Are AI quick wins worth chasing?

What does a quick win usually turn out to be?

In most firms, the quick win is a step, not a job. Someone finds that a chat assistant drafts a decent first paragraph for a client letter. Someone else gets it to summarise a long document before a meeting. A third person builds a prompt that turns bullet points into a tidy paragraph of prose. All of these work. None of them is a process.

The pattern is easy to spot once you look for it. A quick win attaches itself to the part of a job that is easiest to describe in a sentence, which is almost never the part that takes the longest. The drafting is visible, so it gets fixed. The chasing, the checking, the finding of last quarter's version, the handoff to the person who signs it off, the reformatting into the firm's house style: those are invisible, awkward to describe, and they survive untouched.

So the job shrinks a little in the middle and keeps its shape. The fee earner still opens it, still gathers the same inputs, still waits for the same approval, still finishes it at the same hour of the evening. What changed was the unpleasantness of one stage, not the length of the task.

Why do the savings never show up in the accounts?

Because saved minutes do not pool. Twenty minutes freed on a Tuesday afternoon does not sit in an account waiting to be spent on chargeable work. It gets absorbed by whatever was already queued behind it: an email, a call, a colleague's question. Time recovered in small pieces, scattered across a week and across people, behaves like water on a flat roof. It goes somewhere, but not into a bucket.

There is a second reason, and it is harder to admit. Nobody was measuring the before. A firm that cannot say how long its monthly reporting cycle takes from first data pull to partner sign-off has no way of knowing whether it now takes less. The win is reported as a feeling: this is better, people like it. Feelings are worth something, but they do not appear on a profit and loss account, and they do not survive the first difficult month when the old habits come back.

If you want savings you can see, they have to be measured at the level of the whole job. Measuring return on AI sets out how to do that without pretending to a precision you do not have.

What does a scatter of small wins actually cost?

More than nothing, which is the part that gets missed. Every unofficial shortcut someone invents has to be maintained by the person who invented it. When the template changes, their prompt is wrong and they may not notice. When they leave, it goes with them. When a colleague asks how they did it, the answer is a verbal explanation that neither of them writes down.

You end up with a firm where six people do the same job six slightly different ways, all of them faster than before, none of them the same, and no way to tell which outputs were checked properly. That is a quality problem dressed as an efficiency gain, and it is the ground in which shadow AI grows.

There is also an attention cost at the top. Partners have a limited appetite for this subject. Spend it on a run of small experiments that cannot be shown to have changed anything, and the next proposal, the serious one, arrives to a room that has already decided AI is overstated. We have watched that door close in firms that simply started in the wrong place.

What actually compounds?

One finished process. Not improved, finished: mapped end to end, rebuilt on the firm's own documents and templates, tested on live work, and adopted as the way the job is now done. That is the only unit of change that compounds, and it compounds in three directions at once.

The first is time, and it is time of a useful shape. A process that used to take most of a day now takes the first hour of the morning, every time, for every person who runs it. That is a block, not a scatter, and a block can be filled with chargeable work or sent home at a reasonable hour.

The second is knowledge. Mapping a job properly forces a firm to decide what good looks like: what the report must contain, who checks what, which version of the template is the real one. Those decisions outlive the technology. They are the thing that stops templates rotting again in eighteen months.

The third is reuse. The structure built for one process is rarely specific to it. A firm that has sorted out how client reporting gets drafted, reviewed and signed off has already answered most of the questions it will face when it turns to board and management reporting or technical reports. The second rebuild is faster than the first because the arguments are already settled.

Does that mean small experiments are a waste?

No. Experiments are how people find out what these tools are and are not good at, and a firm where nobody has tried anything is in a worse position than one where everyone has. The error is not experimenting. The error is mistaking the experiment for the implementation, and then being surprised that the week looks exactly the same.

The honest framing is that a licence buys capability and nothing else. What a firm is short of is not capability, it is a rebuilt job. That distinction is the whole of a tool is not a process, and it is why firms with assistants on every desk report that nothing much has changed. For what the general-purpose tools genuinely do well, this is a fair assessment.

How should a firm choose the one thing?

Pick the job that is repetitive, high volume, done by expensive people, and currently done differently by each of them. Those four conditions together mean the hours are real, the rebuild has somewhere to land, and the firm will notice when it changes. A job that happens twice a year, however annoying, is not it.

Then do the arithmetic before you start, so you can do it again afterwards. Hours per week, times fee earners doing it, times your own charge-out rate, times forty-six working weeks. The unbillable hours calculator does that for you, and the process priority scorer helps rank the candidates against each other. If you would rather be told, the audit takes about three minutes and names the process to tackle first.

One finished process beats ten half-finished ones, every time. The ten feel like progress. The one is progress.