How do you respond when a stakeholder says 'just use AI' as a solution to an underspecified problem?

Willowgate runs a wedding-planning app that around 40,000 couples are using at any moment, from the first venue tour to the morning of. Handfast is the AI assistant built into it: checklists, vendor research, budget tracking, a running Q&A. Bryony Trelease owns Handfast's roadmap. This is the quarter her CEO's favorite phrase, "just use AI," finally gets its own line on a slide, next to a number that reads almost zero.

How to answer this, stage by stage

Nobody is grading whether you can describe a polite way to push back. They are grading whether you'll notice that "just use AI" is a symptom, not a spec, and catch it before it burns ten weeks proving the wrong build was, in fact, wrong.

Let's learn

Handfast is a chat box tucked in the corner of Willowgate's planning app, the same one every couple sees the day they open an account. It answers questions, nudges couples through their checklist, and compares vendors while they're still deciding. For the first four months of an engagement, before a venue or a caterer is locked in, it's genuinely the best thing in the app: seven in ten couples open it several times a week.

Nothing about the ask itself changed. What Bryony does with the first ten seconds of it changed completely.

Then couples lock things in. A venue. A photographer. Three or four vendors, signed. And that's exactly when Handfast use falls off a cliff: from 71 percent weekly down to 29 percent by the time most couples are five to seven months from their date. Exit surveys all say some version of the same thing: "it keeps telling me things I already did."

Here's the turn. Those aren't extra mistakes piling up. Handfast isn't getting anything wrong, exactly. It's talking to a couple who no longer exists: the one from month one, who hadn't picked anything yet.

At its worst, a couple six weeks from their wedding, the most stressful stretch of the whole thing, gets a push notification suggesting they "compare five florists" for a slot they filled back in March. The assistant didn't get more helpful there. It got loud, at exactly the wrong moment, about a decision that was already made.

What I would leave alone: a one-day experiment nobody's betting the roadmap on doesn't need this ritual. If someone wants to try an AI-written subject line on next week's email, just ship it and look at the open rate. Being wrong there costs an afternoon. The redirect earns its keep on anything expensive or hard to walk back, not on everything that happens to have the word "AI" in it.

The lesson: "just use AI" is not a decision. It's the sound a real decision makes before anyone's done the work of finding it. Somebody has to ask what problem it's actually pointing at, on purpose, every single time, and that costs something real: a day or two of back-and-forth before any build starts, instead of an immediate yes that feels like momentum.

Now here is the same thing as a story

The short version above is what you'd actually say in the room. Read this one when you want to feel exactly what ten weeks costs, not just hear the number.

Every Monday stand-up, part of Bryony Trelease's job was turning whatever leadership had dreamed up over the weekend into a ticket someone else could start on. She was good at it. A vague line in an exec's Slack message would leave her desk by Tuesday as a scoped ask, an owner assigned, a due date attached.

For two years this worked exactly the way it was supposed to. Someone would say "just use AI on the email templates" or "just use AI to sort the support queue," Bryony would route it to the applied-AI pod with a two-line brief, and whatever came back either helped a little or didn't. Either way, a bad week there cost the company a bad week. Nobody minded. She stopped even asking herself whether an ask was ready to be scoped. She just routed it, the way you stop checking a door you've locked a thousand times.

Nobody decided this. It set in over two good years, the way most habits actually do.

Then came the quarter Alderic Brackenmoor, Willowgate's CEO, said the line for what must have been the fortieth time: usage among couples past the vendor-lock stage was still falling, and "this feels like something AI could just fix." Same as always, Bryony wrote a two-line brief and sent it downstream. Ten weeks later, Handfast Concierge shipped: a general-purpose chat assistant that could proactively suggest next steps at any stage of planning. It was genuinely good engineering. The number it was built for, weekly use among couples past vendor lock, moved from 29 percent to 31.

Two weeks after that, for a routine end-of-quarter roadmap review, Bryony did something she'd never done before. She put every "just use AI" initiative from that quarter on one slide, next to the number each one was supposedly for. Six initiatives. Six numbers. Not one had moved by more than a rounding error. The concierge chatbot, the most expensive line on the slide by a wide margin, had moved its own number by two points.

Same six requests, redrawn with the step Willowgate had never built between them and a ticket.

It was never really about any one project failing. Bryony never had a rule for when a vague ask needed defining before it got staffed. She had a feeling: this one seems reasonable, send it. Six reasonable-seeming asks in a row, zero of them defined, zero of them moving anything.

The obvious fix was to just start saying no more, hold the line on anything that showed up without a spec attached. Willowgate actually tried a version of that the quarter before. It didn't work. It just taught people to write vaguer specs to get past her, and twice it read as her blocking ideas that came from the CEO's own mouth, which is a bad way to spend trust you'll need later.

Here's the decision I'd take back instead, and it isn't Bryony's, not really. Back when Willowgate was two founders and a spreadsheet, understanding an ask and staffing an ask were the same five-minute conversation, because everything was small enough that getting it wrong cost a day. Nobody ever split those two steps apart as the company grew, because nothing had gone wrong yet that made the merge visible. The habit that kept the company fast at year one is exactly what let ten weeks disappear at year three.

