The AI-Era Operating Playbook: From First AI Pilot to Production in 90 Days

The Key Points of Every Chapter

The AI-Era Operating Playbook is a practical, vendor-neutral guide for founders, COOs and CTOs at growing and mid-size companies. Below are the key points of every chapter, free to read. The full PDF adds the worked examples, checklists and fill-in templates.

1. Where AI Pays Back First

• Write every idea as one sentence: "[Role] spends time on [task] for [volume], and a good result is [outcome]." If you can't fill it in, the idea isn't ready to score.
• Score each idea 1 to 5 on impact, feasibility and data readiness, then multiply. Multiplying stops one weak factor from hiding behind two strong ones.
• Before you commit, apply four filters: a named owner, a measurable metric, reversible actions, and parts the next use case can reuse.
• The first pilot is rarely the moonshot. Big ideas with weak data go on the roadmap, behind a data project.

2. AI Readiness Checklist

• Readiness belongs to a use case, not to the company. Score it for one specific task.
• 20 questions in four areas: data, systems and process, people, and security and governance. Each scores 0, 1 or 2, so the maximum is 40.
• Four bands tell you the next step: ready to pilot, ready with gaps, foundations first, or not yet.
• Some gaps block the pilot whatever the total: no data access, no baseline, no business owner, or no control over sensitive data. Messy but accessible data can usually be fixed during the pilot.

For a quick self-check you can do today, see our AI readiness checklist.

3. Build vs Buy vs Partner

• A seven-step decision tree, run for one use case: is the task generic, does a product handle your real data, does it meet your data rules, does it cross your systems, is it a source of advantage, and do you have the skills to build and run it?
• Hybrid is the most common answer: a bought model and components, plus custom integrations and business rules.
• Integration (especially write access) and data ownership settle most build-vs-buy debates.
• Compare total cost of ownership over three years across nine cost drivers, including tokens, evals and switching cost.

4. The 90-Day Pilot Plan

• Five phases: baseline and discovery (weeks 1-2), design and evals (3-4), build (5-8), shadow mode and limited live (9-10), then measure and decide (11-13).
• A week-by-week table sets the outputs and owner for every week, and a small team of four to six covers six roles.
• Agree three metrics before building: a value metric, a quality metric and an adoption metric, each with a baseline and a target.
• Kill criteria go into the charter on day zero. They cover quality, safety, adoption, economics and access, each with a date to check it.

5. Agents, RAG, MCP and Voice AI: When to Use Which

• Use the simplest pattern that works: a single model call, then a workflow with AI steps, then RAG for your own knowledge, then an agent for multi-step judgment.
• MCP (Model Context Protocol) is an open standard that connects AI to tools and data. It gives several AI apps the same permissioned access to your systems.
• Voice AI goes on top of a design that already works in text.
• A comparison table sets out what each pattern is best for, how hard it is to test, what drives its cost and the first guardrail it needs. A second table compares RAG, fine-tuning and long context.

6. Evals, Guardrails, Security and Governance

• Build an eval set of 100-300 real cases with known good outcomes in week 2, and run it on every change.
• Anything that must never happen is blocked in code, not requested in the prompt. Guardrails include least-privilege tools, per-user permissions, approvals, output validation and hard limits.
• Autonomy is a dial: suggest, draft, act with approval, act. Raise it one step at a time.
• The key security risks are prompt injection, data leakage, excessive agency, unsafe output handling, the supply chain and runaway cost. Red-team before launch.
• Governance starts with a use policy, a named owner, a risk rating, an audit trail, a change process and an incident process.

7. Measuring ROI

• ROI = (total benefit − total cost) ÷ total cost. Payback = one-time cost ÷ monthly net saving.
• Everything rests on cost per task before and after. The after figure includes human review, model usage, rework and fixed running costs.
• A line-by-line worksheet (with an example clearly labelled as illustrative) shows the arithmetic from tasks per month to 12-month ROI.
• Model conservative, expected and optimistic cases. Saved hours only count if the time is put to other use.

8. Scaling From Pilot to Production

• Production readiness covers quality, ownership, a runbook, monitoring, security, a cost budget, support and a change process.
• There are three operating models: embedded in the business team, a central platform team, or partner-run.
• Token cost ≈ tasks × calls per task × (input tokens × input price + output tokens × output price). Seven levers reduce it, including model routing, caching, tighter retrieval and step caps.
• Monitor quality, cost, performance and safety weekly. Keep the system model-agnostic so a model change needs only an eval run, not a rebuild.

9. How to Choose an AI Partner: 20 Questions

• 20 questions in five groups: experience and team, approach and quality, security and data, cost and commercials, ownership and exit.
• Ask for specifics and documents, not slides. That includes a production system like yours and the named people who will build it.
• Red flags include accuracy guaranteed before the partner has seen your data, demo-only confidence, no eval process, and code or prompts you won't own.
• Score the answers with a weighted scorecard. Ask the same questions of every option, including us.

10. Templates

• Use-case canvas: one page per idea, ending in its priority score.
• Pilot charter: scope, architecture, autonomy levels, metrics, kill criteria, guardrails, budget and decision date, signed by the sponsor.
• ROI sheet: the Chapter 7 worksheet with conservative, expected and optimistic columns.
• Vendor scorecard: eight weighted criteria mapped to the 20 questions.
• Guardrail list: one row per tool, data source or channel. It records permission scope, approval, hard limits, logging and owner.
• Each template maps to a chapter and can be filled in during one working session with the business owner.

Get the Templates and Worked Examples

The full playbook is free. Fill in the short form to download the PDF with every worksheet, decision tree and template. If you would rather talk it through, book a free consultation and bring your best use case.