The evidence that it's worth getting right is real. Tiago Forte summarises a Harvard Business School study of 776 professionals at Procter & Gamble in which a single person working with AI on real product innovation challenges could match the performance of a two-person team. Those using AI also reported more positive emotions than those working without it.
That result comes from people doing their actual work with AI, so start with one person's actual work: the task they'd give away first.
Ask each person which task they would give away
One question, asked one person at a time: "Which recurring task would you hand to someone else tomorrow if you could?" Write the answers down as they say them. Don't suggest AI yet and don't pick for them.
Julie Zhuo's ten-step path for AI adoption starts here, with automating something you hate, and her line for why is the reason to ask rather than assign: "A leader cannot mandate enthusiasm. A leader can arrange for relief. (In the margin: ask, don't assign.)"
Pick the task that is recurring, mostly copying, sorting or summarising, and done by someone who wants it gone. Then, before anything changes, have that person time the next two or three runs: minutes a run, runs a week, how often it gets redone. Zhuo's advice is to write down the old baseline first. Without it, the win at the end is an opinion.
Checkpoint
One task, named by the person who does it, with its time written down.
Fix the process before adding AI to it
the step that pays off
Vas, writing from conversations with over 300 CEOs, CIOs and CFOs, puts the usual failure plainly: most companies put AI on top of broken processes and get the broken result faster.
So map the task first. Have the person who does it describe it in their own words, paste that into the prompt below, and go through the QUESTION and CUT lines together. Steps that exist because someone once asked for them and never looked at the output again are common. Cut them before automating anything.
Map the task and find the steps that should not exist
Checkpoint
A shorter task, with the steps AI could do marked.
Build the prompt with the person who does the task
Sit with them and turn what is left into a prompt they can reuse. The prompt below writes it from the task and their notes. Their notes are the part that matters: what good looks like, what the usual mistakes are, what a new hire gets wrong.
Turn the fixed task into a prompt the team can reuse
Checkpoint
A reusable prompt, written from their notes.
Test it on five past runs
Zhuo's step six is evals, and her definition is the one to use: "An eval is a repeatable test of whether the AI did the job right."
Take five past runs where you still have the input and the finished work. Run the new prompt on the same five inputs, then put the pairs through the check below. Add the rule it suggests, run the five again, and stop when the person who does the task would send the output.
Check the output against work you already trust
Pro tip
Keep the five pairs. Every time the prompt changes, run them again before the team uses it.
Checkpoint
Five test pairs the prompt passes, approved by the person who does the task.
Show the before and after to the team
Run the new version for a week and time it the way you timed the old one. Then show the team both numbers, next to each other, with the person who did the work presenting it.
Zhuo: "A win is a result, not a good performance of transformation." A before and after with real minutes in it is a result. A demo is not.
Checkpoint
Two numbers side by side, presented by the person who did the task.
The takeaway
You end up with one task that takes less time, a prompt the team owns, five test cases that say whether it still works, and a before and after someone on the team can point to. The second task is easier to start because the first one has numbers.
The limit: this fixes one task at a time. It won't tell you which roles change or what the business should reward, which is where Zhuo's path ends up after the wins and the systems. And a task with no clear right answer can't be checked against past work, so start somewhere it can.
Variations
Do it on your own task first. Zhuo's path begins with personal belief. If you run the business, pick your own hated task, run all five steps on it, and present your own before and after. It's easier to ask for the baseline when you've kept one.
One shared process next. Once two or three people have a personal win, pick a small process the whole team touches and nobody owns, and run the same steps with everyone in the room.
Keep a short list of what not to hand over. Anything where a wrong answer costs a customer money or trust stays with a person until the check in step four passes every time.
Don't start with a tool purchase. The first task runs in a chat window with a prompt pasted in. Buy tooling when a task is proven and repeated enough to need it.