Before You Automate Anything: How to Find the Work AI Should Actually Do

Here's a conversation I've had more times than I can count, in one form or another.

"We need to be doing something with AI." Then, a week later, "should we be getting Copilot?" Then, "could we build an agent?" And underneath all three, the real question nobody quite asks out loud: is everyone else further along than we are?

Every one of those starts at the technology. Not one of them starts at the work.

In professional services firms especially, the work that costs the most is the work nobody has ever bothered to name.

I spent about a decade building products for US technology companies before I set up here, a fair bit of it on automation and AI, and I watched this pattern up close often enough to stop finding it surprising. A firm buys a capable tool. Six months later somebody is still trying to find it a job worth doing. The tool wasn't the problem. Nobody had pointed it at anything that mattered in that particular business.

So the question I'd put first is a duller one, and it has nothing to do with AI.

Where is your people's time actually going?

If you can't answer that with numbers, you don't yet know whether you're about to solve an expensive problem or a trivial one. You just know which demo you liked.

What follows is how I'd go about answering it. Six steps, and the technology doesn't turn up until step six.

Step one: find the work that keeps coming back

The work you're looking for is almost never written down anywhere.

Firms document the things that have a name. The client onboarding process has a name. The billing cycle has a name. But the twenty minutes somebody spends every Monday copying figures out of one system so they can paste them into a report has no name, no owner, and no entry in any manual. It's just what Deirdre does. It stays invisible precisely because nobody ever thought of it as a process.

So don't go looking for processes. Go looking for repetition. These are the questions I'd put to a room:

• What did you do this week that you've done almost exactly before?
• What do you copy from one place into another, over and over?
• What are you forever chasing?
• What do you draft again and again from much the same source material?
• Which report do you rebuild from scratch every week or every month?
• What preparation happens before every meeting of a particular kind?
• What do people keep coming to you for, personally, because you're the one who knows?

That last one matters more than it looks. When the answer to "how do we do this?" lives in one person's head, the firm is carrying a risk as well as a workload, and both go on the list.

If you'd rather start from a list than a blank page, I've written up fifteen jobs that come up in almost every professional services firm, with a note on which of them I'd hand over and which I wouldn't.

The people who can answer these questions are the ones doing the work. Not their managers, and not a consultant with a laptop. Ask a partner where the week goes and you'll get an honest answer about the partner's week. The rest of the firm's week is a different picture, and you have to ask the rest of the firm.

Step two: weigh it

Now put numbers on it, because "that job is a nightmare" and "that job costs us four working weeks a year" get very different responses in a management meeting.

The sum is deliberately simple:

How often it happens × how long it takes × how many people do it. Then multiply out to a year.

Say a job takes twenty minutes. Three people do it. It happens four times a week. Across a working year of about forty-six weeks, once you've taken out annual leave and the public holidays:

20 minutes × 3 people × 4 times a week × 46 weeks = 184 hours a year.

If you'd rather not do the sum by hand, the Repetitive Work Calculator does this bit and the sorting that follows it.

A twenty-minute job is not a twenty-minute job. It's four and a half working weeks, and until somebody does that sum it will keep looking like a small annoyance not worth anyone's attention.

Two things people get wrong here.

The first is timing the task and not the tail. The report takes forty minutes to write. It also takes eleven minutes to find the figures, and a further quarter of an hour later on when somebody spots that two of them are wrong. Time the whole loop, interruptions and corrections included, or your numbers will be flattering and useless.

The second is stopping at hours in the wrong place. I'll be straight with you about where I stop: at hours, and at what the firm didn't get done. I don't multiply the total by anybody's salary. The moment you do, the conversation stops being about work and starts being about people, and the frontline staff whose honesty you need for the whole exercise will feel it happen. Everyone in the room can do that sum privately if they want to. Let them.

What I do put beside the hours is the opportunity: what would those hours have gone on instead? That's the number worth arguing about.

Step three: ask how much judgement is really in it

Not everything repetitive should be handed over. The next question is how much of the job is thinking.

