The first question I ask in an adoption workshop has nothing to do with AI. It is this: what work here gets done twice, and who does it the second time?
I have tried a lot of opening questions over the years. That is the only one that has never given me a useless room.
The obvious alternative is to ask where AI could help, and it fails in a particular way. You get the answers people think you want. Summarising long emails. First drafts. Meeting notes. Not wrong, but the same three answers in every organisation I have stood in, so they cannot help you tell one organisation from another. Telling them apart is the entire job.
It also asks two hard things at once. Know what the technology does, then reclassify your own work in terms of it. That is a modelling exercise performed in public, by people who have had a forty-minute overview. Most of them sensibly guess, and a guess offered confidently in a workshop has a habit of turning up later as a requirement.
Duplication is a question about their week, not about the product
Ask about duplication and people stop performing. They start describing their week, because everybody knows precisely which part of their week they are doing for the second time. The answers arrive specific and slightly annoyed, the register you want. Irritation at the work is signal. Enthusiasm about a tool is not.
What comes out is concrete. The figures rebuilt every month by two teams from the same source, formatted differently, then reconciled by a third. The form filled in on paper and typed into a system by somebody else. The status report written for a meeting whose attendees have already read the update it was copied from. The client note keyed into the CRM and again into the practice management system, because the project to join them was descoped in 2019 and nobody has raised it since.
I design learning for a living, so I think about questions as instruments rather than as conversation. A question works when it can be answered from memory instead of from inference. “What gets done twice” is a retrieval cue. It points at an episode, and people are good at finding episodes. “Where could AI help” asks for a judgement about a category that has to be constructed on the spot, and constructing categories in front of your manager is not a task most people want.
The second property that matters is status cost. “What do you find difficult” asks people to describe their own limits, so you get careful answers, correctly hedged and no use to anybody. “What gets done twice” invites people to accuse the process instead, which is safe. Nobody has ever been performance-managed for saying that a spreadsheet gets rekeyed.
The second clause is the one that draws the org chart
Who does it the second time. That half is doing more work than the first half, and it is the half people leave off when they borrow the question.
It is almost never the person who caused the duplication, and it is almost never anybody senior. That is exactly why the duplication has survived. The cost lands somewhere it is not measured, not invoiced and not complained about, on people who long ago concluded that this is what the job is. Duplication that lands on a director gets fixed within a quarter. Duplication that lands on an administrator can run for a decade.
Listen for names. In a good session two or three names come up in unrelated answers from unrelated teams. The person named three times is who you should talk to next, and they are almost never on the stakeholder list you were given, because stakeholder lists are made of people who own things rather than people who absorb things.
You also get the real hand-offs rather than the ones on the process map. A hand-off on a diagram was designed by somebody. One that only surfaces when a person says “and then I send it over and someone in the service team puts it into the other system” was never designed. It grew, and nobody has ever had to defend it in a meeting. Those are the ones worth looking at.
The follow-ups are where the hour is won
The opening question buys you a list. The list on its own is not worth much. What makes it actionable is a small set of follow-ups asked in a fixed order, and the order matters as much as the wording.
- How do you know the second copy is right? Reconciliation is the hidden work and it is usually larger than the copying. If nobody checks, you have found a risk rather than an inefficiency, and that belongs in a different conversation with different people in the room.
- How often does it happen? Frequency, not duration. People know how often. They guess at how long, and a guess said aloud gets anchored by whoever spoke last, then quoted back at you for two years.
- Who asked for it? Frequently nobody living, or nobody still in the role. Reports outlive the person who requested them, and the current author often has no idea who reads it.
- What would break if it stopped for a month? Ask it neutrally and mean it. Sometimes the honest answer is a paper that goes to a committee which no longer meets.
- Where does the first copy come from? This decides whether an agent has anything to work with. If the source sits in a line-of-business system nothing can reach, no amount of building helps, and it is better to know in the workshop than in the third sprint.
- Show me. The best of them and the one most often skipped. People misdescribe their own work, not dishonestly. They describe the documented version, because that is the version that has words attached to it. Ten minutes of watching produces detail no workshop can.
About half of what the room lists does not need an agent
It needs somebody senior enough to say the second copy can stop. I say that in the room, early, and it is not what the person paying for the session hoped to hear.
The reason to say it early is that building the agent makes the duplication permanent. An agent that produces the second copy has a maintainer, a cost line, an owner and a support route. From then on it gets defended, and the meeting where somebody might have asked whether the second copy was needed at all will never happen. You have automated a mistake and given it a budget.
What survives that test is worth doing, and it has a recognisable shape. The two systems will not be joined inside the timeframe that matters, usually for cost reasons rather than technical ones. The second copy is a real change of form rather than a rekeying: different audience, different level of detail, a judgement somebody makes each time. The source is somewhere the tool can actually reach. The work recurs often enough for a build to pay back, and is stable enough to survive next quarter.
The questions I have stopped asking
Four retirements and one demotion.
“What are your biggest pain points?” Retired. You get the rehearsed answer, delivered to its fourth supplier of the year and polished by repetition into something that sounds like insight. It also invites complaints nobody present can act on: too many systems, not enough people, the last reorganisation. All true, all outside the gift of the room, and the room deflates within ten minutes.
“If you had a magic wand, what would you automate?” Retired, and I am embarrassed by how long I kept it. It asks people to design a solution while knowing none of the constraints, so you collect a wish list and spend the second half saying no to things you invited.
“How much time does that take you?” Retired in that form, replaced with frequency. Asking for hours in a group makes people invent a number they then feel obliged to defend, and the first one said aloud drags every number after it. I used to collect those figures and put them in slides. They were real as perceptions and close to worthless as evidence.
“Who owns this?” Retired as an early question. Asked early it makes the room defensive, because ownership questions sound like blame questions when nobody yet knows what you will do with the answer. I ask “who would notice if it stopped” instead, which draws the same map without requiring anybody to admit to anything.
“Walk me through your process.” Not retired, demoted. It produces the documented process, which is a genuinely useful artefact as long as you know that is what it is. Ask it after the duplication question rather than before, and the gap between the two accounts becomes the finding.
How I run the twenty minutes
The structure is the instrument, and it is easy to break by accident.
No slides before the question. It is the first thing that happens, inside the first five minutes, before anyone has been shown what the tool can do. Once people have seen a feature they answer in terms of the feature, and you cannot un-show them.
Two minutes of silent writing before anybody speaks. Skip that and the first confident answer sets the frame, and everything afterwards is a variation on it. This is the cheapest intervention available in workshop design and hardly anyone does it.
Junior people answer first. If a director goes first you get the organisation’s official account of its own work, which is the account already written down in the document you have read.
Capture the answers in the words they were said in. Do not tidy “we type it in again” into “manual data re-entry between systems”. The phrasing is the evidence, and the moment you translate somebody’s sentence into process language you have started agreeing with the process document instead of the person.
Stop at about a dozen items. A longer list is the same list with variants, and it makes the sorting harder without making it better.
Then mark each item as stop, join or automate, in front of everyone, and let people argue about the marking. If I could keep only one part of the session, I would keep the argument.
You can run this without me. It takes a room, a whiteboard, twenty minutes and one person willing to write down an answer that implies their own department is the problem. The last of those is the hard part, and no question I can design will fix it.