Let the problem set the direction
An AI conversation can quickly become a discussion about agents, models and orchestration. Our starting point is different: what is the business trying to change, who is doing the work and what is making that work difficult?
Describe the current workflow
Document review can slow down because of extraction, repeated information requests, inconsistent interpretation or unclear approval ownership. Define which of these problems the workflow needs to solve before choosing its technology.
Define the intended change
A useful requirement explains the users, the task, the inputs, the expected result and the boundary of the work. It also states what should remain under human judgement. This gives business and delivery teams something concrete to review.
Separate assumptions from evidence
Expected time savings, source quality and integration feasibility may initially be assumptions. Write them down as assumptions. Then identify what would validate them and which decision depends on the result. The point is not to eliminate uncertainty before starting; it is to stop uncertainty becoming an invisible commitment.
Choose the simplest suitable structure
Some parts of a process may need conventional automation. Others may benefit from AI-supported extraction or synthesis. Some should remain human decisions. We prefer a design that serves the requirement rather than one that maximises the number of agents or tools.
Make the first increment decision-ready
A first scope should be small enough to understand and meaningful enough to evaluate. Define the task it will complete, the dependencies it needs and the criteria for deciding what comes next.
The question to keep returning to
“What useful change will this make to the work?” is a better organising question than “How much AI can we put into this process?” It keeps the technology connected to the reason the initiative exists.