The chocolate croissant problem

Emily Goligoski opens most of the focus group the same way. She tells participants: if I ask what you did this morning, you might be tempted to say you ran 10 miles and had a green juice. If the actual answer is that you hit snooze and had a chocolate croissant, I would much rather know that, because the croissant is what helps us design for the people we serve.

Emily is a qualitative researcher and audience strategist. She led audience research at The Atlantic, co-founded the Membership Puzzle Project at NYU, was the first head of research at Charter, and co-authored the 2023 Charter and IRC4HR playbook on AI and worker dignity, along with the 2024 follow-up on what companies and workers are getting right. We had her on the latest Workestration episode, and we kept circling back to the same problem from different directions: most of the information companies are using to steer their AI plans is green-juice data, collected through methods that reward people for giving the acceptable answer.

Think about what the evidence base for a typical enterprise AI rollout. An employee survey, which Emily points out people answer differently depending on which meeting they just walked out of or whatever happened to them earlier in the day. A set of employee personas might have been built from that survey a year and a half ago. Adoption dashboards that count the number of logins. A polished executive demo in which nothing goes wrong. All of it records what people wanted to be seen saying, and very little of it records what happened.

Emily has spent her career developing ways to get to what happened. Here is what she would change, roughly in the order we talked about it.

Demote your personas

Emily's piece Kick out your personas takes aim at a tool HR imported wholesale from product design, so Donna, a former product manager, asked her directly: is HR ten years behind product in discovering that personas are broken, or is the HR use case different?

Emily doesn't want to burn the tool down entirely, and said so. An archetype is shorthand that lets a team ask "what happens to persona B if we prioritize persona A", and it works when you can point to the living people represented by it, when it gets refreshed like a living document, and when it is used alongside other methods.

Her worry, and the reason she wrote the piece, is what happens in employment settings specifically. Personas "can be discriminatory inadvertently," their effects can be dehumanizing, and they can limit people "in ways that would hurt their career potential." A customer who gets mis-bucketed might see the wrong ad. An employee who gets mis-bucketed might not get the promotion, or the job in the first place.

For teams starting from zero, she recommends "user needs", or the related jobs to be done framework out of Harvard Business School. Her reasoning - a persona tends to be a nice slide in a deck until someone remembers to update it, a needs list has to change as employees' lives change and as the larger landscape shifts, so it keeps pulling you back into contact with actual people. What she wants the field to talk about more is what she calls "personas plus": personas next to employee resource groups, panels, co-design, whatever else that adds to the integrity of the research.

Close the loop or stop asking

Take a taxi or buy a mattress online and you are immediately asked to rate the experience. Emily's observation is that companies got very good at requesting feedback and stayed bad at sharing "here is what we are doing with what you shared." She runs her projects on what she calls a "no surprises"s policy: nobody should be surprised by the outcome of something they had even a small hand in, including the people whose feedback you decided to ignore. Donna described her version from her product days, which goes a step further: bring a few people from each affected group into the room and tell the 20% whose preference lost why the decision went the other way. In her experience people accept outcomes they disagree with far more readily than outcomes that arrive unexplained, and the alternative is the feedback fatigue ("did they even listen? why did I bother?") that makes every survey after it less useful.

Who never shows up in your data?

There is a second version of the croissant problem, which is the people who never answer at all. Any self-reported study, Emily reminded us, is bounded by who has the time, the inclination, and the technology to participate. On every project she keeps asking one question repeatedly: whose voice is not represented here?

Her 2023 research found that women, workers of color, and workers over 55 are at higher risk of being disadvantaged by workplace AI. When Stela asked her which findings have aged well two-plus years on, her answer was that she has seen the theme continue rather than reverse. And her explanation runs against the usual story about laggards who need more training. The prevailing mood of AI optimism, the "get on board or get left behind", pushes aside concerns that are entirely reasonable: privacy (she brought up the recent spectacle of celebrities urging women to upload personally sensitive data to AI tools, with no mention of what happens to it), environmental cost, and, most of all, caregiving. Time to sit and play with AI is unevenly distributed, and she suggested a test - look at whoever in your workplace is furthest ahead on AI, and check how many obligations they have outside work. She would rather we stop calling the cautious people Luddites and hear their arguments.

