Most leadership teams are past the question of whether to use AI. What they want to know now is what they get back for it. That is where it gets murky. The market is loud, every vendor promises a lot, and from the boardroom it is hard to separate real return from expensive activity.
So before the pitches and the pilots, it helps to pin down one thing: where the return on AI actually comes from. Get that right and the business case almost writes itself. Get it wrong, as most do, and you spend money on tools that never reach the accounts.
Here is the fact that reframes it. Take a pound out of your cost base and, depending on your margin, it is worth somewhere between seven and twenty pounds of new sales. The return on AI is easier to prove than the noise suggests, and larger. It sits where most people are not looking: in the cost of running the business.
Licences buy capability, not working automation
Buying an AI licence for every desk feels like progress. It rarely is. A licence buys your people the capability to build something. It does not buy the something. Your team still has to work out what to automate, connect it to your systems, and keep it running while doing their day jobs. Most never get past the occasional time-saving prompt. You have bought potential and called it change.
Changing how a business runs is different work. It means mapping what actually happens today, task by task, then redesigning those processes for a mix of people and AI. That is where cost comes out and return shows up. It does not happen by handing someone a login.
Pay for outcomes, not seats
There is a quieter problem with the licence model. You pay per seat, whether or not anything useful gets built on top of it. Cost scales with headcount. Value does not.
A working automation is the opposite. It is a defined piece of work, done by a system, measured by what it produces. You pay for the outcome and the capacity it runs, not for the number of people who happen to have access. That distinction matters when you take it to the board. Per-seat licensing is a cost that grows with the organisation. An automated process is an asset that pays back and keeps paying. One is a subscription. The other is a change to the shape of the business.
Where the return actually is
Here is the part finance directors recognise straight away.
Most businesses run a net margin somewhere between 5% and 15%. Cross-industry averages put the middle of that band close to 10%. High-volume sectors like staffing, logistics, grocery and hospitality run thinner still, often low single digits.
Now look at what a margin does to the value of a saved pound. At a 10% margin, a business keeps ten pence of every pound of sales. To add one pound of profit through growth, it needs about ten pounds of new sales to produce it. Take one pound of cost out of the back office instead, and the whole pound reaches the bottom line.
That is the leverage. At a 10% margin, a pound saved is worth ten pounds of new turnover. At a thinner 5% margin it is worth twenty. Even at a healthy 15% it is worth almost seven. In a flat market, where every new sale is hard won, that is the most certain profit a business can buy. And back-office cost is exactly what a human plus AI operating model is built to reduce.
Put real numbers on it
Abstract leverage is easy to wave away, so here is a worked example. The figures are illustrative, but the structure is exactly how we model a real engagement.
Take a business with around £2m of payroll tied up in back-office and support roles. In most operations, roughly half of that time goes on manual, repeatable work: re-keying data between systems that do not talk to each other, reconciliations, compliance and eligibility checks, chasing and cleaning records, producing routine reports, handling inbound documents, scheduling. Call it £1m of cost sitting in manual work. Release half of that capacity and you free the equivalent of £500,000 of time a year.
Now set it against cost. Say the first year, an assessment plus a year of building and running the automation, comes to around £70,000. To cover that entire first-year cost, you only need to turn about 14% of the freed capacity into something real. That is the breakeven. Everything above it is return.
Convert half the freed capacity into value, whether by taking cost out or by redeploying people onto work that earns, and that is £250,000 in year one. Net of the £70,000, the business is roughly £180,000 ahead, and the cost is recovered in under four months. From year two, the running cost is lower and the freed capacity keeps paying.
Then apply the leverage. At a 10% margin, that £250,000 of cost removed is worth the same to the bottom line as £2.5m of extra sales. This is why the cost side of the ledger, not the growth side, is where AI earns its keep.
Be honest about what "freed capacity" is
There is a catch worth stating plainly, because ignoring it is how AI business cases lose credibility. Automating manual work does not hand you a cheque. It hands you time back. That freed capacity turns into money in one of two ways. Either you take cost out, by not backfilling a leaver, dropping overtime, or avoiding a hire as you grow. Or you redeploy it, by pointing people at work that earns, such as fee-earners spending recovered hours on billable delivery instead of admin.
Both are real. Neither is automatic. Which is why any honest model treats the share of freed time that becomes value as a decision, not a given. It is also why the first question we help a board answer is not "how much can we automate", but "what will we do with the capacity once it is free".
The objections worth naming
Boards do not push back on the maths. They push back on the memory of the last attempt. The objections are fair, and each has an answer.
We ran a pilot and nothing changed. Pilots prove a tool works in a corner. They rarely change how the business runs, because no one owns the job of turning the pilot into a production process. The fix is to treat it as an operating-model change from the start, with ownership, not as an experiment.
Is this not just another tool to buy. A tool sits on top of the same process and asks your people to do more. This replaces the process. The test is simple: if the work still lands on the same desk in the same way, you bought a tool.
What about accuracy and risk. A fair concern, and the reason we keep people in the loop on the decisions that carry weight, apply checks to the ones that do not, and measure error rates like any other operational number. AI that runs unchecked is a liability. We do not build it that way.
What happens to our people. The aim is not fewer people doing the same work. It is the same people freed from the grind to do the work that needs judgement, relationships and decisions. Capacity gets redeployed, not simply removed. A business that does this well becomes more capable, not more brittle.
Start small, decide on evidence
The sensible way to test all of this is not to bet the business on it. It is to make the first move small and bounded.
Start with a fixed-fee assessment. Over a few weeks, we map how the business runs, size the opportunity against your own numbers, and hand your board a costed, evidence-based case, along with a working prototype built on one of your own processes. If the return is not there, you have found out cheaply and quickly. If it is, you move into build and run with the maths already agreed. You decide the larger commitment on evidence, not faith.
The return is arithmetic, not magic
The ROI on AI is not a mystery, and it is not magic. It is arithmetic, applied to the cost of running your business, multiplied by the leverage your margin already gives you. Point it at the right place and the numbers are hard to argue with.
If you want to see the shape of it on your own figures, our ROI calculator does it in a couple of minutes. When you are ready to turn an estimate into an evidence-based plan, that is what the AI Opportunity Assessment is for.