AI Strategy: Where to Start When You Don’t Know Where to Start

In 1842, Charles Babbage designed the Analytical Engine — a mechanical computer capable of being programmed with punched cards. It was, by any reasonable measure, one of the most important inventions in human history. It was never built.

Not because the design was flawed. Because nobody could figure out what to use it for first.

The British government had funded Babbage's earlier Difference Engine to calculate mathematical tables. It was expensive but useful — the Royal Navy needed accurate tables for navigation, and human "computers" kept making errors. When Babbage proposed the Analytical Engine — a machine that could, in principle, compute anything — the response was not excitement. It was paralysis.

The machine could do so many things that nobody could agree on which thing to do first. The Admiralty wanted navigation. The Treasury wanted actuarial tables. Babbage himself wanted to prove a mathematical principle. Each stakeholder wanted the machine to solve their problem before anyone else's. Funding stalled. Arguments calcified into positions. The project died. The computer would not be reinvented for another hundred years.

Ada Lovelace, who understood the machine better than almost anyone alive, wrote that the Analytical Engine "weaves algebraical patterns just as the Jacquard loom weaves flowers and leaves." She saw the potential clearly. But seeing potential and knowing where to start are not the same thing. They never have been.

This is the AI strategy problem in 2026, in miniature. The technology can do a remarkable number of things. Most businesses are stuck on the same question Babbage's funders were stuck on: where do we start?

The answer is simpler than the question implies.

Why Most AI Strategies Fail

The typical AI strategy process at a mid-sized business goes something like this. The leadership team decides they need an AI strategy. This decision usually follows a board meeting where someone mentioned that a competitor is "doing something with AI," or a conference where every speaker used the word "transformative" without specifying what was being transformed.

They hire a consultancy. The consultancy spends 8–12 weeks interviewing staff, mapping processes, and benchmarking against industry peers. They produce a 60-page document. The document is well-researched. The formatting is excellent. The framework has a name — something like the "Digital Maturity Index" or the "AI Readiness Quadrant."

The document recommends 15–20 potential AI use cases across the business, ranked by impact and feasibility. Each use case has a two-paragraph description, an estimated ROI range so wide it's meaningless ("£50K–£500K in potential savings"), and a traffic-light risk rating.

The leadership team reads it. They feel overwhelmed by the number of options. They schedule a follow-up meeting to prioritise. The follow-up meeting gets postponed because it's quarter-end. When it's finally held, the team can't agree on which use case to pursue first. The CFO wants to start with finance. The COO wants operations. The CTO wants infrastructure. Nobody wants to go second.

Six months later, they're still talking about starting. The 60-page document is in a shared drive folder that three people have bookmarked and nobody has opened since May.

This happens because the strategy was designed to be comprehensive rather than actionable. A list of 20 things you could do is not a strategy. A strategy is: the one thing you will do first, why you're doing it, what it will cost, what it will return, and how you will know it worked.

The One-Question AI Strategy

Every business that has successfully adopted AI — not talked about adopting AI, not produced a strategy document about adopting AI, but actually deployed it and measured the results — started with a version of the same question:

That's the strategy. Not 60 pages. Not a quadrant. One question. The answer tells you where to start.

The reason this works is that it selects for high-impact, low-risk targets automatically. Expensive processes justify the investment — you don't need to build a business case because the cost is already visible. Repetitive processes are the easiest for AI to handle — they follow patterns, and pattern recognition is what AI does best. Manual processes have the most room for improvement — you're not trying to optimise something that's already efficient.

If you can find a process that is all three — expensive, repetitive, and manual — you have your first AI project. And in every business I've been inside, there are at least five of them hiding in plain sight.

How to Find Your First AI Project in 48 Hours

You do not need a consultancy for this step. You need a whiteboard, a calculator, and two hours with the people who actually do the work. Not the managers who describe the work — the people who sit at the desks and do it. They know where the time goes. They've been meaning to tell someone for years.

Step 1: Map the data movement

List every process in the business that involves a human copying data from one system to another. This is the single richest vein of automation in any business, and it exists in every company I've ever assessed. Invoice processing. Report assembly. Data entry from documents. Bank reconciliation. Client onboarding forms. Payroll preparation. Compliance filing. Management accounts. Write them all down.

Don't filter at this stage. Don't decide what's important. Just list everything where a human is the bridge between two systems that should be talking to each other.

