Playbook

Projects that should stop rarely do, because nobody wants to declare failure. This playbook covers recognising the signals early, stopping cleanly, and capturing the learning so the next one benefits.

Most failing AI projects fail for the same few reasons, and most are recoverable if diagnosed honestly. This playbook covers finding the real cause, narrowing scope, and deciding whether to continue.

Regulatory inquiries about AI systems are won or lost on evidence assembled beforehand. This playbook covers the first hours, assembling what is asked for, coordinating the response, and what to avoid.

Discovery in regulated environments has to establish constraints before opportunities. This playbook covers surfacing obligations early, testing data access honestly, and scoping something that survives security review.

Maturity assessments produce scores nobody acts on. This playbook covers assessing against evidence rather than self-reported opinion, and finishing with a sequenced plan rather than a radar chart.

Bringing an AI system in-house fails when the transfer covers code and not knowledge. This playbook covers what to acquire, what will need rebuilding, and how to transfer the understanding that matters.

Vendor audits ask capability questions and miss the ones that matter. This playbook covers what to establish before signing, what evidence to require, and what to review while the relationship runs.

Training data decisions are close to irreversible, which makes provenance the control that matters. This playbook covers establishing rights, recording where data came from, and handling personal data carefully.

AI documentation fails by describing what was built rather than what someone operating it needs. This playbook covers the sections that matter, what to keep current, and how to stop it going stale.

AI data incidents spread through paths conventional response plans miss: prompt logs, vector indexes, and provider retention. This playbook covers containment, scope, notification, and prevention.

AI contracts carry clauses that conventional software agreements do not. This playbook covers data rights, model change, output ownership, liability for wrong results, and the exit terms worth fighting for.

Change control for AI has to cover things conventional processes miss while staying fast enough to use. This playbook covers classifying changes, gating by risk, and producing evidence automatically.

AI systems change through prompts, models, and retrieval corpora, and most teams version only the code. This playbook covers versioning all three so behaviour changes are traceable and reversible.

Escalation paths fail when they route to someone who cannot help. This playbook covers choosing triggers by consequence, giving reviewers the context and authority to resolve, and measuring whether escalation works.

Feedback buttons collect noise nobody acts on. This playbook covers capturing feedback that is specific enough to act on, routing it to where it changes something, and converting it into evaluation cases.

AI coding agents raise output volume and move the constraint to review. This playbook covers sanctioned tooling, repository preparation, review capacity, and measuring whether delivery actually improved.

AI user research has traps conventional research does not: novelty effects, over-trust, and demos that hide failure. This playbook covers studying the real task and testing with realistic errors.

AI systems need support that differs from conventional on-call, because quality failures are not outages. This playbook covers what the rotation covers, who can serve, and how to keep it sustainable.

Acceptance criteria written for deterministic software do not work for AI. This playbook covers writing criteria that can actually be tested, including thresholds, refusal behaviour, and operational requirements.

AI cost forecasts are usually wrong because they extrapolate token prices rather than modelling task cost and volume. This playbook covers building a forecast that survives contact with reality.

AI budget requests fail on framing more than on merit. This playbook covers connecting to existing priorities, staging the ask, pre-empting the standard objections, and preserving credibility for the next request.

Model provider negotiations focus on price when the terms matter more. This playbook covers knowing your usage precisely, what is genuinely negotiable, and the non-price terms worth spending leverage on.

AI capacity is constrained by provider quotas more than by your infrastructure. This playbook covers modelling demand honestly, understanding the real limits, and deciding in advance what degrades under load.

Board presentations on AI fail when they explain the technology. This playbook covers framing the decision, presenting exposure honestly, showing what is already happening, and what to commit to reporting.