02 · work · ai in delivery
AI in delivery: a lab, a workflow and three projects
I tested AI delivery in a four-week lab, then led the staffing and workflow changes for three client projects. All three delivered full scope within committed dates and projected financial margin with one engineer working with agents instead of four developers.
Role: VP of Engineering, Teravision
The lab
I started the labs. The bench had idle time between placements, so I used it. I began with a document that stated my hypotheses and how we would measure them, using the evidence-based management approach from my PAL-EBM training. Then I defined one project, chose three squads based on what I knew of their technologies, set up spaces for them to collaborate and gave them four weeks.
All three squads built the same tool, so we could compare the results and the code quality. One used Flutter and Node, one used Next.js full stack and one used .NET with React. The stacks ranged from familiar to new on purpose.
All three finished. The best result came from the squad with the most senior engineers. They gave the AI room to build, then reviewed the output against the requirements. The other squads paced the AI step by step. Some engineers did not trust the output and checked every move, and that cost speed. Where their knowledge of the stack was thin, the extra caution did not improve the quality.
It was one experiment with three squads over four weeks. Seniority and stack varied across the squads, so I read the result as a signal and tested it next on real projects.
The workflow
I used the lab findings to change staffing and introduce AI tools on client projects, then checked whether the new way of working was actually being used. The first iterations fell short. Training and conversations helped, and some engineers still redid by hand what the AI had produced. Development time barely moved. I found out by asking how each increment was built. I tracked adoption with a short checklist and treated it as something to measure.
The teams evolved to a fully agentic workflow where one engineer runs the whole process: planning, premortem, automated tests, independent code and security review, and human approval before release.
The result
Three client projects moved from teams of four developers to one engineer working with agents. All three delivered the full scope within the committed dates and the projected financial margin. Scope and complexity matched what the earlier four-person teams had handled. Defects in the development cycle fell by about 40%, user acceptance quality improved, and the same setup carried into maintenance. The numbers cover scope, dates, margin and defects. They do not include the time spent on review and oversight.
The engineers delivered the work. My part was the staffing decision and the design of the workflow.
From the lab to a framework
What we learned in the labs became the starting point for the Cognitive Leader Framework, which our CEO wrote up in a book. I refined the concepts for implementation and built the platform that organizes it: the content and activities, progress tracking across sessions and a PMO module for following each client’s implementation. I also co-deliver weekly executive workshops with our CTO, about twenty sessions across two engagements in twelve weeks, one with a paying client and one with a partner organization.