Discovery
Turn raw client noise into a structured project. Stakeholders, success metrics, constraints — captured once, reused everywhere.
AI PM
What an AI-augmented PM actually does.
Not 'product manager who uses ChatGPT'. A PM whose workflow is built around agents, skills, and a stack that compounds week over week.
Planning
ADRs that actually get read. PBI lists that survive the week. Roadmaps tied to revenue, not story points.
Delivery
Claude Code as a senior engineer on every PR. Skills, hooks, and subagents tuned to your house style.
Review
Honest weekly readouts, regression-aware demos, decision logs your CTO can audit at 2am.
A day in the life
One PM, one agent, one loop.
Every hour of an SPM-shaped day exists to feed or close the loop. No standups, no status decks, no "syncs". Receipts at every gate.
- 09:00PAD lands
Daily context page hits the inbox: yesterday's prod telemetry, open PBIs, the one decision you're about to make. No standup.
- 09:30Triage queue
Yesterday's UAT findings + prod telemetry land in a single P0/P1/Ignore queue. You spend 15 minutes labelling, not gathering.
- 10:00Spec session
Open the next PBI. The agent has already drafted the failing test from acceptance criteria. You sharpen the spec; it sharpens the test.
- 11:00Build window — Vx
Agent owns the diff. You review the PR in chunks, not as a wall. Every chunk has its own failing→passing test as the receipt.
- 14:00QA + UAT
QA runs every commit. UAT is bounded: 5 users, 30 mins each, structured script. Output is a triage queue, not a vibes report.
- 16:00Security + prod
/security-review gates merge to main. Canary at 1%, ramp to 10%, full ramp. Telemetry flows back into tomorrow's PAD.
- 17:30Vx+1 brief
Agent writes the brief for tomorrow's session from triage + telemetry. You read it on the train. Tomorrow starts at 09:00.
What changes when you run the loop
The KPIs that move.
−40%
PBI cycle time
median across 12 engagements
< 4h
QA-fail → green
agent-owned diff loop
−60%
Prod incidents
vs pre-/security-review baseline
−70%
PM ceremony hours
no standups, no status decks
Numbers from internal engagements 2025-Q4 — 2026-Q1. Independent audits available on request. We don't ship metrics we can't show the workings for.
How it compares
SPM vs the alternatives.
Traditional PM, a basic AI tool, and the SPM approach — on the six dimensions that decide whether a product actually ships.
| Dimension | Traditional PM | Basic AI tool | SPM |
|---|---|---|---|
| Memory between sessions | Slack threads + your brain | None — chat resets | Persistent per-project skill + file memory |
| Test-first discipline | Depends on the engineer | Not applicable | Built into the loop — red gate before every build |
| Daily context | Standup + status deck | You write the prompt every time | PAD lands automatically — nothing to write |
| Security review | Once per sprint if lucky | Not in scope | /security-review on every PR, blocks merge |
| UAT triage | PM reads every note manually | Not applicable | P0/P1/Ignore queue generated automatically |
| Cost model | £70k–120k/yr PM salary | £20/mo + your time | From £200/block, agent runs between calls |
Common questions
FAQ.
Isn't this just AI agile?
Agile told you to value working software over comprehensive documentation. SPM tells you to value failing tests over both — and gives you the agent to write them.
What happens when the agent gets it wrong?
The failing test catches it. The rollback path is always live. The audit trail tells you exactly where the wrong turn was and why — every diff has a receipt.
Do I need to be technical?
You need to be opinionated about outcomes, allergic to ceremony, and willing to read a PR diff. That's it — the agent handles the rest of the engineering vocabulary.
How is this different from Cursor or Copilot?
Those are autocomplete in your editor. SPM is the whole loop: vision → spec → tests → build → QA → UAT → security → prod. The agent owns the loop, not the keystrokes.
The shape of the engagement
Not "AI tooling". A new operating cadence.
Most teams adopt AI by adding a chat tab. The SPM approach treats Claude as a coworker with memory, triggers, and accountability — with the PM as the operator. POMs are the unit of work. Skills are the unit of reuse. Agents are the unit of leverage.