Skip to content

Phase 0: Readiness

Duration: 2–6 weeks, depending on company size. Goal: decide whether, when, and under what conditions to adopt AI-assisted development, and make sure the preconditions are in place before the pilot starts.

Why Phase 0 exists

Most companies skip this phase. They jump to Phase 1 because "the developers are already using Copilot anyway" or because a board has asked "what's our AI story?" Skipping Phase 0 is why most AI-assisted-development programs later have to unwind incidents publicly: Samsung banning ChatGPT after engineers pasted source code into the public model1, the Replit agent wiping a production database during a code freeze2, the Chevrolet dealer chatbot agreeing to sell a truck for $1 after a prompt injection3.

Phase 0 buys you the right to go fast in Phase 1 by getting the cheap decisions out of the way early.

Phase 0 outputs (what "done" looks like)

Output Who owns it Worksheet
Leadership charter: a one-page document saying whether the company will adopt, under what guardrails, and who sponsors it. Executive sponsor (CTO / VP Eng / equivalent) Leadership charter
Interim Acceptable Use Policy: a one-page document covering what data must never be sent to AI tools, which tools are approved, and how to report an incident. CISO or Head of Security (delegated to AppSec if needed) AUP drafting
Workload risk classification: a short list of the systems the company runs, each tagged Critical / High / Medium / Low per the risk matrix. Architecture / risk function Risk classification workshop
Incident response addendum: existing IR plan updated with AI-specific scenarios (prompt injection, hallucinated dependency, agent-gone-wrong, model-output exfiltration). IR lead / CISO Incident tabletop
Metrics baseline: current DORA four-keys and, where possible, SPACE / DevEx baselines. You cannot measure improvement without this. DevEx or platform team Metrics baselining

Exit criteria

All five must be true:

  • Charter signed by a named executive sponsor with authority to stop the program.
  • Interim AUP published and acknowledged by anyone who uses or will use an AI coding tool.
  • At least the top-tier workloads (authentication, payments, PII systems, regulated systems) classified. Lower tiers can be filled in during Phase 1.
  • IR plan updated; tabletop run once; owners of each response step identified.
  • Baseline metrics recorded with a date. Missing data is itself a recorded finding.

Sizing guide

  • Small (< 50): Charter, AUP, risk list, and baselines fit on four pages total. One person usually authors all of them; a peer reviews. Skip the formal tabletop; just walk through the IR plan verbally.
  • Mid (50–500): Distinct owners per output. Charter and AUP go to a single review meeting, not a committee.
  • Enterprise (500+): Each output goes through its normal approval chain (Legal, Privacy, Security, Procurement). Run an exec-level tabletop with IR, Legal, and Comms in the room.

The moves, in order

1. Name an executive sponsor (day 1)

One named person who owns the program, has authority to stop it, and reports on it to the board or equivalent. Not a committee. Committees do not stop programs.

2. Hold the leadership charter meeting (week 1)

Use the Leadership charter worksheet. The meeting produces a single-page answer to:

  • Why are we adopting AI-assisted development? (What does success look like?)
  • What are we explicitly not doing? (What's out of scope?)
  • What would cause us to pause or roll back?
  • Who owns go/no-go decisions?
  • What review cadence will the program follow?

Output: signed charter + Decision Record. If the answer is "don't adopt," you stop here; the charter documents why.

3. Publish an interim AUP (week 1–2)

Use the AUP drafting worksheet. The interim AUP is deliberately small, ideally one page, because it exists to be honoured, not admired. It covers:

  • What data must never be sent to any AI tool (e.g., customer PII, regulated data categories, source code for certain components).
  • Approved tools. An initial list of 1–3 tools you're explicitly allowing, with the rest implicitly disallowed.
  • Reporting incidents. One email address or Slack channel for incidents.
  • How the policy will be updated. Review cadence and owner.

Output: AUP v0.1, published to a durable location (wiki, intranet page), acknowledged by engineering staff.

Don't let perfect block this

Companies spend six months drafting the "real" AUP while engineers are pasting source code into public models on day 1. A 70%-correct policy published in week 2 beats a 95%-correct policy published in month 6.

4. Classify your workloads (week 2–3)

Use the Risk classification workshop. You don't need to classify everything. Start with systems that touch one of: regulated data, external customers, money, or production identity.

