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 asked a junior to "look into it" for a quarter. Twelve months later the business has spent real money and changed nothing about how it runs on a Tuesday morning.

That pattern is not an AI problem. It's a starting-move problem. The first decision in an AI rollout — what you pick to deploy first — determines whether the next twelve months compound or stall. Pick well and you'll have three workflows running inside a year, a team that trusts the process, and a budget that keeps getting approved. Pick badly and you'll have a Slack channel full of people sharing screenshots of ChatGPT for the rest of eternity.

This chapter is the one you want if you're an MD or founder running 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. No strategy deck required. Just a cleaner answer to the question: what do I actually start with, and what do I leave alone for another quarter?

The shape of the workflow you want first

The first AI workflow in your business should be four things: reversible, measurable, admin-shaped, and owned by one person. Miss any of those four and the project slips.

Reversiblemeans you can turn it off on a Friday and the business keeps running on Monday. That rules out anything customer-facing, anything in the finance sign-off chain, anything touching pricing decisions. You want a workflow where the downside of it breaking is "we do it manually for a week", not "we've embarrassed ourselves with a guest" or "we've mis-invoiced a supplier".

Measurablemeans you can put a number on it before and after. Hours per week, error rate, cycle time, cost per ticket. If the success criterion is a feeling ("it felt faster"), you will not get budget for the second workflow.

Admin-shaped means the workflow is mostly rearranging text or data — summarising, classifying, drafting, extracting. Not making judgments. Not replacing a person. The jobs AI is genuinely ready for at SME scale in 2026 are the ones where the person was already reading something, writing a version of it, and sending it on.

Owned by one personmeans a named human who loses something if it doesn't ship. Not a committee. Not "the ops team". One person with a calendar entry for 20% of their week and explicit permission to spend up to £500 on tooling without asking. If you can't name that person inside five minutes, you aren't ready to start — but you are ready to assign them, which is chapter five.

Three workflows that reliably pay back inside 90 days

Across the businesses we run ourselves — four pubs, a coffee brand, a chefs' marketplace, a hospitality software product — three workflow shapes have paid back inside a quarter every single time.

The first is supplier invoice extraction and reconciliation. Pipe incoming PDFs and email attachments into Claude via a simple script or an n8n workflow, pull out the line items, match them to the purchase order, flag discrepancies for human review. At Coal & Co this cut the finance team's admin load by six hours a week and caught about £800 a month of supplier pricing drift that nobody had the time to notice before. The tooling: Claude for extraction, n8n to orchestrate, Xero as the system of record. Monthly cost, all-in, under £80.

The second is first-draft response generation for repeat inbound enquiries. Every SME has an inbox full of questions the business has answered a thousand times — booking policies, delivery timelines, product availability. A Claude prompt plus your actual past replies turns a cold inbox into a queue of 80%-done drafts the human just edits and sends. Note the word drafts. You are not replacing the human. You are deleting the blank page. At Need, our coffee brand, this turned a two-hour-a-day B2B inbox task into a 25-minute one, with no customer-facing visibility of the change.

The third is meeting-to-action conversion. Record the internal meeting, pipe the transcript into Claude with a prompt you've tuned to your house style, get back a decisions-and-owners list in three minutes. Post it into Notion or your wiki automatically. The tooling here can be as simple as Otter or Granola for capture and a single Claude call for the summary, or as neat as a Cursor-built script that writes directly to your project tracker. We run this on every weekly leadership meeting across the portfolio. The cost is rounding-error. The benefit is that nothing gets lost, and the new hire in month two can read back through every decision the team has made for a year.

Three workflows. Each one reversible. Each one measurable. Each one admin-shaped. Each one owned by one person. None of them involves a Head of AI, a strategy deck, or a custom model. Total monthly tooling spend across all three, at the businesses we actually run: under £200.

What to ignore (for at least another quarter)

The same honesty that picks the right first workflow has to pick the wrong ones out. Here's what we watch SMEs waste money on every month, and what to do instead.

Bespoke chatbots on your public website.The tooling is cheap; the content moderation risk isn't. The return on making a customer talk to a bot before they talk to you is almost always negative at SME scale, because your volume doesn't justify it and your brand does the work of a bot better. Skip it. Come back to this one in 2027.

"AI strategy" engagements from large consultancies.If someone is charging you £40k to tell you where AI fits in your business and they have never run a P&L the size of yours, they cannot give you a useful answer. They will give you a deck. The deck will sit on a shared drive. You know this because you've already had one of these done, or a peer has, and nothing shipped out of it.

