AI won't grow your business. It can make growth cheaper to deliver.

A mid-sized company in a competitive market has growth as its main priority. Capital goes to whatever moves the top line. So when AI becomes the hot topic, it looks like one of two things. Either a distraction from the real job, or something you are supposed to be doing without ever being told why.

Both readings are wrong. The same idea corrects them.

AI is unlikely to grow your business directly. Inventing new revenue is hard and risky. It is rarely where a mid-sized company has an edge over larger rivals or nimbler startups. What AI can do reliably is make growth cheaper to deliver. It is a smaller claim, but a more useful one.

The point of the back office

Take the work that scales with the business but does not win customers on its own: procurement, invoice processing, financial analysis, the rest of the support functions. As you grow, that work grows too. Normally its cost grows roughly in step. Sometimes that arrives as a direct cost: double the invoices and soon you are paying double to process the invoices. More often it shows up first as time. The work takes longer and a backlog builds. The delay then sets off its own chain of problems: payments slip, early-payment discounts are lost, the month-end close drags.

AI's value in those functions is that it can break that link. Done well, the cost of the workflow grows slower than the volume running through it. More of every new pound of revenue then survives the journey to the bottom line. That is operating leverage. For a company trying to grow without its overhead growing just as fast, it is worth more than most revenue-side AI experiments.

There is a condition, and it is easy to skate over. The capacity you free has to go somewhere useful. Redeploy it to work that grows the business, or use it to avoid or delay new hiring as you scale. Freed capacity that turns into slack is not a saving. It is a cost you have not noticed yet.

There is a second shape to this. It matters just as much. Sometimes AI does not make an existing task cheaper but gives you something you never had, because it was too expensive or too specialised to do at all. A live view of what your outstanding receivables are costing you at your own cost of capital. A quick check on any invoice of whether an early-payment discount beats that cost or quietly destroys it. These are not efficiencies. They are decisions you were making blind, or not making at all. The economics still work the same way. The value grows as the business grows, while the cost of producing the insight stays small. This is operating leverage too. It arrives as a better decision rather than a lower processing cost. The same discipline still applies. An insight that changes no decision and no number is a dashboard, not a source of value.

So the question stops being "should we adopt AI" and becomes "which workflows will let AI make our growth cheaper to deliver, and which are worth the money and the effort to find out". That needs a way to decide. Here is the one I suggest.

Start with the workflow, not the tool

Most AI adoption goes wrong at the first step. It starts with the tool and then hunts for somewhere to put it. Starting with the workflow turns that around. Make each candidate earn its AI investment. I call the test GRADE, because you are grading a workflow for whether it is worth investing in, the way a lender grades a borrower.

Five gates, each one more expensive to fail than the last:

• Gather the candidates: the high-volume, repetitive support work where growth creates the leverage opportunity, and the decisions you make blind for want of affordable insight.
• Rule out: a fast, cheap screen that most candidates should fail.
• Architect: reimagine the workflow as if you were designing it today from scratch, then remove what stops you.
• Divide the costs: separate the one-off investment from the ongoing cost of running the newly designed workflow.
• Evaluate: two decisions, both of which must pass. First, the marginal running cost against business growth. Second, the opportunity cost of AI-enabling the workflow against your other growth options.

The discipline is in the order. Each gate costs more than the one before. So you only pay for the deep work on the few workflows that survive the early, cheap steps. Most of your effort belongs at the top of the funnel, where saying no is quick and free.

The cheap test that saves the expensive one

The screen is the gate that earns its keep and it costs almost nothing.

For a given workflow, start from the assumption that AI will not help. Assume it will not make the cost grow slower than the volume. Put the burden of proof on anyone who thinks otherwise. Get the people closest to the work in a room. Put one person alongside them who understands what today's AI can and cannot do. The workflow experts often know the process cold and the technology not at all. Give them a couple of hours to make the case in words, without a spreadsheet: here is the specific thing AI changes, and here is why. It breaks the link between cost and volume, or it delivers value that grows with the business while its own cost stays flat.

Then let people who are not invested in the answer attack it. If the case survives, it earns its place in the next step. If it does not, shelve it and write down why. A no today can turn into a yes as your volumes grow or the tools improve.

