Eight chapters on shipping AI in a real operating business.

The series

Read in order, or skip to the one you need.

Chapter 1 is the cornerstone — the one to read if you're starting from scratch. The rest are being written and will land over the coming weeks.

What each chapter covers

The long view.

A three-paragraph outline of each chapter so you can decide which one is worth your 10 minutes on a Tuesday.

Chapter 01

Where to actually start with AI in an operating business (and what to ignore)

Most businesses that tried AI in 2024 tried the wrong thing first. They wrote a strategy deck, ran a Copilot pilot, or hired a junior to 'look into it'. None of those approaches ship anything. Twelve months later, the business has spent real money and changed nothing about how it runs on a Tuesday morning.

This chapter is the one you want if you run an operating business — pubs, a retail chain, a services firm, a coffee brand — and you've been circling AI for six months without a single workflow deployed. It's written for the MD or founder who needs to pick the first battle well, because the first one sets the pattern.

The framing is simple: AI is already useful at SME scale if you pick one reversible, measurable, admin-shaped workflow and actually deploy it. It is not useful if you treat it as a strategic bet, an all-hands announcement, or a six-month transformation programme. By the end of this chapter you'll know where to start, what to ignore, and whether your business is actually ready.

Chapter 02

Picking the first workflow to automate — a decision framework

The wrong first workflow wastes three months and kills internal appetite for AI. The right one pays back inside a quarter and funds the next two. The difference isn't the tool — it's whether the workflow was chosen by a framework or by whoever shouted loudest in a meeting.

This chapter walks through the exact 30-minute exercise we run at the start of every audit. You list the admin tasks in your business, tag each one with its weekly cost in hours, rank them by complexity, and apply a single decision rule. The rule is counter-intuitive — it's not 'automate what costs the most', it's 'automate what pays back fastest with the lowest emotional load', because emotional load is what kills rollout.

You'll get four worked examples — one hospitality group, one coffee brand, one professional services firm, one SaaS — showing what the exercise actually looks like when an operator does it on their own business. The output in each case is a shortlist of three candidates, one recommended first pick, and a reason to start Monday instead of writing another spec.

Chapter 03

Building vs buying AI tools — when each one wins

Every AI tool sales call pushes you to buy. Every developer on LinkedIn pushes you to build. Neither is right universally. The answer depends on three things: how core the workflow is to your P&L, how often the vendor's incentives stay aligned with yours, and how much unique data or process lives inside the task.

This chapter is the decision tree we use when a client asks 'should we pay HubSpot for this or build it?' We walk through three real calls from the last 12 months — one where we bought, one where we built in n8n, one where we wrote 80 lines of TypeScript — and we show the full maths, including the hidden costs most people miss (integration, change management, vendor lock-in, and the compounding cost of mediocre tooling).

The rule of thumb at the end is simple enough to remember the next time a vendor pitches you: buy the commodity, build the edge, and never pay for a feature that will be commoditised inside six months. You'll leave this chapter able to say 'no' to 80% of sales calls with a reason, not a feeling.

Chapter 04

The hidden cost of doing nothing

'We'll look at AI next year' is the most expensive sentence in the UK SME economy right now. It sounds cautious. It feels responsible. It's actually a large, compounding bet — one most operators have never modelled, because the loss shows up in places accountants don't flag.

This chapter does the maths. We run the spreadsheet on three business archetypes — a four-pub hospitality group, a £6M services firm, a £12M DTC brand — and show what 'wait and see' actually costs them across four vectors: margin compression from competitors automating faster, talent drift as the strongest operators leave for AI-native peers, customer-expectation gap as buyers get used to instant responses elsewhere, and founder time lost to admin that should have been automated 18 months ago.

By the end you'll have a one-page model you can run on your own business. The numbers are not scary if you act. They are genuinely scary if you don't. The goal of the chapter isn't to panic you — it's to give you the ammunition you need to move a board, a co-founder, or a risk-averse FD from 'not yet' to 'let's start with one workflow'.

