Most enterprise advice on getting started with AI is written for organisations that already have a data team, a Chief AI Officer, and a half-million-pound discretionary innovation budget. That is not most businesses. The rest — owner-operated firms, mid-market businesses, second- and third-generation family companies — are paralysed by a different problem. They know AI is real. They do not know what to do on Monday morning.
The temptation, and the standard consultant pitch, is to commission a strategy. Do not. A strategy without a working example will sit unread and unfunded. The first investment should buy proof, not analysis.
Pick a workflow, not a strategy
The right starting point is a single workflow that meets four tests. Each test is doing real work — it filters out the use cases that look good in slide-ware but go nowhere in practice.
- High-volume. You do it a lot — hundreds or thousands of times a year. Anything below that volume will not justify the change-management effort, regardless of how impressive the demo looks.
- Repetitive. It follows a recognisable pattern. The inputs vary but the shape of the work does not. If every instance is bespoke, you are buying a different problem.
- Document-heavy. Text or structured data goes in. Text, summaries, classifications, or decisions come out. This is what generative AI is good at, this year. Use cases that require physical action, real-time control, or novel scientific reasoning belong in a different conversation.
- Low-risk if it goes wrong. Reversible. Reviewable by a human before anything irreversible happens. Customer-facing failures, financial misstatements, or compliance breaches cost vastly more than the upside of automating them, especially the first time round.
Workflows that consistently pass these four tests in mid-market businesses include: drafting first-pass responses to customer email enquiries; summarising long client meetings into structured action notes; producing first drafts of internal reports from standard data sources; categorising and routing incoming invoices; turning informal voice or text briefs into formatted quotations or proposals; and extracting structured data from semi-structured documents like contracts, RFPs, or planning applications.
Workflows that look attractive but reliably fail these tests: any customer-facing chatbot that can speak unsupervised; any system making credit, hiring, or pricing decisions without a human in the loop; any "AI strategy" that touches the entire firm at once.
The eight-week recipe
For a business that has never used AI, eight weeks is enough to prove value or learn that this workflow was the wrong one. Either outcome is useful.
| Weeks | What happens |
|---|---|
| Week 1 | Pick the workflow. Document how it works today, who does it, and how long it takes. Establish the baseline. |
| Weeks 2–3 | Use an off-the-shelf tool — Claude, ChatGPT, or Copilot — to build a working version. No custom code. No data engineering. No vendor selection. |
| Weeks 4–5 | Run the new workflow in parallel with the old one. Same inputs. Compare outputs. Measure quality, time, and cost. |
| Weeks 6–7 | Refine the prompts, the structure, the human checkpoints. Train the people who will own the workflow. |
| Week 8 | Decide: roll out, refine, or stop. Document what you learned about the firm's appetite, capability, and constraints. This becomes the basis for the strategy. |
Why this works when bigger plans do not
A workflow-first start gives the executive team something they cannot get from a strategy deck: a real artefact, in their own firm, doing real work, with real numbers attached to it. That artefact becomes the basis for everything that comes next — the funding case for scale-up, the conversation with the rest of the leadership, the credibility to make harder decisions about data, governance, and capability.
The fastest way to learn whether AI works for your firm is to use it for one thing, well, in eight weeks. Everything else is preamble.
The firms that move fastest into the next phase — building capability, integrating into the operating model, scaling responsibly under EU AI Act obligations — are almost always the ones that did this small thing first. They are not the ones that bought the biggest strategy.
For the executive sponsor
Before commissioning anything else, identify three candidate workflows that meet the four tests above. Pick one. Give it eight weeks, a named owner, and a baseline metric. Treat the cost as tuition. The lesson it teaches you about your own firm is worth more than the workflow it automates.