Start with the task
An AI agent is a software system that uses an AI model to work towards a goal. Depending on its design and permissions, it may choose steps, use tools, review intermediate results, and adjust what it does next. The word “agent” does not tell you which systems it can access or which actions it is allowed to take. Those details need to be defined for each implementation.
For a business team, the useful question is concrete: what piece of work should this support, and what would a good result look like? A clear task is a better starting point than a broad request to “add AI”.
Agents and workflows
A fixed workflow follows a sequence chosen in advance. An agent can be designed to make more of the decisions about the next step. A project may combine both: predictable steps around a more flexible piece of reasoning. This distinction is explained in Anthropic’s guide to building effective agents.
More flexibility also creates more choices to check. Begin with the simplest approach that can deliver the required result. If a clear set of rules solves the problem, there may be little reason to add open-ended decision-making.
Define a useful boundary
Write down the input, the expected output, and the point at which a person takes over. “Help with projects” is difficult to evaluate. “Turn an approved set of project notes into a draft brief, with open questions highlighted” is much easier to discuss and review.
Decide what belongs outside the task as well. A brief-drafting pilot could read the information you provide and prepare a document. Sending that document to customers, changing deadlines, or assigning people to work would be separate decisions requiring their own scope and permission.
Keep review visible
Give someone clear ownership of checking the result. The reviewer needs enough context to recognise a missing fact, an unsupported conclusion, or an action that should not happen. They should also know when to stop the workflow and ask a colleague.
For an early pilot, keep consequential actions behind an explicit approval step. Define what approval means in practice: who reviews the work, what they check, and how a correction gets incorporated. “A human is involved” is not a complete process.
An example to discuss
Illustrative workflow
A team supplies approved discovery notes and its existing brief template. The workflow produces a draft brief, separates confirmed facts from assumptions, and lists unanswered questions. A project lead reviews it before it becomes the team’s working plan.
This example is a way to frame a conversation, not a promise that every implementation has the same capabilities. The information sources, permissions, tools, and review steps must fit the actual work.
Choose the next step
Pick one recurring task with a visible result and a person who can judge its quality. Gather a few representative examples, including an awkward one. Then describe the boundary and the evidence you will use to decide whether the pilot is useful.
Continue with choosing your first AI workflow, then use the pilot evaluation guide to define your measures before you begin.
Let’s start with your work.
A goal, a challenge, or an example is enough to begin exploring what useful support could look like.