Chapter 05

How to roll AI out without breaking your team's trust

AI projects don't fail on tech. They fail on people. Specifically: on the gap between what leadership announces in the all-hands and what actually happens at the desk on Monday morning. Close that gap and adoption follows. Leave it open and you'll have a beautifully-configured tool that nobody uses.

This chapter covers the three hardest pieces: the comms sequence leadership needs to run before anyone touches a tool (what to say, what never to say, who speaks when), the internal ownership model that works in a 20–200-person business (hint: it's not a new hire, it's a 20% carve-out from someone who already has credibility), and the 4-hour training template that turns a sceptic into a user.

We also cover the two people every rollout needs to handle well: the team member who's worried AI is coming for their job (it isn't, but the fear is real and ignoring it corrodes trust) and the enthusiast who thinks AI is a magic wand (the bigger risk, because they'll ship low-quality work with confidence). The goal is a team that trusts the process enough to tell you when a workflow isn't working — because that honesty is what lets you fix it fast.

Chapter 06

Measuring ROI on AI integration — what to actually track

Most AI 'success stories' published in 2024 and 2025 measured the wrong thing. 'Generated 10,000 drafts' is a vanity number. So is 'saved 200 hours a week' when nobody can name which 200 hours or what actually changed for the customer. Unmeasured AI is just theatre, and theatre gets cut at the first budget review.

This chapter is the honest guide to what to track. Three numbers that matter — hours reclaimed (counted, not estimated), error rate (the real one, measured post-deployment by someone who isn't the owner), and cycle time (end-to-end, including the human review step). Four vanity metrics to ignore — generations, tokens, engagement rate on AI-assisted comms, and 'team adoption' measured by logins.

We finish with a one-page dashboard you can build in 20 minutes in a spreadsheet, a weekly review cadence that takes 15 minutes, and the one question you should ask every quarter that most operators never ask: 'If we turned this off tomorrow, who would complain, and how loudly?' If the answer is 'nobody', you've built a solution in search of a problem. Kill it, and use the budget for something that actually runs.

Chapter 07

Hiring for AI-native operations (or upskilling what you have)

The 'Head of AI' job ad is everywhere on LinkedIn and almost none of it is right for an SME. A hire at that level costs £120k+, takes six months to settle, and usually arrives with a playbook built for enterprise — not a business where the MD still signs the invoices. For most 20–200-person businesses the right answer is not a hire at all.

This chapter covers the three realistic paths. The 20% internal owner: an existing high-performer (usually an operations or product manager) who gets a formal carve-out of their week to own AI tooling, with budget, mandate, and a clear target. The AI-fluent mid-career hire: what to screen for in a CV, what to ask in an interview, and the one practical task to set in stage two that reveals whether they've actually shipped anything. The fractional option: when an external operator on retainer beats a full-time hire, and the contract shape that keeps it honest.

We also cover upskilling the team you already have. The two-hour weekly practice session that builds real fluency. The internal 'show and tell' cadence that compounds shared knowledge. And the uncomfortable question about who on your team will not make the transition — and what to do about that, humanely and early, rather than pretending it won't come up.

Chapter 08

The next 18 months — how to stay ahead without chasing every release

Every model release gets a Twitter frenzy, a round of 'everything changes now' posts, and a week of board members forwarding articles to founders. Almost none of it matters at SME scale inside 30 days. Some of it matters in 12 months. Your job as an operator is to separate the two without spending half your week reading.

This chapter is a thesis, written to be useful to a £5M–£50M business making a budget this year. We cover the three shifts that will materially change operator economics in the next 18 months — pricing collapse on foundation models, the real (not hyped) state of agentic workflows, and the regulatory tightening coming from the UK's evolving position on AI. We also cover the four things you can safely ignore for another year without falling behind.

The posture we recommend at the end is the one we use ourselves: run one new workflow a quarter, kill anything that isn't paying back inside 90 days, and treat the hype cycle as noise unless a tool passes a specific test we describe in the chapter. This is how you stay current without burning your week on releases, and how you keep compounding small wins instead of chasing the next headline.