Working concept ·
AI in every step, a person accountable for each one
The Enterprise AI Delivery Loop. A working model for AI-assisted delivery in seven stages, built on clear product intent, human accountability, and evidence at every step.
What happens when AI becomes part of the entire development lifecycle, not just the code editor?
Teams already use AI to draft specs, write code, and generate tests, and the speed is real. What breaks is accountability. When one model wrote the requirements and another wrote the code, who decided the customer problem was worth solving, and how does anyone know it worked?
The loop below is my working answer. It builds on existing thinking about the AI-driven development lifecycle and holds every stage to three rules:
- Clear product intent. Nobody builds until the problem, the outcome, and the baseline are written down.
- Human accountability. Every stage has a named owner. AI drafts; a person decides.
- Evidence at every step. Each stage ends with an artifact someone else can check.
The loop, stage by stage
01 Frame
Define the customer problem, business outcome, and a baseline before deciding what to build.
- What AI does
- Sizes the problem from support logs, reviews, and usage data, and drafts the brief.
- Who owns it
- Product owns the problem and the decision to invest.
- Evidence
- An opportunity brief with a measurable outcome.
02 Specify
Turn the bet into requirements that people and machines can both check.
- What AI does
- Drafts user stories and acceptance criteria from the brief, and flags gaps and edge cases.
- Who owns it
- Product and engineering leads agree on the spec and on what is out of scope.
- Evidence
- Acceptance criteria written as checks a test can run.
03 Design
Shape the experience and the architecture before writing production code.
- What AI does
- Generates flow, interface, and architecture options fast enough to compare several.
- Who owns it
- Design and engineering choose an option and own the tradeoffs.
- Evidence
- A decision record: the options considered and why one won.
04 Build
Write the code against the spec, in small pieces a person can review.
- What AI does
- Writes and refactors code, and explains each change in the pull request.
- Who owns it
- Engineers review every change and own what merges.
- Evidence
- Reviewed pull requests linked back to the spec.
05 Validate
Prove it does what the spec says, and nothing it shouldn't.
- What AI does
- Generates tests, runs evaluations, and checks results against the acceptance criteria.
- Who owns it
- Quality and product decide what good enough means for this release.
- Evidence
- Test and evaluation results against the acceptance criteria.
06 Release
Ship in stages, with a tested way back.
- What AI does
- Watches error rates and early signals during the rollout and flags anything unusual.
- Who owns it
- A named release owner makes the go or no-go call.
- Evidence
- A rollout plan and a rollback path, tested before launch.
07 Learn
Compare the outcome to the baseline from Frame, then decide.
- What AI does
- Summarizes usage, feedback, and support signals against the baseline.
- Who owns it
- Product decides to keep, change, or stop, and that decision opens the next loop.
- Evidence
- An outcome-versus-baseline readout that becomes the next brief.
Where it comes from
None of this is theoretical. At The Home Depot, the case for a shared release pipeline started at Frame: a baseline of 17 manual release activities before anyone chose a tool. At AT&T, the network health score was an AI model inside a product whose question, “is it me?”, belonged to people. And GovHub is the loop at the scale of one: I built it alone in a few months, with AI tools doing much of the building and me owning every decision.
What I’m still testing
The loop is deliberately heavy for big bets and too heavy for small fixes. A copy change should not need an opportunity brief. The open question is where the line sits: which changes can skip straight from Specify to Release with automated evidence, and which ones need a person at every gate. I’m testing that on my own products first.