AI implementation starts with the work.
Choose a bounded workflow before choosing a model.
All perspectives · · 2 min read
Describe the decision, not the tool.
A useful AI brief begins with the work someone needs to finish. Describe the incoming request, the information required, the decision to make and the destination for the result. Identify the person who remains responsible when the information is incomplete. This makes the scope concrete enough to estimate, review and test.
A support reply and a supplier payment may both involve reading documents, but their permissions and consequences differ. Define the boundary of the workflow before treating them as the same automation problem.
Separate context, judgment and execution.
List which business facts are stable, which records change and which actions affect customers or money. An agent needs access to the right context, but access to information does not automatically justify permission to act. Agree on the points where someone reviews a proposal or approves a change.
Model choice follows these requirements. OpenRouter distinguishes model routing from provider routing; the decision includes availability, capabilities and delivery characteristics, rather than a single model name.
Ask for a reviewable first release.
The first delivery should show an input, the proposed action, the resulting record and a useful failure path. Review realistic exceptions alongside successful examples. A short demonstration of an easy case says little about how the workflow behaves when a record is missing or a request is ambiguous.
Agree on a baseline, a small set of success measures and who will monitor the result. Expand the scope after the first workflow produces evidence that it is useful.
Further reading.
Put the brief into practice.
AI integration · Discuss the work your business needs to do.