mickey.bushman
menu

// case study — ai-augmented product management

How I run a multi-person product scope solo

I was drowning in overtime covering four product areas alone. Instead of working more hours, I built an operating model: a coordinated team of specialist AI agents with contracts, quality gates, and one human in charge.

01

The problem

My scope was the workload of a multi-person product team: several extremely complex products, each demanding detailed specifications with rich user-interface walkthroughs to keep high-performing development teams continuously fed; ongoing market research; a living product roadmap; executive-level presentations; contract reviews. All of it owned by one person — me.

The math didn't work. I was covering it with extreme amounts of overtime, and overtime doesn't scale. Worse, being spread that thin meant quality was living on borrowed time: every hour spent writing a spec was an hour not spent validating one.

02

Why “just use AI” wasn't the answer

A single general-purpose AI assistant helps with volume, but it fails in a specific, predictable way: it has no boundaries. Ask it for requirements and it will happily pick your database. Ask it for a UX flow and it will rewrite your business rules along the way. Output goes up; trust goes down — and untrusted output just moves the bottleneck to review, which was exactly the hour I didn't have.

The insight was that my problem wasn't a writing problem, it was an organizationalproblem. What I was missing wasn't a faster pen — it was a team. So I designed one.

The problem wasn't a faster pen.
It was a missing team.
03

The design: an org chart of specialists

I built a multi-agent system modeled on how a well-run product organization actually works: one coordinator that only routes work, and eight specialist agents — Product Manager, Business Analyst, Project Manager, UX Expert, Solution Architect, Development, QA, and Technical Writer — each owning exactly one artifact type.

Agents own outcomes, not implementations.

Every handoff carries an objective, constraints, and acceptance criteria — never instructions on how. The receiving agent is the expert in its domain and rejects anything that crosses that line.

Never trust the previous agent.

Every deliverable is independently validated on receipt against four checks — completeness, consistency, ownership, authority. Verification is mandatory at every handoff, not optional.

Corrections return to the owner.

No agent ever edits another agent's work. QA files defects; it doesn't fix code. Ownership never changes hands.

Iteration caps and criteria-bound acceptance keep review loops from ping-ponging forever, and a named human — me — is the circuit-breaker for anything irreversible or high-risk.

One boundary matters more than all the others: the agents don't decide what to build. Customer judgment, prioritization calls, and scope trade-offs stay with me — the PM agent drafts briefs and enforces scope discipline, but product direction is a human call. The agents produce; I direct, judge, and own the results.

04

Right-sizing the intelligence

One deliberate design decision that pays for itself daily: each agent runs on the smallest model sufficient to do its job at a high level — and nothing more. A routing coordinator doesn't need frontier-model reasoning; a solution architect weighing non-functional trade-offs does. Matching model capability to role keeps costs proportionate, responses fast, and — counterintuitively — quality up, because an over-powered model given a narrow job is exactly the one most tempted to overreach its lane.

05

The results

The overtime is largely gone. In its place: significantly higher output and higher quality — detailed specs with UI walkthroughs that keep development teams continuously supplied, current market research, an actively maintained roadmap, and executive presentations, all produced at a pace no single unassisted PM could sustain. I now carry what is normally a two-to-three-person product scope as one person, without the burnout math that used to make it possible only on paper.

This isn't a thought experiment. This operating model runs my day job. The linked guide is the actual system — charters, handoff contracts, validation checks, and loop controls — written so any team could build it.
06

What I learned

Boundaries beat intelligence.

The biggest quality gains came not from better models but from scoped authority — agents that know what they're not allowed to decide.

Verification is the product.

Generation is cheap; trust is expensive. Building independent validation into every handoff is what turned raw output into work I could put my name on.

Loops must be bounded by design.

Left alone, agents will politely reject each other's work forever. Iteration caps, criteria-bound acceptance, and a human escalation endpoint aren't nice-to-haves — they're what makes the system converge.

Be honest about enforcement.

Some guarantees are real mechanisms (scoped tool access, automated gates); others are conventions held by well-written charters. Knowing which is which is the difference between a demo and an operating model.

07

The roadmap: treating the system like a product

The operating model isn't finished — and I manage it the way I'd manage any product: with an honest assessment of its gaps and a prioritized roadmap. Currently on it:

Instrument everything.

The state layer already tracks first-pass yield, rework rate, and cycle time per work item. Next is baselining and trending them — so iteration caps get tuned from data, and the system's ROI is provable rather than felt.

Harden conventions into mechanisms.

Rules held by well-written charters should graduate to rules held by the platform: write-scoping and pre-write hooks on every agent, and schema validation on every handoff contract so a malformed handoff is rejected by a script, not by judgment.

Structured state.

Lifecycle state moves from free-form markdown to structured data updated through a small CLI — one bad edit shouldn't be able to drift the whole orchestration.

Regression-test the agents themselves.

A charter edit is an untested deploy. A set of golden work items gets replayed after any charter or steering change to catch behavioral drift before it reaches real work.

Close the loop post-ship.

The pipeline currently ends at documentation. An outcome-review step — did the success metrics actually hit? — will feed results back into the roadmap, so the system verifies not just built right but worth building.

Re-bid the models periodically.

Right-sizing decays as models improve and cheapen. Tracking cost-per-artifact per agent puts model assignments on a review cadence instead of leaving them frozen at whatever was optimal on day one.

Want the full system? The complete operating model — agent charters, the handoff contract, the four validation checks, loop controls, and a step-by-step build guide — is available as a downloadable PDF.

download the guide (pdf) ↗