The book

The Compounding Company

Building an AI-native business that gets better every time it works.

Most AI programmes cannot prove they changed anything — not because the technology failed, but because nobody wrote down what they were changing before they started. This is a measurement discipline for people who would rather know.

What it argues

The counterfactual disappears as it passes.

Every quarter you run a programme without a baseline is a quarter that can never be measured retroactively. That is the whole problem, and no amount of additional technology solves it.

The book is organised around a diagnostic. Chapter 3 hands you the whole instrument and asks you to stop reading and score your own organisation. Every part after that takes one dimension of that score and covers what moving it actually requires — what to build, what it costs, what breaks, and what to measure.

It weights technology capability at eighteen points out of a hundred and defends the choice. And it publishes its own weights as a hypothesis with a stated test, on the grounds that a book demanding evidence should be willing to supply some.

The six movements

Fifty chapters, six questions.

Prove itDid AI actually change anything?
Remember itCan the company recover its facts, signals, decisions and precedent?
Automate it safelyHow much autonomy has the system earned?
Measure the economicsDid human effort and cost per outcome actually fall?
Change what you sellDoes greater efficiency destroy revenue under your existing model?
CompoundDoes each deployment make the next one cheaper, faster and better?

If you read nothing else, read the first and the fifth. They are the two most organisations get wrong, and the fifth is the one almost nobody is asking.

Contents

Nine parts, fifty chapters, seven appendices.

Part I — The Argument
  1. 1Nobody Can Prove It Worked
  2. 2Procurement Is Not Transformation
  3. 3Score Yourself Before You Read Any Further
  4. 4Why Technology Is Eighteen Points
  5. 5If You Are Under Two Hundred People
Part II — What Your Company Remembers
  1. 6What a Company Actually Knows
  2. 7Facts Are Not Signals
  3. 8Retrieval Is Not Search
  4. 9The Decision Is the Artefact
  5. 10Permissions Are Where Enterprise Rollouts Die
  6. 11Freshness, Provenance and Trust
Part III — The Execution Layer
  1. 12The Autonomy Ladder
  2. 13Your First Agent Should Be Boring
  3. 14Designing the Human in the Loop
  4. 15When Agents Fail
  5. 16Orchestration, Handoffs and the Agent Registry
Part IV — The Machine Room
  1. 17Choosing Frontier Models
  2. 18Routing, Fallback and Never Being Locked In
  3. 19Running Your Own Models
  4. 20Build, Buy, or Partner
  5. 21MCP and the End of Bespoke Integration
  6. 22Evals — The Discipline Nobody Wants
  7. 23Your MCP Surface Is a Liability
  8. 24Containment, Not Detection
Part V — The Department Playbooks
  1. 25The Transferable Pattern
  2. 26Engineering
  3. 27Support
  4. 28Sales and Marketing
  5. 29Finance and Legal
  6. 30Operations and Delivery
Part VI — The Company That Runs It
  1. 31Coherence, Not Average
  2. 32Who Owns This
  3. 33Governance That Doesn't Stop Everything
  4. 34Measuring the Cost You Impose
  5. 35The Agent Supervisor
  6. 36Budgeting AI as Infrastructure
Part VII — The Economics
  1. 37Revenue Per Employee Is the Score Outside People Use
  2. 38Human Minutes Per Outcome
  3. 39No-Touch Rate
  4. 40What This Actually Costs
  5. 41Reporting This Upward
Part VIII — What You Sell
  1. 42Everything Above This Was About How You Work
  2. 43The Boundary Is Moving
  3. 44The Billable Hour Paradox
  4. 45Quantifying Your Own Deflation
  5. 46Repricing Before You Are Forced To
  6. 47Selling Something That Could Not Have Existed
Part IX — The Compounding Company
  1. 48The Self-Learning Loop
  2. 49Time to Launch
  3. 50What This Looks Like in Five Years

Appendices: the ANS-100 instrument in full · the interview protocol · the evidence checklist · a worked assessment · the calibration programme · where this book is contested · the three canvases, blank and completed.

Honest about the edition

This is a first edition, published while the research is still running

The diagnostic assigns weights to seven dimensions. Those weights are a reasoned hypothesis derived from practice and checked against published research. They have not yet been validated against outcome data.

When the data arrives, some of them will change. The book commits in print to publishing the corrections, including where they contradict the chapter that defends the current weighting. Revisions are maintained here.

The extended anecdotes in the book are composites, and it says so. One chapter — on people and recruiting — is deliberately absent pending employment-law review, and it says that too.

Buy

Paperback and Kindle.

Bulk orders and workshop editions: get in touch through the assessment page.