Step 2: Estimate the hours

For each process, estimate the hours per week. Not the hours you think it takes — ask the person who does it. There is always a gap between what management believes a process takes and what it actually takes. The gap is usually 40–60% — the process takes significantly longer than anyone above the person doing it realises.

Be specific. "Invoice processing" might be 2 hours a week in a quiet period and 15 hours a week at month-end. Capture both. The peak number is often where the real pain is.

Step 3: Calculate the cost

Multiply the weekly hours by the average hourly cost of the people doing it. Include salary, NI, pension, and overhead — the fully-loaded cost, not just the salary. For a qualified accountant in the UK, this is typically £30–£45 per hour. For a senior manager, £50–£75.

Now multiply by 48 weeks (allowing for holidays) to get the annual cost. Sort the list by annual cost, highest first.

The most expensive process on the list is your first AI project. Not because it's the most interesting or the most technically exciting — because it's the one where the numbers make the investment obvious to everyone, including the person who controls the budget.

Step 4: Check the rules test

Ask the person who does the process: can you explain exactly how you do this? If the answer is yes — if there's a set of rules, even informal ones, even ones that live entirely in one person's head — then AI can almost certainly do it. Rules-based work is what AI excels at.

If the answer is "it depends" or "I use my judgement" — if the process genuinely requires intuition, creativity, or contextual reasoning that changes every time — park it for now. Start with the processes that follow patterns. Those are the quick wins that fund the harder problems later.

This exercise takes two hours. At the end of it, you know exactly where to start, exactly what it's costing you, and exactly what "success" looks like. No consultancy required for this step.

What "AI Strategy" Actually Means for a Business Your Size

For a business with 10–200 people, an AI strategy is not a document. It is not a framework. It is not a quadrant or a maturity model or a transformation roadmap. It is a sequence of bets, each one small enough to afford, each one measured on its own terms, each one informing the next.

The first bet is usually the cheapest and the most important. It's not really about the technology at all. It's about building the organisational muscle of identifying a problem, testing whether automation solves it, measuring the result, and deciding whether to continue. Once you've done that once, the second project is easier because you've built the muscle. The third project is obvious because you can see the pattern. By the fourth, AI is not a strategy initiative any more — it's just how the business operates.

The companies that are years ahead of their competitors right now did not start with a grand vision. They did not commission a 60-page strategy. They started with one expensive, repetitive process and a willingness to try. Everything else followed.

The Most Common Mistake

The most common mistake is trying to build an AI strategy that covers the entire business before doing anything. This feels responsible. It feels thorough. It feels like due diligence. It is paralysis disguised as planning.

The reason it's a mistake is not that comprehensive planning is bad. It's that comprehensive planning, in the context of a technology that is changing every six months, produces a document that is partially obsolete before the ink is dry. The use cases you mapped in January may not be the right ones by July, because the tools available in July didn't exist in January. The cost assumptions from Q1 are wrong by Q3 because API pricing dropped 40%.

The better approach is to build in small, measured increments. Each increment teaches you something about your business, your team's readiness, and the technology's capabilities that no amount of upfront planning could have predicted. The strategy emerges from doing, not from documenting.

Babbage designed the perfect machine. He spent the rest of his life perfecting the design. It was never built. Meanwhile, less elegant machines — built quickly, for specific purposes, by people willing to start before the design was perfect — changed the world.

Your AI strategy does not need to be perfect. It needs to be started.

The Second Project Is Where It Gets Interesting

Most writing about AI strategy focuses on getting started. That's important, but the real value emerges at the second and third projects. Here's why.

The first project proves the model. It demonstrates that AI can work inside your business, with your data, with your team. It produces a specific, measurable result. It builds confidence.

The second project is where compounding begins. The infrastructure you built for project one — the data connections, the team knowledge, the vendor relationship — makes project two faster, cheaper, and easier. You're not starting from scratch. You're building on a foundation.

By the third project, the pattern is established. Your team knows how to identify automation candidates. They know how to measure outcomes. They know what works and what doesn't. At this point, AI stops being a project and starts being a capability — something the business does naturally, the way it does payroll or client onboarding.

The practices and businesses that are furthest ahead right now are not the ones that started with the best strategy. They're the ones that started earliest, learned fastest, and compounded the most. Their advantage is not intelligence. It's iteration.

Not sure where AI fits in your business? The AI Readiness Assessment takes 3 minutes and shows you where to start. If you're in accountancy, see how practices are already automating compliance work, or visit our accountancy industry page.