Digital transformation services: what it means for a small firm

The owner described her in more colourful language than I am going to use here. The short version was that one person ran the finances for the whole group, and without her they would be in serious trouble.

What he wanted, in his words, was a digital version of her.

He did not call that digital transformation. The companies selling digital transformation services would, and they would price it very differently. So this is what it actually looks like, three weeks in, at a healthcare group with several clinics and around thirty staff.

Who sells digital transformation services

I read the pages ranking for this before writing it, which seemed only fair.

One of the largest IT consultancies on the list gives example budgets of $100,000 to $500,000 to get started, and up to $5,000,000 for a company wide programme. Another is written for multinational banks.

Even the one aimed at small firms, a London consultancy writing for businesses of twenty to five hundred people, does not describe a single real client.

Every one of them sells transformation, which is fair enough. So do I, in a smaller way. I sell AI consultancy and training, and you should read this knowing that.

The difference I would point to is that what follows is a real business, anonymised, and I am going to tell you the part that did not work.

What the phrase means at thirty staff

At this size, digital transformation is not a programme. It is two pieces of software that disagree with each other, and one person who knows how to make them agree.

At this group, the appointments and invoices live in a clinic booking system. The money lives in an accounts package. The two do not talk. A lot of the billing goes through insurers, so a payment that lands in the bank has to be matched back to an appointment in the other system, by hand.

The owner told me in our first meeting that they had never really trusted the numbers. That is not a technology problem. It is what happens when the only complete picture of a business lives in somebody's head.

Five jobs, one person, two systems

The second session was with her. I asked one question. Which five jobs take most of your week?

She did not need to think about it.

Matching insurer payments against appointments across the two systems. A monthly turnover report built from an export that cannot be filtered, so she tidies it by hand. Contractor timesheets checked against the diary, appointment by appointment, at the start of every month, which is already her busiest week.

Card payments in the accounts that she cannot identify. And a set of company wellness appointments that have to be rebilled by hand every month.

That turnover export deserves a sentence of its own. Anyone with insurance appears in it twice, and home visits turn up in a clinic's figures because that is the diary they happened to be booked into.

Then there is the daily check. Every morning she looks back at yesterday's clinics to see what was not paid. An unpaid appointment shows up as a slightly lighter green. Reception goes home at six. The clinic runs until seven.

None of this was written down anywhere. She listed all of it in about two minutes. Nobody had asked.

That list is the transformation

This is the part the service pages skip, probably because it does not fill a proposal.

The list is the work. Once the five jobs are written down you can see which ones start in one system and finish in the other, which ones are easy to check, and which one is dull enough to hand over first.

It is also the honest answer to the owner's worry. A digital version of her is not a clever piece of software. It is her five jobs written down, with the dullest of them running without her.

The second problem: every check happens too late

Writing the jobs down showed something else, and I think it matters more than the time.

Every one of her checks happens after the event. An unpaid appointment, an insured patient booked without their policy attached, a card payment nobody can explain. All of them get found days later, by her, rather than at the time by the person who could have fixed them in thirty seconds.

A missing policy is the expensive one. By the time it is noticed, the authorisation can have run out, and the invoice has to be cancelled and raised again.

So some of the work is not doing her checks faster. It is moving the check to the moment the mistake happens, so it lands with the receptionist who made it that afternoon rather than with her the following Tuesday.

What I tried, and what did not work

I had planned to solve the insurer payments job there and then, in the room, as the quick win.

It needed the AI to work through the booking system in a browser. The signal in the building was poor, and I made things worse by pulling the wrong login out of my password manager. After a few minutes I said we would have to leave it and do it another time.

I had picked the most impressive thing to show rather than the most certain one. I had also picked the one job that needed a good connection, in a room without one.

So we tidied the turnover export instead. Strip out the columns she had highlighted, remove every line marked no charge, and ask the AI two things at the end. What is missing from this request? What should I keep for myself?

It worked first time, in front of her. That was a better start than the one I had planned, which is not something I would have said while the signal was going.

Where the AI stops

One of her five jobs ends with changing a batch of appointments in the booking system to free of charge. Could the AI do that as well?

Technically, yes. We agreed it would not.

The AI has its own read only login. It works from the reports and the diary, and it does not go near patient records. A data processing agreement and a data protection impact assessment were drafted before anything touched live data.

That was the most important decision in the whole project, and it was about what the AI is not allowed to do. In a business holding health information it has to come first, not as a caveat at the end.

The same principle runs through the jobs themselves. On the contractor timesheets, the AI compares each one against the diary and lists only the differences. She decides who gets paid what. On the wellness billing, it prepares the month's list and amounts. She raises the invoices.

It reads and it matches. The decisions that cost money stay with the person who is accountable for them.

What the research says about bigger programmes

The most honest figure I know on this is from the UK government. DSIT found that 75% of businesses using AI reported better workforce productivity, but 77% reported no change in revenue, and only 12% an increase.

Productivity is not profit, even when it is real. That is worth knowing before anyone sells you a programme on the strength of the first number.

Deloitte's State of AI survey of more than 3,000 leaders across 24 countries, mostly in large organisations, found 37% using AI at surface level with minimal process change, and only 30% redesigning key processes around it. Deloitte sells exactly that redesign, so its conclusion is also its pitch. The underlying point still stands.

And the scale is worth a sentence. A Federal Reserve Bank of Atlanta survey of US finance chiefs found the median small firm expected to spend about $12,500 on AI in 2026, against $300,000 for a large one. That is a long way from a five million dollar programme.

What it costs at this size

Most of the cost is time, not software. The valuable part of this project so far has been two sessions sitting with the people who do the work.

The software is the cheap part. The AI assistant is around twenty pounds a person a month, and the overnight workflow that moves figures from the booking system into the reports costs about ten pounds a month on the setup I use.

What I will not quote is a total, because it depends entirely on how many of the five jobs you take on, and in what order.

What you could do on Monday

Find the person in your business who would be hardest to replace. Ask them one question: which five jobs take most of your week?

Write the answers down in their words. Next to each one, note which system it starts in and which system it finishes in.

Then pick the one that is dullest and easiest to check. Not the most valuable. The one where you can tell in under a minute whether the output is right.

That list will be worth more than most of the proposals you are sent this year.

Before you change anything, keep a tally for two weeks of roughly how long each of those five jobs takes. A tally on paper is fine. Without it you will never know whether the change saved anything, and neither will anybody you are paying.

The honest limit

We are about three weeks in. I do not have the before and after number yet, and I am not going to guess at one.

The only one of her jobs with a time on it so far is email: twenty to thirty minutes a day, and an hour or more on the morning after two days off. Everything else is a two week paper tally she is keeping now, so that there is a real before figure to measure against rather than an impression.

None of this makes the two systems agree if the diary is wrong. It does not replace her judgement about which insurer to chase this week, or which contractor query is genuine. And it does not remove the risk of relying on one person until her jobs are actually running without her, which they are not yet.

When I have the number, I will write it down here.