Run the same kind of quarter again, six weeks later, with the new habit in place. Alderic drops a new one in the roadmap sync: "just use AI on RSVP, our catering partners keep calling the week of the wedding to fix meal counts." Same tone, same shrug, same three words. This time Bryony doesn't reach for a ticket. She asks, right there in the meeting: what outcome are we changing, and how do we know it worked. It takes the rest of that meeting, not ten weeks, to get an honest answer: coordinators keep calling because the dietary notes couples type into a "Notes" field on their vendor page never once reach Handfast's guest summary. Wrong data, not a weak model. The real fix ships in eighteen days. Meal-count complaint calls drop from about 118 a month to 31 within four weeks of it going live.

What I'd tell the version of myself who built that first intake process, back when two founders and a spreadsheet were the whole company: the shortcut that gets you through year one doesn't announce when it's stopped being a shortcut. Somebody has to go looking for the moment it turned into a cost, on purpose, because it will never once raise its own hand.

FLIPS, or the five questions ten weeks could have skipped

Not a trick to sound structured. It's the difference between a habit that survives a good quarter and one that only survives a bad one.

Four setup and payoff letters, and one question that only had to be asked once it actually mattered.

The whole answer, in one picture. Nobody designed a dial. Everybody got a switch, and for two years nobody had to notice.

The AI-specific failure worth naming plainly is missing grounding: Handfast wasn't wrong so much as blind to a couple's real state, which vendors they'd actually booked, what a coordinator had already logged, so a technically fluent assistant kept producing technically fluent advice for a couple who no longer existed. The guardrail is a small eval set built from real couples at different planning stages, checked before every release, holding guest-summary accuracy to a target rate rather than promising it's always right. There's a real trade-off, accepted on purpose: the redirect costs a day or two of back-and-forth before any build starts, in exchange for staffing only asks actually pointed at something. For a cheap, one-day experiment, that overhead costs more than just trying it, which is exactly why it's a redirect for expensive asks, not a rule for every idea in the building.

And if you want to be sure it really works, try it somewhere else

Same five letters, a completely different flip family this time. Nobody hands an ambiguous ask down a chain here. A process just quietly stops being used, and nobody notices for a year and a half.

Greavesmoor Civic Systems sells Stampwell, an AI tool that helps city permit offices sort incoming applications: which ones are complete, which are missing a document, which need a second look. Cosmina Thrumley owns Stampwell's roadmap. When a council member says "just use AI on the backlog," Greavesmoor used to have an actual answer for that: a one-page intake form asking for the specific bottleneck, the target number, and who owns it. For the first year, everyone filled it out.

Same five questions, a completely different way the flip hides. This time the process that should have caught it just went quiet.

Then it stopped. Not all at once. A council aide would call Cosmina directly instead of filling out the form, because the form felt like paperwork for what seemed like an obvious ask. Cosmina, wanting to help, would just start the work anyway. Within eighteen months, nobody had opened the form in ten straight requests. It still existed. It just wasn't part of how anything actually got asked for anymore.

Both stories run F through S. Only one letter, the I, tells you which way the habit actually broke.

F · Cosmina Thrumley, who owns Stampwell's roadmap at Greavesmoor Civic Systems.
L · She stopped insisting on the intake form once council aides started calling her directly instead, because refusing to help on the phone felt worse than skipping a form.
I · The abandonment flip, a different shape from Bryony's. Old setting: every "just use AI" ask goes through a written form naming the bottleneck and the number. New setting: the form sits unopened while requests arrive by phone, by email, in a hallway, anything but the form, and nobody's first move defines anything anymore. Nothing in between: either the form is genuinely how asks arrive, or it's decoration nobody uses.
P · Building a real intake step but never making it the only door, so a live person on the phone was always going to be an easier ask than a form.
S · Cosmina keeps the same question, but asks it live now, on the call, instead of pointing at a form. A council aide says "just use AI on the backlog." Cosmina asks: which permits, and slower than what. Most of the "backlog" turns out to be resubmissions, permits Stampwell had already flagged as incomplete, sitting with no visible reason why, because nobody had ever told Stampwell which applications were resubmissions versus first-timers. Three weeks of labeling fixed what ten more weeks of "smarter sorting" never would have touched.

Swap the trigger and it still runs.
Speed: an interviewer caps you at ninety seconds. Skip straight to it: an AI ask that's never had to define its own outcome will always find a build to attach itself to, no matter how good that build is.
Cost: no budget to build a target-number template from scratch. Reuse the outcome question itself; it costs nothing extra, just refusing to let a ticket exist without an answer to it.
The model got better, for real: say Stampwell's sort accuracy climbs on its own next release. Doesn't matter, maybe matters more. A model getting quietly better is exactly when nobody thinks to check whether it's still solving the actual bottleneck.

Where people run it wrong.
They treat "the form exists" as proof the process is still alive, instead of checking whether anyone's actually filling it out.
They let a live phone call feel like a shortcut around defining the problem, instead of just asking the same question out loud.
They wait for a public, embarrassing moment to force the question, when the whole point of asking early is that you don't need one to show up first.

How to use it live. If an interviewer asks how you'd handle a vague AI request, ask yourself one thing before answering: if I said yes right now, could I write, in one sentence, what number would prove I was right? If the honest answer is no, that's the whole question, answered.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0