Map the process before you pick a model
DRIVE is how I take one business process from a written map to a tested release your team runs. It starts with the process, and the model comes later.
A process is ready for AI when its owner can explain it
Write down the inputs, decisions, tools, exceptions and acceptable output before choosing where AI belongs. That gives the build a target you can measure, and it shows which decisions need a person.
Here is my most-viewed n8n template, mapped that way. A person decides at two points, and the AI drafts between them.
- 1YouSubmit a topic.
- 2AI agentWrites a post for each platform from SerpAPI search results.
- 3GmailDelivers the drafts as a formatted email. Nothing continues until you approve them.
- 4YouApprove again before anything is published.
- 5n8nPosts the approved drafts and their images to Instagram, X, Facebook and LinkedIn.
- 6Email and TelegramBring the results back to you.
The first deliverable is a map like this one and an eval set. The software comes after.
DRIVE takes one process from map to production
-
1Define the process
Interview the people who run it, map each step and record walkthroughs as they happen, so the judgment calls are kept.
Done when the process and the exact output of the first release are written down.
-
2Roadmap the first release
Rank the candidate features by value and feasibility, then pick the smallest release that can prove or disprove the approach.
Done when success and stop conditions are agreed.
-
3Implement the smallest working slice
Choose the simplest design that meets the acceptance criteria, whether that is one model call, an n8n workflow or an agent with MCP tools.
Done when the slice runs end to end on real inputs.
-
4Validate it against an eval set
Run representative inputs and expected outputs on every change. Track wrong answers apart from format problems, because they need different fixes.
Done when the release meets its acceptance criteria.
-
5Evolve it in production
Roll it out to a first group, train the team and watch quality and the cost per completed task. Widen the rollout only after that group meets the criteria.
Done when the next change comes from production evidence.
Your business and the models keep changing, so the cycle runs again.
Every decision needs a name beside it
Every implementation needs someone to choose the work, someone to build it and someone to judge whether the process now works. One person can hold more than one role, but each decision needs a named owner.
AI Visionary
Owns: Picks the business problem, sets the budget, names the owner, approves the target and stop condition, and removes organizational blockers.
Usually held by: An executive such as the CEO, COO or responsible vice president.
Where I fit: 1:1 planning and implementation review. See AI for CEOs.
AI Implementer
Owns: Architecture, integrations, tools, deployment, monitoring and maintenance.
Usually held by: Someone with a software engineering or IT background and context engineering, API integration and production review skills.
Where I fit: On Premium I work hands-on in the build. On the other plans I guide your technical team or advise on the build.
AI Operator
Owns: The process specification, acceptance criteria, rollout and feedback loop.
Usually held by: Someone in operations, product or program management.
Where I fit: I help produce the specification and the eval set with the DRIVE method.
The Implementer may direct AI coding agents, but remains accountable for the code and system behavior. When one person is both Implementer and Operator, write the two sets of acceptance criteria down separately. Technical completion and a working process are different tests.
Saved time needs a destination
A tested automated step can run more often without more people. Compute, review, exceptions and maintenance still cost money.
The people in the process stay in charge of its decisions. Give their saved hours a named destination, such as reviewing exceptions or talking to customers. If nothing is removed or improved, the automation hasn't created value.
Pick one process and name its owner
Tell me which process you want to change and what is stuck. We will name its owner and agree a measurable result before any build starts.