It runs on a spectrum rather than a switch:

• Low judgement. The output is predictable, the sources are known, and two competent people would produce much the same thing. Good candidate to hand over.
• Some judgement. Assembling it is mechanical, but deciding what matters isn't. AI does the preparation and a person makes the call.
• High judgement. The thinking is the work. This is what your clients are paying for, and it stays with your people.

How do you know whether a task is safe to automate?

Six questions place a job on that spectrum:

• Is the output predictable enough that you'd recognise a wrong one?
• Is the information it needs actually available, and reliable?
• Can somebody check the result quickly?
• If it goes wrong, is it reversible, and would anyone notice?
• Does doing it well require knowing this client, this history, this relationship?
• Who carries it when it's wrong?

That fourth one deserves more weight than it usually gets. A job that fails loudly is safe to hand over, because the failure announces itself and somebody fixes it. A job that fails quietly, where nobody would notice for six months, is a different proposition entirely, whatever its frequency looks like.

Step four: the five-way sort

Every job on the list gets one of five answers, and I've come to think of this as the five-way sort. This is the part of the day I'd defend hardest, because it's where an honest count stops being a wish list.

Automate it. It repeats, it follows a method, the material is there, and a person can check the result. Prepare it automatically and put it in front of somebody.

Assist it. The assembly goes to AI and the decision stays with a person. Most professional work that involves a client lands here, and that's the right place for it.

Fix the process. This one matters more than the other four combined, and it's the one nobody expects.

Plenty of jobs are painful because they're badly designed, not because they're being done by hand. Three people touching the same document in sequence. An approval step somebody added after an incident in 2021 that everyone now works around. A form that asks for information the firm already has. Automate any of those and you get the same bad process, faster and in greater volume.

The answer is often that a half-day with the four people involved would fix it for nothing, and the automation would have cost real money to make the problem permanent.

Keep it human. The human involvement is the point. Delivering bad news. The first reply when a client is unhappy. Anything written about a person. Advice you're professionally answerable for.

Leave it alone for now. It's genuinely repetitive, but a system's being replaced in the spring, or the volume isn't there yet, or the firm has enough change on its plate this quarter. This is a real answer and it should be said out loud, with the reason attached, so somebody can come back to it when the reason expires.

Firms usually walk in expecting most things to land in the first box. In my experience the middle three are where the useful answers are.

Step five: rank by what you'd actually get back

Now you've got a sorted list, and you still can't do all of it at once. Nor should you.

The instinct is to rank purely by hours saved. That'll steer you wrong, because the biggest number often sits on the hardest and riskiest job in the firm, and starting there is how first AI projects die.

How should you prioritise AI automation opportunities?

I'd weigh six things together:

• The hours it consumes, annualised.
• How often it happens, because frequent work builds trust in the thing quickly.
• Whether the people doing it want rid of it, or quietly like it.
• Whether the client ever sees the output.
• How hard it would be to do, honestly assessed.
• What happens if it gets it wrong, and how fast you'd know.

Put those together and the ranking changes. The best first automation isn't the cleverest one available. It's the one that gives real capacity back, quickly, without anyone having to hold their breath. Get one of those working and the firm will trust you with the harder ones. Start with the hardest one and you'll spend your credibility before you've earned any.

A good first automation is frequent, heavy enough to matter, easy to check, low-risk when it's wrong, and wanted by the people currently doing it.

Step six: only now go looking for the technology

Six steps in, and here's where a tool finally enters the conversation.

By this point you know what the job is, what it weighs, how much judgement is in it, and what you want done about it. That's enough to answer the question properly, and the answer is quite often not AI at all:

• A setting in software the firm already pays for.
• A template, so the thing gets built right the first time instead of fixed three times.
• Two existing systems talking to each other.
• A plain workflow automation with no language model in it anywhere.
• A language model doing the drafting or the summarising.
• A small agent that runs on its own and hands its work to a person.
• Or the process fix from step four, and nothing else.

