Skip to main content

The MD’s guide to AI that pays

For managing and finance directors. Where AI earns its keep in a mid-sized business, where it does not, and how to find out without betting the year’s budget.

Max Revitt, Founder and Director.

Most managing directors we speak to have been told that AI will change everything. The question they bring us is narrower. Which jobs in my business would it do for less than they cost now, and how do I find out safely? This guide answers that, using work we have done for named clients.

Where AI pays

AI pays where the same kind of work arrives every day, follows rules a person could write down, and is checked before it matters. Much of the gain comes from connecting your systems, with AI doing the parts that need reading or judgement. In a mid-sized business that usually means five places.

Admin. Retyping, chasing and copying between systems. At Miranda, a building supplies firm, invoices are now prepared every day from the orders, and the operations team approves each one before it is sent. No order goes more than 24 hours without an invoice.

Orders and invoices. Large customers set the terms: order digitally, invoice in our format, by our deadline. Everything Boxed faced that from its biggest customer. Orders now arrive digitally and are checked against its own data, and invoices go back in the customer’s format. It kept the customer.

Customer replies. Enquiries wait for someone to find time. Every enquiry to our own website now has a researched, personal reply drafted within three minutes, day or night. A person checks it before it goes.

Reporting and checking. AI is good at sorting a long list of differences into the ones that matter and the ones that do not. Global HR, a payroll consultancy, used to prove two payrolls matched by hand, over weeks. Now AI sorts each difference, every decision is recorded, and employees’ personal details never reach the model.

Stock. Stock problems are usually data problems first. When the till, the shop and the warehouse disagree, no forecast will help. Mitchell’s Vintners runs two online shops fed from its till. One fault we found made a sales report show no revenue for products that had sold. Get the numbers to agree first. Forecasting and reordering pay once they do.

Exhibit 1. Where AI pays in a mid-sized business
The jobWhat changesProof
Admin: invoices raised by hand, days after the orderInvoices prepared every day from the orders. A person approves each one before it goes.Miranda
Orders and invoices by email, spreadsheet and phoneOrders arrive digitally and are checked. Invoices go back in the customer’s format.Everything Boxed
Customer enquiries waiting for someone to replyA researched, personal reply drafted within three minutes, checked by a person before it is sent.Revitt (our own)
Reporting and checking: weeks of spreadsheet comparisonAI sorts each difference into explainable or needs investigating. Every decision is recorded.Global HR
Stock and product data that disagree between systemsConnect the till and the shops first, so the numbers agree. Forecasting only helps after that.Mitchell’s Vintners

Source: Revitt client work and our own systems, 2025 to 2026. Each name links to the full story.

Where it does not pay

Some jobs look like a fit and are not. We tell clients to leave these alone, or to fix something else first.

  • Work that rarely happens

    If a job comes up once a month, setting up and checking an AI to do it costs more than doing it by hand.

  • A process nobody can describe

    If three people do the job three different ways, AI will copy the confusion. Agree the process first.

  • Answers you cannot check

    If nobody can tell whether the output was right, you cannot measure the saving or catch the mistake.

  • Data that disagrees with itself

    Five spreadsheets with five versions of the customer list will give you five answers. Fix the data first.

  • A tool bought because a competitor has one

    A licence nobody uses is a cost with no return. Start from the job, then choose the tool.

How to start safely

The safe way in is small and measured. One job, with the controls in place before anything goes live.

  1. 1

    Pick one job

    The one that costs the most time and happens every day or every week. Name the person who does it now.

  2. 2

    Measure it before you change it

    Count how often it happens, how long each one takes and how often it goes wrong. A couple of weeks of honest counting gives you a baseline to judge the pilot against.

  3. 3

    Put the controls in first

    Decide what data can go into the AI, and keep personal details out where they are not needed. Anything that sends, changes or deletes waits for a person to approve it. Every step is recorded.

  4. 4

    Keep private data private

    If the data cannot leave the building, the AI does not have to either. Open models such as Llama, Qwen, Mistral and Gemma run on a machine you own, with no calls to an outside provider and no per-use fees. We benchmark them on your hardware before you commit.

  5. 5

    Widen it only when the numbers are in

    Compare the pilot with what you measured. If it pays, widen it. If it does not, stop. You have spent one pilot, not a year.

What it costs

There are three costs, and only one of them is the AI itself.

The model. Hosted AI from OpenAI, Anthropic, Google or Microsoft is charged per use or per seat. Private models on your own hardware have no per-use fees, but you buy and run the machine. A router in front of several models can send each job to the cheapest model that does it well, with every call logged and costed.

The work around it. Fitting the AI to how your business runs, connecting it to your systems, and building the approvals and the record. The vendors sell the model. This is the work that makes it pay.

Keeping it running. Models change, prices change and providers change their terms. Someone has to watch cost, quality and failures every month, and move you to a better model when one arrives.

Ask any supplier for a fixed price for the first step and a monthly figure for running it. Ours starts with the AI Profit Review at £495 + VAT, credited in full if you order the first piece of work within 30 days.

How to write the business case

Keep it to one page. Your board needs five things.

  • The job

    What it is, who does it and how often.

  • What it costs today

    Hours spent, multiplied by what those hours cost you, plus the cost of mistakes: late invoices, missed orders, rework. Use what you measured.

  • The change

    What the AI does, what a person still does, and what needs approval.

  • What it will cost

    The build, the monthly running cost, and the time your own staff will put in.

  • Payback and risk

    How many months to pay back, what could go wrong, and how you would know.

Some of the return is not hours saved. Miranda and Everything Boxed each kept a customer that required digital ordering and invoicing. Put the revenue at risk in the case, next to the hours. If the payback only works when everything goes right, the case is not ready. Use the figures you measured, not a supplier’s estimate.

What to ask a supplier

A good supplier will answer each of these plainly, in writing.

  • Which model will you use, and why that one? A supplier tied to one vendor will recommend that vendor.
  • Where does our data go, who can see it, and is it used to train anything?
  • What needs a person’s approval, and where is the record of what the AI did?
  • What is the fixed price for the first step, and what will it cost each month to run?
  • How will we measure whether it worked, and what happens if it has not?
  • Can we move to a different model later without starting again?
  • Who runs it after launch, and who do we call when it breaks?
  • Which named clients can we speak to?

Not sure where your business stands? The free AI readiness check asks ten questions and gives you a score out of 30 and your next steps.