When I sit down with a team to figure out what they should automate, I don't open with a technology question. I open with this one: "What do you not enjoy doing?"
An HR lead at one of our clients answered it last Monday. "Sickness statistics. I compile them twice a year by hand. Pure admin time-sink. But it's how I catch the early signs that someone's struggling."
That answer is close to perfect, because it contains everything you need to make a good automation decision: a tedious task, a clear rule behind it, and a human purpose hiding inside the admin work. In this article I'll walk you through the simple test we use with Swiss SMEs to decide what to automate first, and just as importantly, what to leave alone.
Why "What Do You Not Enjoy?" Beats "What Can AI Do?"
Most automation conversations start backwards. Someone has seen an impressive demo, the management team asks 'where can we use AI?', and suddenly the company is hunting for problems that fit a tool. That approach produces pilots that impress nobody and quietly die after three months, because they were never anchored in a real pain point. I've seen Swiss SMEs spend five-figure budgets on exactly this kind of solution-in-search-of-a-problem.
Asking people what they don't enjoy doing flips the logic. Tasks that people dislike are usually disliked for a reason: they're repetitive, rule-based, and mentally draining without being mentally demanding. That's almost a textbook definition of what today's AI and workflow automation handle well. The emotional signal ('I dread this every quarter') is a surprisingly reliable proxy for technical feasibility.
There's a second benefit. When the first automation removes a task someone genuinely hates, you don't have to sell the change. The HR lead who no longer compiles statistics by hand becomes your internal champion. Adoption, the place where most AI initiatives actually fail, takes care of itself.
The Three-Question Test
Once the dislikes are on the table, two filters decide whether a task is a candidate for automation. They fit on a Post-it. A third question (the one most consultants forget) decides whether you should actually go ahead.
Can you describe it?
Not do it. Describe it. Inputs, steps, decision factors, clearly enough that a new colleague could follow your written instructions and arrive at the same answer you would. If you can't articulate the rule, no one else can either: not a colleague, not a piece of software, not an AI model. Fuzzy judgment calls that live entirely in one person's head are poor first candidates. Crisp, describable rules are gold.
Does it happen often enough that anyone cares?
A task you do 200 times a year is a different animal from a task you do twice. Building an automation costs roughly the same either way: the build cost doesn't move much with frequency. The payback does. High-frequency tasks pay back the investment in weeks; rare tasks may never pay it back at all. Volume is what turns a clever automation into a worthwhile one.
Do you actually want it gone?
This is the question that separates what can be automated from what should be. Some tasks carry hidden value precisely because a human does them. Before you automate anything, ask the person who owns the task whether they truly want to give it away, and listen carefully when the answer is no.
The first two questions tell you what can be automated. The third tells you what should.
A Real Example: Four Tasks, Four Verdicts
We ran the HR lead's recurring work through all three questions. Here's what came out, including the one task where the right answer was to keep doing it by hand.
Automate
Keep human
Sickness statistics
Clear thresholds (green if absences stay under 42 hours a week, then yellow, then red). Done by hand, twice a year. Automate it, and she can run the report quarterly instead. The time saved is the smaller win. Spotting earlier who's struggling is the bigger one: the automation doesn't just remove work, it improves the outcome the work was for.
CV screening
48 dossiers for one engineering role, 180 for another, over the last six months. The criteria were crisp: required qualifications, language skills, experience thresholds. Crisp criteria plus high volume is the strongest possible automation signal. An AI-assisted first pass ranks dossiers against her own written criteria; she still makes every actual decision.
Second-interview scheduling
Coordinating a second interview across five internal calendars, every single week. Clear rules, recurs constantly, and nobody on earth enjoys calendar Tetris. Automate.
Onboarding questions from new hires
Describable? Yes. A chatbot could handle 'how do I log my time?' easily. High volume? Definitely. And she said no. Because that small, silly question is the door to the real one: 'How are you settling in? Feeling okay?' Verdict: keep it human.
When the Friction Is the Value
The last task on that list matters most. By the first two filters, onboarding questions were a perfect automation candidate. A chatbot would have answered them faster and around the clock. Technically, it was the easiest build of the four. On a spreadsheet, it would have been the obvious next project.
But in a small company, those small interruptions are where relationships form. The new hire who asks where to log their time gets a two-minute answer and a 'how's your first week going?' on top. Those cheap, daily moments are how an HR lead actually knows what's going on in the company. Automate them away and you lose a sensor you didn't know you had.
This is why I distrust automation roadmaps sorted purely by ROI. Efficiency is a means, not the goal. The goal is freeing people for the work only they can do, and sometimes the 'inefficient' task is exactly that work in disguise. The friction IS the value.
Start with the tedious. Free yourself for the strategic and the human. That's the whole game.
How to Run This Exercise with Your Own Team
You don't need a consultant to apply this. You need two weeks and an honest team conversation. Here's the format we use in our workshops, scaled down so you can run it yourself:
Log recurring tasks for two weeks
Ask each team member to keep a simple list: every recurring task they touch, how long it takes, and how often it recurs. No tooling needed: a note on the phone works. The goal is to surface the invisible admin work that never appears in any process documentation.
Score every task against the three questions
In a one-hour session, run each task through the test. Can the owner describe it precisely? Does it recur often enough to matter? Do they actually want it gone? Only tasks with three clear yeses go on the candidate list. Expect surprises: the loudest complaints rarely score highest.
Pick one or two pilot candidates, not ten
Choose the one or two tasks with the clearest rules, the highest frequency, and the most motivated owner. Resist the urge to start everything at once. One automation that demonstrably works builds more momentum than five that are 80% finished.
Measure before and after
Before you build anything, write down the baseline: hours spent, error rate, turnaround time. After four weeks of running the automation, compare. Hard numbers turn a nice experiment into a business case for the next one, and tell you honestly when an automation isn't earning its keep.
Teams that run this exercise usually find three to five strong candidates within a fortnight. More importantly, they also find the tasks they consciously decide to protect, and that decision is worth as much as any automation.
Frequently Asked Questions
- What tasks should a small business automate first?
- Start with high-frequency, rule-based admin tasks that nobody enjoys: compiling recurring reports and statistics, first-pass document screening, scheduling across multiple calendars, copying data between systems, and drafting routine correspondence. These tick all three boxes (describable, frequent, and genuinely unwanted), so they pay back quickly and build trust for bigger projects.
- What tasks should you NOT automate?
- Don't automate tasks where the human contact is part of the value: onboarding conversations, sensitive personnel topics, complaint handling with upset customers, and judgment calls you can't write down as rules yet. If removing the task would also remove a relationship or an early-warning signal, keep it human, even when a bot could technically do it.
- How do I know if a task can be automated with AI?
- Apply the describability test: write instructions so precise that a new colleague could do the task and reach the same result as you. If you can do that, today's AI tools can almost certainly handle it. If the instructions keep collapsing into 'it depends' and 'you'll know it when you see it', the task needs more structure before it's automatable.
- How much does a first automation project cost?
- Less than most SMEs expect. A focused first automation (one task, one team, clear rules) is typically a matter of days of implementation work, not months. The bigger investment is the thinking beforehand: describing the task precisely and agreeing what 'correct' looks like. That's exactly why frequency matters: a task that recurs weekly pays back even a modest build cost within the first year.
- Should I survey my team before choosing what to automate?
- Yes, but ask about feelings, not technology. 'What do you not enjoy doing?' produces far more useful answers than 'where could we use AI?'. People can reliably name the tasks that drain them; they can't be expected to know what current AI tooling can or cannot do. Pair their dislikes with the three-question test and the shortlist almost writes itself.
How we work · Work · Insights