Start with the work your team hates.
Ask a team where AI should go first and you will usually hear about something impressive: a customer chatbot, a forecasting model, a tool that writes proposals from scratch. Ask the same people which part of their week they dread, and the answers get much more useful. Somebody re-keys intake forms into the scheduling system. Somebody else spends Monday morning reconciling invoices against purchase orders, line by line.
The dull jobs make the best first projects.
A first AI project has a narrow brief. It has to work, it has to be visible to the people doing the work, and it has to leave them wanting a second one. The ambitious ideas tend to fail the first test, because they touch customers and usually depend on clean data nobody has.
The dreaded chores pass all three. They are internal, so a rough first version harms no one. The person who does them knows exactly what a correct result looks like, because they have produced it by hand a few hundred times. And when a task that took, say, two hours shrinks to twenty minutes, the person who used to lose that time notices right away.
Hated work has a recognizable shape.
Most of the tasks people complain about share a few traits. They repeat on a schedule, weekly or monthly or every time a new client signs. They follow rules that someone could write down, even if nobody has. And they involve moving text or numbers from one place to another, reshaping them on the way: copying fields from an emailed form into a spreadsheet, turning meeting notes into a status update, rewording the same reply to the same question for the fortieth time.
That shape is close to what current AI tools handle well. Reading a document and pulling out specific fields, drafting a first version from a template, summarizing a long thread into five lines, sorting an inbox into categories: these are the jobs where a model saves real time and where a person can check the output in seconds.
The person who hates a task is the best judge of whether AI has done it right.
A town clerk who formats the same permit summary every week can tell at a glance whether the draft is usable, and so can a project manager at an engineering firm who compiles the same weekly report from five people's updates. Years of doing the dull work by hand are what make the experiment safe to run.
People will tell you, if you ask the right way.
"Where could we use AI?" is a hard question to answer on the spot. It asks people to imagine a technology they may barely have used. "What do you put off until the last possible minute?" is easy. Everyone has an answer, and most people have three.
A few prompts help: the task you would hand to a temp on their first day, the job that eats your Friday afternoon, the thing you redo because the first version is always slightly wrong, the email you write over and over with small changes. Collect the answers in writing. Spoken complaints tend to drift into general frustration, while a written list stays specific enough to act on.
The list also shows you patterns across the team. When four people independently name the monthly board report, or the same intake spreadsheet, you have found a shared bottleneck, and the fix will be felt by everyone who named it.
Some of the worst tasks are not AI problems.
Expect part of the list to be the wrong kind of work. Some tasks are hated because the process behind them is broken, like a form that asks for the same information twice or an approval step nobody remembers adding. AI layered on top of a bad process usually produces a faster bad process. Fix those directly, and give the team credit for spotting them.
Other tasks are dreaded because they are hard in a human way. Telling a family their application was declined, or handling a complaint from a long-standing client, is unpleasant because it carries judgment and consequence. AI can help someone prepare for that conversation, but the decision and the delivery should stay with a person.
Set those aside and you are usually left with a short list of repetitive, rule-shaped tasks that each have a clear owner, which is the list to start from.
Why relief builds momentum.
Early AI work often stalls because the benefit is abstract. A leadership team hears that tools are getting better, a few enthusiasts experiment on their own, and the rest of the organization waits to see if it matters. Removing a specific chore from a specific person's week settles the question in a way a presentation cannot.
The person whose task got shorter becomes the most credible voice in the room. Nobody suspects them of selling anything. They have their Monday mornings back, and colleagues who hate a similar task start asking how it was done. That is a better route to the second and third project than any rollout plan, and it keeps the work grounded in what people actually do all day.
A 30-minute worst-tasks workshop
Run this with one team at a time, in a room or on a call, with a shared document everyone can type into.
- Five minutes, alone and in silence: each person writes down the three tasks they most dread, one line each, with how often they come up.
- Five minutes: everyone reads the full list. Mark duplicates and anything named by more than one person.
- Ten minutes: sort each task into one of three groups. Broken process (fix the process), human judgment (keep it with a person), or repetitive and rule-shaped (an AI candidate).
- Five minutes: from the AI candidates, pick one task that is internal, happens at least weekly, and has an owner who can judge a correct result at a glance.
- Five minutes: write down who owns the experiment, what a good output looks like, and when the team will look at the results together, ideally within two weeks.
Keep the full list. The tasks you did not pick this time are the next round of candidates, and the broken-process group is a to-do list for whoever runs operations.
