How do you keep a research team connected to user problems without constraining their exploration?

Emberfield Systems builds Wattloom, an app that quietly shifts a home's dryer, dishwasher, and EV charging into cheaper electricity hours and nudges the thermostat to match, without the homeowner noticing a difference in comfort. Nikolina Grewal leads the applied research pod chasing a harder version of the same idea: guessing which appliance is running from the smart meter alone, so Wattloom can work in the millions of homes that only have one meter. A product manager wants to embed one of Nikolina's researchers into Wattloom's two-week sprint, the same fix a sibling pod tried six months ago. Hjalmar Askeland, VP of Product, has to decide what Nikolina's pod actually gets first.

How to answer this, stage by stage

Nobody is grading whether you can name four ways to keep a team aligned. They're grading whether you can tell a real connection from one that only performs connection, and defend the order you'd build them in.

Let's learn

The quarterly deck lived in a shared folder called Research_Alignment_FINAL_v3, four pages, the same four category tags every time: appliance issue, schedule mismatch, comfort complaint, billing question. It was laminated once, that's how permanent it felt.

Wattloom is Emberfield's app for keeping a home's electricity use cheap without anyone having to think about it. It quietly delays a dryer cycle by forty minutes, moves a dishwasher run past the evening price spike, and shifts an EV charge to the middle of the night, all while holding the thermostat steady. It works because it can already tell, from separate submeters, exactly which appliance is drawing power. Most homes don't have those submeters. They have one meter, for the whole house.

So Nikolina's pod is chasing a harder model: guess which appliance is running from the single smart-meter stream alone, no submeters needed. That only works if the model has actually seen the messy real cases: a window air conditioner cycling on its own schedule, a solar panel and battery quietly cancelling out half the signal, a space heater running at 2am in a house with an elderly resident who keeps odd hours. None of that shows up in a category tag that just says "appliance issue."

This was the only door into real households the pod had, for a year, and nobody had built it to be a door.

Here's the turn. A product manager wants to fix this by embedding one of Nikolina's researchers directly into Wattloom's two-week sprint, full time, same as a sibling pod did six months ago for a dynamic-pricing model. It sounds like the obvious fix: get research physically closer to the product. But closeness to a calendar is not the same thing as closeness to a real household. Ruaridh Tindale's pod proved that already, the expensive way.

Ship it blind and the cost isn't a slightly worse model. Nikolina's pod could spend a whole quarter refining a beautifully engineered version of a model that only works on the plain, single-meter, no-solar homes, exactly the homes where the old, simpler model already worked fine. Every home that actually needed the new model, the ones with solar, with window units, with two appliances running at once, keeps guessing wrong, the same way Ruaridh's billing model broke on the accounts nobody on his roadmap had ever mentioned.

What I would leave alone: a separate research thread inside the same pod is testing whether a smaller version of the model can run directly on the thermostat's own chip, to cut how long a recommendation takes to appear. That's a pure speed question. No household behavior in it at all, so none of this connection question applies there.

The lesson: access is not the same claim as calendar. A pod can sit in every meeting in the building and still never touch what's real, if what reaches the room has already been rounded off before anyone in it opens their mouth.

Now here is the same thing as a story

The short version above is what you'd actually say out loud. Read this one for the two quarters a whole research pod spent believing it was finally aligned.

Emberfield's applied research floor is quiet by six most nights, except for the two weeks a quarter when demo prep runs late and someone's ordering in food at eight. Nikolina Grewal has run the load-disaggregation pod for a year and a half, four researchers, one open question: can a single smart-meter stream tell you what's running in a house, without a single extra sensor.

Down the hall, Ruaridh Tindale ran a different pod, chasing a model that could predict when a household was about to get a shockingly high bill, before it happened. Six months ago, his pod got what everyone called a win: full seats on the billing team's two-week sprint. Standups, tickets, demos, all of it. For a while it genuinely felt like alignment. Ruaridh told people it was the best thing that had happened to his research in a year.

Nobody told Ruaridh's pod to stop exploring. Their calendar just never had room left for it.

What actually happened was smaller and slower. Every hour of every sprint already had a ticket number attached to it. Nobody on the pod had time left over to go read a real, messy bill dispute, the kind where a customer calls furious about a charge nobody can quite explain. Instead, the pod worked from the billing team's own success dashboard: clean numbers, clean cohorts, a pilot group of suburban single-family homes on simple flat-rate plans.

The bill-shock model shipped. It worked beautifully on the pilot cohort. Then it hit real accounts: apartment buildings on shared meters, households that switched pricing plans mid-month, customers on a low-income assistance program with its own separate billing rules. Nobody on the pod had touched a single account like that in two quarters. Not because anyone hid those accounts. They just never showed up on a billing-team roadmap, so nobody in the room ever had a reason to mention them.

Now, in the next building over, a Wattloom product manager is proposing the exact same fix for Nikolina's pod: embed a researcher full time, same as Ruaridh's team. Hjalmar Askeland, watching what happened down the hall, isn't ready to say yes just because it looks like the fastest way to show commitment.

Valdis Pellegrino's actual ticket, and the four words that were all Nikolina's pod ever saw of it.