Autonomous agents for customer-facing work.The technology is real and improving. The reliability at SME scale, today, is not where it needs to be for anything that touches a paying customer. Agents are a 2027 bet for most operating businesses. Keep half an eye on the space, but don't spend money on it this quarter.

Enterprise Copilot-style licence rollouts before you've shipped anything bespoke.Microsoft and Google will sell you £20-a-seat-a-month AI for your whole team. At a 50-person business that's £12k a year, and in most businesses it produces almost no measurable output because nobody has a defined workflow to plug it into. Ship one bespoke workflow first. Then revisit whether the blanket licence adds anything on top.

Hiring a Head of AI.Unless you're above £20m and have a specific capability gap, the right shape is not a hire. It's a 20% carve-out from an existing operations-savvy team member plus a fractional operator on retainer to keep them moving. We'll come back to this in chapter seven, but if you're currently writing a job spec for this role, put the pen down.

The stack you actually need to start

People ask us what tools to buy before they've decided what to deploy, which is backwards. But assuming you've picked a workflow using the test above, here is the opinionated starter stack that covers about 80% of the workflows an SME will ship in their first year.

Claude(Anthropic's API) as the default language model. Better at long-form instruction-following than the alternatives at the time of writing, better at refusing the bad outputs, and priced cleanly for operator workflows. A £20-a-month Pro subscription covers the owner; API access costs pennies for most SME volumes.

n8n or Make to glue it to your existing tools. Both are node-based workflow builders; n8n is self-hostable and better value as you scale, Make is faster to get started and has a cleaner UI. Pick one and stay on it. For most SMEs, £20–£50 a month is enough.

Zapieronly if you already have it and the workflow is simple. It's the most expensive of the three and the least flexible when you start needing real logic. Fine for a one-step trigger. Not where you'll end up.

Cursor when you need to build something bespoke with actual code. If your workflow needs a small script — a data cleaner, a custom report, a domain-specific parser — Cursor with Claude makes that a two-hour job for a non-engineer with some technical appetite. £20 a month.

That's the starter stack. Total monthly cost for a 50-person business running the three workflows described earlier: about £140 a month in tooling, plus the time of one internal owner. Not £5,000. Not £50,000. £140.

The mindset that separates operators who ship from operators who don't

The technical piece of this is the easy piece. The harder piece is the operator mindset that turns a single shipped workflow into a compounding capability. Three things matter.

Ship before you're ready.The first version of every workflow we run was visibly worse than its third version. That's the point. You cannot tune what you haven't deployed. An 80%-good workflow running on Monday is worth more than a perfect one that ships in six weeks, because you spend the difference iterating on real output instead of imagined output.

Kill things fast.If a workflow isn't paying back after 90 days, turn it off. Don't limp it along. Don't add a feature. The willingness to kill is what lets you start the next one without inheriting the debt of the last one. Every AI project in our portfolio has a named kill-date on it. About one in four gets killed on that date. That's a feature.

Measure quietly.Don't announce the workflow. Don't put it in the all-hands. Don't build a dashboard. For the first 60 days, one person watches the number. If the number goes the right way, you celebrate on day 90 with the whole team. If it doesn't, you kill it quietly and nobody's political capital gets spent. Big announcements before proof is the single biggest predictor of a failed AI project at SME scale.

What the 90-day checkpoint actually looks like

Three months in, the review meeting is 20 minutes long. You look at one number — the one you nominated before you started. You ask the owner three questions. Did it save the hours you said it would? Did anything break that we didn't expect? Do we double down, hold, or kill?

If you double down, the next 30 days is about making it resilient — writing the runbook, handing it off to a second person, building the monitoring you didn't need while it was owned by one human. If you hold, you freeze the workflow at its current capability and pick the next one. If you kill, you turn it off on a Friday, write a one-page note on what you learned, and pick again — this time with better information. All three outcomes are wins. The only losing outcome is the one where nothing happens on day 90 because nobody scheduled the meeting.

Where to go next

If you're ready — three yeses on the readiness test, a named owner, a workflow candidate in mind — the next chapter walks through the exact 30-minute decision exercise to pick the right first workflow from your specific shortlist. If you're not ready, that's useful too: the gap is almost always a named owner or a named kill-date, and both are solvable this week.

The businesses that will have three workflows running by this time next year aren't the ones that read the most. They're the ones that picked a small first move and shipped it. Everything else is commentary.