Composing your own agentic workflow: what the canvas gives a legal team
The vendor’s roadmap stops being your constraint; your own discipline takes its place. That is the whole trade.
Legal AI: build, buy, or configure?, published here earlier this month, decides which path fits a use case. It stops one question short. When the answer is configure, who draws the workflow? A bought legal product ships someone else's: its stages, its review points, its idea of where a lawyer belongs. Composing your own agentic workflow on a general-purpose canvas reverses that. The intake stage is your intake. The gate where a person signs sits where your risk actually is, not where a product manager guessed it would be.
The canvas matters because it makes the mandate visible. Every tool the agent holds, every source it may read, every place it may write is a drawn connection that a lawyer can point at, question and remove. That is the control the prompt injection piece argues for, delivered by construction rather than by vendor assurance: Meta's Agents Rule of Two caps an agent at two of three risky properties in a session, and on a canvas that cap is something you can see.
For an in-house team, the gain is fit and custody. The workflow mirrors the stages the team already runs instead of bending the team around a product, and the run history sits in your perimeter as a record you can show a regulator or the board. For a law firm, practice groups get matter workflows shaped to their own review depth, and the same records let partners price the work rather than defend the hours.
The honest cost is ownership. A composed workflow is yours to maintain: every model, prompt or connector change needs a release gate, and the workflow definition joins the asset list you would take with you when you leave a platform. That is not a reason to buy someone else's workflow. It is the discipline that makes owning yours defensible.
Published