Four-phase AI maturity model

A four-phase model for any team's AI-transformation journey, distilled from one reference deployment that has walked through all four phases in ~18 months. Each phase has its own success criteria, its own dominant failure mode, and a specific transition trigger to the next. Lets a founder honestly answer "what phase are we actually in?" and design the next move accordingly — rather than chasing flashy features from a phase three steps ahead.

The biggest reason transformations stall is mis-sequencing — trying Phase 3 (cross-team workflows replacing process) before Phase 1 (everyone using Cursor / Claude Code daily) is solid. The phases are not a marketing taxonomy; they're prerequisites. Each unlocks the next.

Phase 0 — Foundation ("Cursorization" of the company)
Three preconditions; without all three, nothing else compounds:

1. Every team lead personally uses Cursor / Claude Code daily for their own work. Founders who say "I'll just hire an AI lead" mostly fail. The founder must set up the tooling themselves and solve their own tasks first.
2. Single Shared Workspace (#162) holding all company knowledge. Department repos under Docs/. Personal files in Downloads/Desktop are killed.
3. Central navigation file (CLAUDE.md / AGENTS.md at root) — a distributor, not a rules dump. Two-level hierarchy: root → department.
   Success criterion: Every team lead can answer a cross-department question from their workspace inside 60 seconds.
   Failure mode: Founder delegates the setup. Don't.

Phase 1 — Workflows shared across the team

• Department heads adopt Cursor for personal work.
• They share "aha moments" with their team (~15–30 minutes in every leadership meeting; sometimes a whole meeting).
• Teams build small skills/workflows for themselves — research, prep, summaries.
• Force-function: deprecate the legacy wiki ("4 weeks to migrate; killing the knowledge base was the best admin tool I used"). Rotate one team lead who didn't adopt; replace with a senior who did. "~12 weeks of attention, 2×/week, until the point of no return."
  Success criteria: −50% Slack/Telegram messages, −15% internal meetings; meeting content shifts from "syncing" to "deciding". Sales-assistant headcount drops by 1; some analysts cut; one salesperson leaves and isn't replaced.
  Spend signal: AI subscriptions ≈ ~4% of payroll, growing.
  Failure mode: Stalling at "some power users." Without the deprecation event you stay at 40% adoption forever.

Phase 2 — Cross-team workflows replace the process

• An owner (head of analytics / head of CS) writes one big "workflow skill" — a ~200-line markdown file describing the entire end-to-end process (e.g., client onboarding) with links to sub-instructions for each step.
• Operator says to agent: "start it" (one word, just go). Agent walks the 200-line plan. When it stalls, the human fixes the workflow file, not their head.
• All learnings flow back into workflow files, not into people's heads or chat history.
• Org effect: Integration team and Product team start working on the same artifacts (the agents). Time to integrate a client: 6 weeks → 2 weeks → 1–2 days.
  Success criterion: ≥1 end-to-end process has been replaced by a workflow skill running unattended for ≥2 weeks.
  Transition trigger to Phase 3: Founder realizes ≥2 heads of function are writing prompts that direct their team's daily priorities.

Phase 3 — AI Overlord (proactive workflow agent directs the team)

• The head of a function writes a long prompt that directs the team.
• A per-project agent runs nightly/regularly, reads everything (correspondence, tasks, transcripts, comments, data), and produces a per-project priority list / dashboard / instructions for human operators.
• People work the prompts instead of dashboards. People are happier (clear priority list, less ambiguity). The head needs less daily discipline to spot the right priorities.
  Success criterion: A team member's morning ritual is "read what the agent decided is most important today" rather than "open Linear and Slack and try to triage."
  Failure mode: Framing it as "AI manages humans." Frame it as "AI prioritizes, human decides" — people feel relief, not surveillance.

Phase 4 — Selling agents

• "Software used to be a cost-saver. AI software directly makes/saves money — so you can charge as a service provider with success-based pricing." (See #196.)
• Two product flavors:
• Copilot for the client's team ($5–15K/mo at reference deployment; lots of small clients): agent connects to client's systems, runs ad-hoc analyses, suggests rules. Human at client still owns.
• Outsourced agent-driven service (high prices, success-based): the company takes responsibility for the outcome; 90% of work done by agents, ~1 customer-success person per client.
• Inside structure: layers = data → product → agents → tools → skills → automations.
  Success criterion: Real paying pilots in either tier.

Self-diagnosis card — fill in honestly, then place yourself:

Foundation:
[ ] Every team lead opens Cursor/Claude Code ≥3×/day for their own work
[ ] Single Shared Workspace exists, every employee is in it
[ ] Root CLAUDE.md/AGENTS.md works as a navigation distributor

Phase 1 signals:
[ ] AI subscriptions ≈ 3–5% of payroll, centralized billing
[ ] Slack/Telegram volume down ≥30% YoY
[ ] At least one role's headcount didn't grow with revenue
[ ] Legacy wiki has been formally killed (not just discouraged)

Phase 2 signals:
[ ] ≥1 end-to-end process has a 100–300-line workflow skill that ships work unattended
[ ] Time-to-integrate (or analog metric) has dropped ≥3×

Phase 3 signals:
[ ] ≥1 head of function has a "directing prompt" that team members work from daily
[ ] People say their mornings are less chaotic, not more surveilled

Phase 4 signals:
[ ] You've packaged an agent service externally (copilot or outsourced)
[ ] At least one paying pilot exists

You are in the lowest-numbered phase where any signal is unchecked.

Phase-by-phase rate-limiting question to ask weekly:

Phase 0: "Have I personally solved a real work task with Cursor this week?"
Phase 1: "Did I move one important file out of Notion/Slack/Drive into the workspace?"
Phase 2: "Did I extend the workflow file when the agent stalled, or did I do the work manually?"
Phase 3: "Did the team work from the directing prompt today, or from Linear?"
Phase 4: "Do I have a customer paying for an outcome, not a seat?"

• Don't claim Phase 3 with Phase 1 infrastructure. Most companies that say "we have AI agents directing the team" actually have one prompt that runs nightly and is mostly ignored. That's Phase 2 attempted poorly. Fix the workflow files first.
• The 4% rule. AI subscription spend ≈ 4% of payroll at maturity; budget 5–10% for next year. Below 1% you're under-investing; above 8% without phase-aligned ROI you're overspending without process gains.
• 80–90% of new-hire success has nothing to do with AI skills. Speed, judgment, domain fit, team match. AI skills are picked up in ~8 weeks if the person isn't resistant.