I'd rather say "you don't need me for this one, change that setting" than sell a firm an agent for a job a template would have solved. The firm remembers which of those two things you told them, and it decides what they let you near next.

What this looks like in an eighteen-person firm

Here's the shape of it, worked through. The figures below are illustrative rather than from a live engagement, because I don't publish client work. The proportions are what I'd expect, not a promise.

Eighteen people. Take a day, get both ends of the firm around one table, and you'd typically end up with something like seventy recurring jobs written down. Most are small. That's the point.

Weigh them, and perhaps eighteen are heavy enough to be worth an argument. Sort those eighteen and they'd fall out roughly like this: eight to hand over, five to assist, three broken processes to fix, and two that stay human.

Now rank them, and the first two are rarely the ones the room expected that morning.

First: chasing outstanding information. Two people, about twenty-five minutes a day each, five days a week. That's 192 hours a year spent asking clients for things they already agreed to send. Nobody in the firm had ever thought of it as a job, because it's never on anyone's task list. It's just the thing you do between other things. It's repetitive, it follows an obvious sequence, and a person approves the wording and takes over the moment it turns sensitive. It goes first because it's dull, safe, and enormous.

Second: the weekly status pack. Four people, fifty minutes each, every week. 153 hours a year assembling something from notes those same people already wrote. Hand over the assembly, and the person who owns it still writes the interpretation and approves what goes out.

And the one everybody was sure would top the list? The proposal that takes three days to write. That's a fix, not an automation. It takes three days because three people rewrite each other's sections in sequence and nobody owns the final version. Automate the drafting and you'll produce a mediocre first draft faster, then still spend three days on the rewriting. Sort out who owns it and you get most of the time back for nothing.

That last one is why counting first is worth the day. It's also the finding that makes people trust the rest of it.

You can run this yourself

Everything above is a method, not a trade secret. You could run a version of it with a whiteboard, a stubborn afternoon, and someone willing to write down what they actually do all week. Some firms should, and I'd rather they did that than nothing.

What's hard about it isn't the knowledge. It's getting a firm to describe its own work honestly, out loud, with both ends of the room listening.

That takes four things a whiteboard doesn't supply. Structure, so the day doesn't drift into a general moan about systems. Facilitation, so the quietest person in the room gets counted as carefully as the loudest. Neutrality, because a partner running the session changes what people are willing to say about the partner's workflow. And discipline, so that when the count says the answer is "fix the process" or "wait", it gets written down that way.

That last one is the real difficulty. A firm counting its own work has an interest in the answer.

Count the work before you buy the technology. Everything else follows from that, including the decision not to buy anything at all.

And once the counting starts turning into building, a quieter question follows: who keeps making these decisions next month? One experiment more or less minds itself. Three or four need somebody minding them. That's a different purchase from a build, and I've written a guide to it.

And when it does go the other way, what you get isn't a firm that feels automated. It's the chasing already gone out before anyone had to steel themselves for it, the pack sitting in the drafts folder waiting to be looked over rather than written, and your best people spending Thursday on the work your clients are actually paying for.

Want someone to run this with your team?

That's the Census. One facilitated day inside your firm: your boardroom, your people, both ends of the table, and a printed roll running its length. No laptops, no recording, and nothing digital leaves the room. Every piece of unbilled work goes on the roll in the handwriting of whoever actually does it. By lunchtime you know what the work is. In the afternoon you weigh it and run it through the five-way sort.

The day ends with a direction rather than a decision, and nobody signs anything. The written AI Roadmap lands within 48 hours. The roll stays with you, on the wall of whatever room your firm argues in, and it keeps being true whether or not you ever work with me again.

It's €1,850, and the whole of it comes off the deployment if you go ahead within ninety days. A fair number of firms finish the day and hear "wait", and that is worth knowing before anybody builds anything.

How a Census day runs →

More on what I do and the firms I do it for: AI consultant for professional services firms in Ireland.