Problem
The first vendor-procurement validation pass drifted from the implementation-plan-first workflow, which made the second pass a process-design problem.
Product management proof
Turning Validation Work Into A Repeatable Playbook
I turned a one-off validation push into an auditable playbook where an AI agent could do structured validation work while I retained responsibility for review and decisions.
What I owned
I moved the work into WBS-first execution with GitHub tracking, source hierarchy, work logs, quality gates, response monitoring, baseline comparison, and allocation-decision stages.
Result
The validation play became repeatable enough to produce evidence, preserve learning, and end in a decision instead of activity.
Internal issue numbers, tracker details, and work logs have been summarized for public use.
The situation
The vendor-procurement validation work started as a market test: could procurement or subcontractor activation generate stronger signal than the hiring-led positioning?
The first pass exposed an operating problem. The work was directionally right, but it drifted from the implementation-plan-first workflow. That made the second pass a product-operations problem: how do we make the play repeatable enough that it produces evidence, not just effort?
The approach
I helped move the work into a stricter operating system where an AI agent could do repeatable validation work without becoming the decision-maker. The second pass used deterministic GitHub tracking, a synced work breakdown structure (WBS) tracker table, explicit opening and closing comments for each work item, source hierarchy, deliverables, work logs, and clearer stages from send operations to response monitoring to baseline comparison.
The important decision was what the play had to end with. It could not end with “we sent the messages.” It had to end with an allocation decision.
That was the human-in-the-loop design. I was defining the objective, source hierarchy, quality gates, claim boundaries, and decision criteria. The agent could push the validation work forward, but the PM still owned what counted as evidence and what decision the evidence supported.
What I built
- WBS-first validation workflow
- source hierarchy and quality gates
- GitHub tracking and work-log receipts
- send, response-monitoring, baseline-comparison, and allocation-decision stages
- human-in-the-loop review criteria for agent work
Why it matters
This is AI-agent work as product management. The agent could move the validation forward, but the PM still owned the objective, gates, review standard, claim boundaries, and decision criteria.
Result
The validation play became more traceable and reusable. The team could inspect what happened, preserve the learning, and decide whether to continue, cut, or change direction.
What I learned
I learned to record the sources, actions, and results of each test so the team could reuse the findings and inspect how we reached the recommendation.