This part of the conversation did not stay theoretical for long. In the weeks before we recorded, Cloudflare had attributed 1,100 job cuts to AI while reporting record revenue, and Bolt's CEO was explaining that he fixed his HR problems by eliminating the HR department. These calls are being made mostly by people who carry the lightest caregiving load, and they set the example for everyone in similar seats.

Emily's alternative frame is dignity as authorship. What helps someone do their job well and produce work they are happy to have their name on? You can research that question but most companies never ask it.

Widen the aperture

Stela asked what Emily would build in place of the usual pilot metrics, the time saved and adoption rates and fractions of FTEs reclaimed. Emily suggested to start before designing the survey questions, at the level of method. A snapshot inherits every croissant problem at once, because the employee who just left a bad meeting will tell you this is the worst place they've worked, and the one prepping their resume will tell you it's the best. So she reaches for what she called "increasing the aperture": time-use studies, diary studies, longitudinal research that follows the same people long enough for patterns to emerge. Even one week in the life of a single employee, she argued, tells you more than 15 minutes each with twenty of them.

Then come the questions, and this stretch of the episode turned into the hosts and guest trading favorites. A few worth stealing:

What feels like it takes an unnecessary amount of effort, that you would want to see transformed in the next one to three years? (Emily's staple. She likes it because it invites an imaginative answer.)

What is your most embarrassing prompt? (Stela's icebreaker. People are already using AI, usually for different reasons than the company assumes, and as she put it, those uncovered use cases are where you need to lean in. The laughing helps people say them out loud.)

How can I tell when you are frustrated? In the flow? Overwhelmed? (Donna's list, inherited from her favorite manager 15 years ago, now pointed at AI: when are you happy handing AI a decision, and when does it detract?)

When is the last time AI did a great job by you? (Emily's own, invented more or less live on the show. Her answer involved AI fact-checking her research synthesis and giving her back two hours of a weekend.)

Fail out loud

The buttoned-up executive demo has its place, and Emily granted that there is a case for showing people what good looks like. But in her follow-on research after the 2023 study, the executives who stuck with her were the ones who said some version of: there is nothing like me failing out loud in front of my employees, hitting an error, and turning it into a teachable moment, sometimes by crowdsourcing what to try next. Stela has watched the same dynamic from the field. Top-down sets direction and opens doors, but the learning spreads sideways, one colleague showing another over Slack what they got working.

Emily runs her Columbia classroom on the same principle. Her graduate students are expected to try AI, to disclose on every assignment what they used, and, if they opt out, to come talk through their reservations with her (she promises not to convert them, she just wants to understand). The payoff this year was that vibe coding finally freed her business-of-media students from obsessing over how polished their final prototypes looked, and they put the recovered hours into the market research and the business case, which is what she wanted graded all along.

She connected this back to Donna's closing question, what leaders will regret three years from now, and her answer was one sentence: believing that AI workplace intelligence comes from the senior levels. The people who will actually teach your workforce what is exciting and what saves time are scattered across the org chart, and most organizations have no mechanism for giving them the platform to role model, to teach others.

Where to start

Swap one persona deck for a living user-needs list, and put a review date on it.

Open your next listening session the way Emily opens hers: tell people the unimpressive answer is the useful one, and that skepticism is a service, not a risk.

For the last survey you ran, publish what you heard and what you are doing about it, including where you went against the feedback and why. While you're at it, look at who never responded, and go find out what that silence is telling you.

Trade one pilot metric for one longitudinal method, even a one-week diary study with ten people.

And in your next rollout meeting, ask the room for their most embarrassing prompt.

Emily closed the episode by quoting Mary Oliver on your one wild and precious life. Stela's parting instruction from her end-of-show summary was to go have the chocolate croissant. Emily accepted, with one amendment: disclose the croissant eating.

Listen to the full conversation at Workestration.ai or on Apple Podcasts, Spotify, YouTube.

References and further reading:

Get the latest episodes directly in your inbox