Two things make this work. Both are easy to get wrong. The people in the room have to be able to disagree without it costing them anything; a screen run by people who cannot say "I don't believe this" is theatre. And the case, either way, has to be specific and not a feeling. "AI could probably help here" clears no bar. "AI removes the manual matching step that today rises one-for-one with invoice volume" is something you can argue about.

Redesign, don't just automate

A workflow that survives the screen has earned the expensive question: what should it actually look like.

The instinct is to take the process you have and bolt AI onto it. Resist that. The process you have is years of evolution and accumulated workarounds. Automating it only makes those workarounds run faster. Design the perfect version on paper first, as if no constraint existed and knowing what AI can now do. That greenfield version is your target. It can include capabilities you never had: the visibility or the analysis that was always worth having and never worth the cost.

Then list the constraints that stop you reaching it, and sort them into two piles. Some sit outside your control, at least for now: a regulation, a core system you cannot replace this year, data you do not own. Those set the ceiling on how close to the ideal you can get. The honest thing is to state that ceiling in the business case, not pretend it is not there. The rest are controllable: an approval hop that exists because someone once got burnt, a manual handoff, poor data that could be cleaned. Those are your redesign. So the workflow you actually build is the ideal design, minus the uncontrollable constraints you must live with, minus the controllable ones you clear. What you are left with is neither the fantasy nor a faster version of the process you started with.

Two buckets, two decisions

Now, and only now, the numbers.

Split the cost into two buckets and keep them apart. The first is the one-off investment to reach that design: cleaning the data, building the integration, redesigning the process, and the change management. That last one is the part almost everyone under-costs, and the part that usually decides whether the thing works. The second is the ongoing cost of running the workflow as the business grows: the usage that scales, the maintenance, the oversight. Do not fold the two into a single per-unit figure. That hides one inside the other, and makes the business case look either falsely cheap or falsely unviable.

Two decisions follow. Both have to pass.

The first is about the ongoing bucket on its own. Does the running cost grow slower than the top line as you scale? Is there real operating leverage here, independent of the one-off spend? If not, stop. No one-off investment, however small, justifies a workflow whose unit economics get worse as it grows. And stress-test that ongoing cost against the chance that today's AI prices are being subsidised and will rise. A case that only works while the discount lasts is not a viable case.

The second is about the one-off bucket. For a growth company it is the harder one. Does spending that money to remove those controllable constraints beat the next best use of it: a salesperson, a product, a new market? In a company where money is scarce and growth is the priority, that comparison is the real gate.

What makes it work, and what breaks it

None of this is heavy if you keep it in proportion. The greenfield question is an afternoon at a whiteboard, not a study. The screen is one meeting. Match the effort to the size of the bet.

Attribution is the honest difficulty. Isolating what the AI did, as against the ten other things that changed at once, is hard. The answer is to baseline first and change one thing at a time. Do not wave the problem away. And the whole method rests on something no framework can supply: a culture where a well-placed sceptic can say "I still don't believe this" and keep their standing. Without that, the screen rubber-stamps whatever reaches it. Every step afterwards just decorates a decision already made.

A growth-phase tool

One word on where this framework fits. The engine of GRADE is growth. It is designed for a company whose volumes are rising and whose aim is to grow without its costs rising in step. That describes much of the mid-market. But it is not confined to the mid-market: a fast-growing large company or a scale-up faces the same dynamic and can use the same test. A mature business in a flat market is a different problem. When volume is not climbing, the logic changes. Operating leverage is a weaker argument. Freed capacity becomes a cost-cutting decision rather than fuel for growth. And the capital choices and constraints look different enough that the test would need rebuilding. GRADE is a growth-phase tool.

Where this leaves you

Growth stays the priority. AI's job, in a company that has to grow and to watch its money, is simple. Make the growth you are already chasing cheaper to deliver. Loosen the link between how fast you grow and how fast your costs grow with you.

GRADE is only a way to keep that honest. Start with the workflow. Make it prove the leverage cheaply. Design it properly. Make the money compete for its place. Run it as a standing practice. The workflows you shelve today are worth another look as your volumes climb and the tools improve. The answer can change with them. The ones that pass now will be few. That is the test working.