transformation happens, an action is taken, and the result is recorded. If nobody can describe that path, the problem is process clarity, not AI.
3. Does the required information exist?
An agent cannot reliably use knowledge that is missing, inaccessible, outdated or contradictory. Data readiness decides more outcomes than model selection does, and it is usually the constraint people discover last.
4. Can the agent's authority be bounded?
A production agent should know what it may read, what it may change, and what requires approval. Start with narrow permissions and widen them only after evaluation gives you a reason to.
5. Can success be measured?
Pick a baseline before you build: handling time, review time, completion rate, escalation rate, error rate, user satisfaction. Without one, you will not be able to tell whether the agent helped or simply moved the work somewhere less visible.
A practical starting pattern
The most reliable first agent usually follows the same shape:
- 01Receive a well-defined request.
- 02Retrieve information from approved sources.
- 03Prepare a recommendation or draft.
- 04Ask a person to approve a consequential action.
- 05Record the decision and outcome.
In short
The right first AI agent is not the one with the most autonomy. It is the one attached to a valuable workflow, grounded in reliable context, limited by clear controls, and measured against a real operating baseline.
Autonomy is something you earn from evaluation, not something you specify up front.