Does this mean the AI tools themselves are not worth using?

No, they work. The point is that a tool used by one person in a browser produces a personal saving, and the firm only sees it when the task around the tool changes: where the trigger comes from, where the data comes from, who approves the output and where it lands. Same tool, different result.

How much data centralisation does a small firm actually need?

Less than the phrase suggests. For one workflow you need the information that workflow touches to be reachable from one place, in a consistent format, without a person re-typing it. That is often a single spreadsheet or one system's export rather than any large project.

What is a reasonable result to expect from the first attempt?

One task that runs without anyone remembering to run it, and a before-and-after number you trust. Hours per week is enough. If a first project cannot produce that, it was scoped too broadly, and cutting it back usually costs nothing.

Two-thirds are using AI. Five per cent get anything back

Why does using AI every day produce nothing measurable?

Because the use is personal and the work is shared. Someone drafts faster in a chat window, saves fifteen minutes, and hands the result into the same queue it always went into. The queue did not get shorter, and nobody wrote down how long it was to begin with.

Three habits keep it that way.

The tool sits beside the work rather than inside it. A person copies text out of one system, pastes it into a browser, and pastes the answer back. That works, and it stops working the week that person is away.

The saved time is not reclaimed. Fifteen minutes returned to a full day becomes fifteen minutes of something else. Time shows up in the accounts only when it changes what the firm can take on, or when the firm stops paying for it.

And the output still needs the same sign-off it always needed. A draft written in four seconds that then waits three days on the principal's desk was never held up by the writing. The constraint sits somewhere else.

What does an enabled business actually mean?

Three things, and none of them is a piece of software. In Deloitte's framing, an AI-enabled business has a strategy embedded in its core processes, trains its people to use the tools, and holds its data in one centralised place. Put plainly: the work knows where AI sits, the people know how to use it, and the information it needs is somewhere it can reach.

Two of those three are not technical problems. Most small firms fail the training and the data long before they fail the strategy, and neither gets fixed by buying anything.

How do you tell which side of the line you are on?

Run the test on one task rather than on the whole firm. Pick something the business does at least weekly, then answer these honestly.

Can you name the task, or only the tool? "We use AI for admin" is a tool answer.
Would it still run the same way if the person who set it up took a fortnight off?
Is how it works written down anywhere, or does it live in one person's chat history?
Does the AI reach your own data, or does somebody paste the data in each time?
Do you know what the task cost before? Any number counts, including a rough hours-per-week guess.
Has anything downstream changed: a shorter queue, an earlier reply, a report nobody has to build?
If it produced something wrong tomorrow, who would notice, and how?

Mostly yes means you have a system. Mostly no means you have a tool that one person is fond of. Neither is unusual, and plenty of firms have run for years on the second without it costing them anything they could see.

What does a system look like on a single task?

Take a weekly report that somebody assembles by hand. In the tool version, that person asks a model to write the commentary and pastes it into the usual document, and it still takes most of Monday morning.

In the system version the numbers are pulled from the source on a schedule, the draft is built with last week's figures and this week's movement, and it is sitting in the owner's inbox before they arrive. They read it, correct what is wrong, and send it. The report gets produced whether or not anyone remembered, and the person who used to lose the morning to it is doing something else. That shape has its own article because it is the most common first build.

The difference is not the writing. The trigger, the data, the assembly and the delivery have all moved, and the human step has shrunk to the part that needs judgement.

Isn't building that a bigger project than a small firm can run?

Not while it stays at one task. The usual failure is not moving too slowly, it is opening with a strategy document that covers everything the business does. Those take a month, get filed, and produce nothing anyone can point at.

One workflow taken all the way to running unattended teaches a firm more than a plan covering ten. It also produces the thing that makes the second one easy to argue for: a real number for what the first one saved.

Where should a firm start?

In this order. Each step is small enough to finish inside a week.

Pick the task, not the tool. The one costing the most hours, not the one that irritates you most. Frequency beats annoyance almost every time.
Write down what it costs now. Hours a week times a realistic rate. Rough is fine, and there is a method for it if you want one.
Map the current steps in plain sentences. Where the trigger comes from, what gets looked up, who decides what, where it ends up. A good share of the waste is visible here before any tool is involved.
Automate the assembly and leave the decision. Fetching the data, drafting the output and putting it in front of a person is machine work. Deciding whether to send it is not.
Hand it to somebody else for a fortnight. If it only runs when the person who built it is there, it is not finished.
Measure the same number again. Then decide whether the next one is worth doing.

Two and six are the steps firms skip, and skipping them is a fair part of why the five per cent figure looks the way it does. Without a before and an after, the second project has to be argued for on faith.

What to do next

Take one task this week and do step two only. Hours, a rate, a total, written down. It takes about twenty minutes and it either surprises you or tells you that you picked the wrong task, and both of those are worth knowing before you spend anything.

If the number is big enough to act on, the next move is mapping the steps rather than shopping for software. That is where Shift starts as well, mostly because the map usually shows that several of the steps do not need to happen at all, and removing a step is cheaper than automating it.