He pulls one real example to make the point concrete. Valdis Pellegrino, a Wattloom user, filed a ticket that read: "my dryer keeps kicking off mid-cycle, right when it's supposed to be running on the cheap window." Somewhere in the support tool that became "Appliance issue, other." That's the version that would have reached even an embedded researcher, unless the actual pipe carrying real language into the room got fixed first. Sitting in the sprint room doesn't fix a filter sitting upstream of the sprint room.

So here's what Hjalmar decides not to do: build the quarterly curated deck that a product marketing colleague half-offered as a compromise. Same failure as Ruaridh's dashboard, just delivered slower and with nicer formatting. And here's what he does instead. Nikolina's pod gets self-serve, low-friction access to the real ticket text and the real usage traces, anonymized by account, not by content. Alongside it, an optional monthly drop-in slot on Wattloom's calendar, no one required to attend, no notes assigned, just an open half hour where a real problem or a real research question can get put on the table. Sprint embedding stays on the shelf, not thrown out, just not reached for first, because it's the one move nobody can cleanly take back once it's running.

What I'd tell myself, watching Ruaridh's pod get praised for finally being aligned: a seat in the room was never the thing that mattered. What reached the seat was.

ORDER, for deciding what actually earns a research pod's trust first

PICK would fit if this were only two options and an even trade. But there are four candidates here, and the real question is which one earns the quarter first, and which one you hold back. That's ORDER's job.

Every box after the first one only works if the first one actually happened.

Reversibility isn't a reason to avoid the harder one forever. It's a reason to be sure of it first.

Only the first milestone was ever forced. Everything after it was a choice, checked against real use before the next one got built.

Three things worth stating directly, since the real judgment sits here. The alternative worth naming and rejecting is the quarterly curated deck, built by product marketing as a middle-ground compromise. It loses because Ruaridh's pod already ran that exact experiment with a live sprint seat standing in for the deck, and a sanitized summary is a sanitized summary no matter how it's delivered or how often it's refreshed. The AI-specific failure worth naming is silent relevance drift: a research pod's whole sense of what's worth exploring quietly stops matching real households, because its own evals got built from the same filtered categories feeding everything else. The guardrail is a standing check, a monthly sample of raw tickets pulled against whatever the pod currently believes its top problems are, so a mismatch shows up as a number, not as a broken model three months later. And the trade-off is accepted on purpose: raw access costs real engineering, an anonymization pipeline, a privacy review, ongoing upkeep, in exchange for a pod that isn't quietly guessing at problems from a summary, which is the more expensive failure and the harder one to see coming.

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

Same five letters, a farm instead of a home, and this time the sanitized summary wears a different name: the quarterly agronomy report.

Loambrook Agritech builds Cropscope, a tool that predicts when a field needs water from soil-moisture sensors and satellite imagery, so a farm can irrigate the right acres on the right day instead of running every line on the same fixed schedule. Svetlana Kirkland, Head of R&D, is deciding what her irrigation-timing research team gets, after a regional agronomist asked to have a data scientist join the farm-ops team's weekly rotation full time.

The report is cheap to cancel and useless anyway. The two options worth keeping sit in the top half of this chart, not the bottom.

Same steps, mapped onto Loambrook. Outcome: protect whether an irrigation recommendation actually holds up on real, mixed-crop, mixed-soil farms, not just the one tidy pilot farm the model was tuned against. Reversibility: a drop-in field-notes session, dropped in on a farm-ops call once a month, is trivial to cancel; a data scientist rotated into weekly farm-ops planning is not, since crews start building their week around that person showing up. Dependency: the model only earns trust if it sees raw sensor drift and the agronomist's actual field notes, not the quarterly agronomy report's rounded-off "irrigation needs improved in the northeast block." Evidence: Svetlana's team shipped self-serve sensor and note access first and watched who actually opened it without being asked, exactly the same voluntary-use test Hjalmar ran. Rank: self-serve raw sensor and field-note access ships first, the drop-in session rides alongside it, full rotation into farm-ops stays held, and the quarterly agronomy report gets cancelled rather than kept as a fallback.

Swap the trigger and it still runs.
Speed: an interviewer caps you at ninety seconds. Skip straight to the rank: ship the raw-access point first, it's cheap and it's the one thing every other option secretly needs anyway.
Cost: there's no budget to properly anonymize the raw complaint text this quarter. Ship a visibly reduced version, real language kept, only the account identifiers stripped, and say so plainly, rather than quietly shipping the old flattened category tags again under a new name.
The model got better, for real: say the lab accuracy on the disaggregation model already looks great. Keep the same access gate anyway. Lab accuracy on curated data was never the same claim as accuracy on a real, messy home.

Where people run it wrong.
They pick the option that photographs well on a roadmap slide, an embedded headcount, over the one that actually fixes the data gap underneath it.
They treat a meeting's attendance as proof of connection, instead of checking who used it when nobody made them.
They let "protect privacy" quietly turn into "strip out everything specific," when the real fix is anonymizing the identifiers and keeping the actual words intact.

How to use it live. Before ranking anything, ask out loud: "which of these can I undo next quarter without anyone getting hurt, and which one, once it's running, is nobody going to want to be the one who cancels it?" Whichever answer feels uncomfortable to say is usually the one to rank last, not first.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0