The five factors from the Risk Classification Matrix:

Factor 1 2 3 4
Data sensitivity Public Internal Confidential Restricted / regulated
System impact Non-prod Staging Prod, non-critical Prod, critical
User exposure Internal only Small internal Broad internal External / customers
Regulatory requirement None Internal standard Contract / SLA Legal / regulatory
Security requirement Standard Hardened Hardened + audit Highest + attestation

Sum the factors. Bands: Critical (16–20), High (11–15), Medium (6–10), Low (5). Critical and High workloads will require enhanced verification in Phase 1 (V.E.R.I.F.Y. Level ⅔).

Output: classification table, stored somewhere durable. Revisit every 6 months.

5. Update the IR plan and run one tabletop (week 3–4)

Use the Incident tabletop worksheet. At minimum, walk through these scenarios:

Scenario Why it matters
A developer pastes a secret into a public model. Near-universal "first incident."
A model suggests a package name that doesn't exist; an attacker has registered it ("slopsquatting"). Documented attack vector4.
An AI agent with write access makes a destructive change in production. Replit incident precedent.
A customer-facing AI makes a commitment the company can't honour. Air Canada chatbot ruling5.
The AI vendor has an outage or data exposure. DeepSeek ClickHouse exposure6.

Output: updated IR runbook, assigned owners per step, dated.

6. Baseline your metrics (week 4–6)

Use the Metrics baselining worksheet. The minimum you need before Phase 1:

  • DORA four-keys (deployment frequency, lead time for changes, change-failure rate, MTTR). See the 2024 DORA report for definitions.
  • SPACE five dimensions if your survey infrastructure permits; if not, at least a developer-satisfaction single-question quarterly survey7.
  • Security signals. Open CVEs on the target system, rate of critical findings per month, time-to-fix on prior quarter.
  • Code-quality signals. Churn rate (pieces of code reverted within 2 weeks), PR review time, merge/rebase conflicts.

If you don't have these today, that's a finding, not a blocker. Document what you can measure, what you cannot, and a 90-day plan to close the gap.

Common objections

"We don't have time for Phase 0; the CEO wants a demo next Friday." → A one-hour leadership charter meeting + a one-page AUP is achievable inside a single day with the right people. Friday's demo is the pilot's goal, not its start.

"We'll write the policy based on what we learn in the pilot." → An interim policy bounds the learning. Without it, the pilot will expose you to risks (data leakage, IP contamination) that the pilot itself is supposed to study.

"Our devs won't follow a policy they weren't involved in drafting." → Involve them. The AUP worksheet has a "working group" role; staff it with developers. An AUP written by a CISO alone is the one that gets ignored.

"This framework is for enterprises; we're 20 people." → For a 20-person company, Phase 0 is four meetings over a week, producing four one-page artifacts. Not adopting the framework doesn't remove the risks; it just removes the preparation.

Anti-patterns

Don't

  • Don't start procurement in Phase 0. Tool selection belongs in Phase 1 or early Phase 2. Buying a vendor before you know your risk profile inverts the decision.
  • Don't over-specify. A ten-page AUP will be ignored. A one-page AUP that says "no customer PII in prompts, approved list is X, Y, Z, incidents to email@company" will be followed.
  • Don't let Phase 0 become a forever-running steering committee. Three weeks is normal. Three months means the committee has become the deliverable. Set an end date.

Next

  1. Bloomberg, "Samsung Bans Staff's AI Use After Spotting ChatGPT Data Leak", May 2023. ↩

  2. Replit AI agent deletion incident, Jul 2025; see Evidence: security incidents. ↩

  3. Business Insider, "A Chevy dealership's AI chatbot agreed to sell a Tahoe for $1", Dec 2023. ↩

  4. Lasso Security, "AI Package Hallucinations"; Vulcan Cyber, "Can you trust ChatGPT's package recommendations?". ↩

  5. Moffatt v. Air Canada, 2024 BCCRT 149, decision online. ↩

  6. Wiz Research, "Wiz Research Uncovers Exposed DeepSeek Database", Jan 2025. ↩

  7. Forsgren, Storey, Maddila, Zimmermann, Houck, Butler, "The SPACE of Developer Productivity", ACM Queue, Mar 2021. ↩