AI empathy maps skip the interview

If a tool wrote the sticky notes on your empathy map, those notes describe a generic user. Your customer may be a different person, with a different failure, and the map will point the team at the generic one.

Rachel Krause makes the case for Nielsen Norman Group in AI Can't Replace Real Research in Empathy Mapping. She defines an empathy map as a shared picture of what a team knows about a particular user, sorted into what that person says, thinks, does and feels. Teams usually build it from sessions, interviews, tickets, call logs and visits.

She describes the shortcut that shows up when the workshop is booked and the interviews are still ahead. Someone asks the tool to think like a frustrated user, and the wall fills. The tool has no access to your users, to the particular way your product fails, or to the situation someone was in when it failed them. The sentence could describe a frustrated user of almost any product in the category.

She sets those generic lines next to the kind of sentence real research tends to produce. For a grocery-delivery app, the tool says, "I wish the app would ask me before swapping out my items." Beside it she puts someone whose oat milk was replaced by a gallon of whole milk, charged before they saw the notification. For online banking, the tool says, "I want to feel confident that my money is secure." Beside it she puts a person who transfers $1 first to check the account number, then sends the rest. A scheduling tool offers "It's frustrating when the calendar doesn't sync properly," beside someone who found a 2 p.m. double-booked because they were still on the calendar from before a coworker moved the meeting. Course registration offers a complaint that the system is slow, beside four people on a group call at 6

a.m., refreshing together so someone can grab the spot.

You can change a product from the milk sentence, including when the charge happens and what a substitution is allowed to be. "Ask me first" could sit on any app in the category, so a team with only that line still has to guess the rule. A more detailed quote from the tool is still an assumption. Put the $1 transfer on the map after a customer has told you they do that.

Once the evidence exists, she gives the tool a real job. Cluster notes you have already collected. Clean up wording after an insight is tied to something a person said or did. Summarise patterns across a large set of real quotes. Draft a first grouping that a human will correct. She treats four uses as a broken map: quotes attributed to people who were never spoken to, filler in a thin study, a "typical user" standing in for research, and a whole quadrant produced because that box was empty.

Before a note goes on the wall, she wants three questions. Am I handing the tool real data to organise, or asking it to generate data I do not have? Can I trace this note to a session, a ticket or a visit? If I had written this line myself, would I call it a finding or an assumption? An uncomfortable answer means you need more research.

1. Start from a recording, a ticket, a session, or time next to the person doing the task. If the only source is the tool, leave the space blank.
2. Hold the note up to Krause's pairs: a swapped gallon of milk, a $1 test transfer, a 6
   a.m. refresh. If the sentence could move onto a competitor's wall unchanged, go back to the recording.
3. Let the tool cluster and summarise after that material exists, and keep someone who heard the sessions in the review.
4. When a quadrant is empty, book the conversation this week and leave the blank box in the readout.

A customer experience workshop is a good place to do this on one live failure. Bring the tickets. Leave the empty boxes empty until someone on the team has spoken to the customer.