We made the energy bill clearer. The habit never stuck.
A first-month redesign at Lumen House Energy improved comprehension, then failed its more important test: giving households a reason to come back.
A useful failure at Lumen House Energy.
Current was Lumen House Energy’s companion app for households on time-of-use tariffs. We believed one achievable recommendation each day would make the product useful between bills. I designed that experience, helped launch it, and later recommended removing most of it.
What I owned
I led the first-month experience and interaction design. Nadia co-moderated the diary debriefs; I synthesized the patterns and designed the comprehension study.
What I worked with
The meter integration, billing model, and a business hypothesis that personalized advice would create a weekly habit. Three engineers shipped the mobile work.
The people I worked with
- Priya Shah
- Product lead · activation and retention
- Tom Ellis
- Energy specialist · tariff interpretation
- Rosa Vega
- Mobile lead · meter lifecycle
- Nadia Cole
- Research partner · household diary study
We could not control appliance schedules or change the recommendation service in the first release. The “one useful action” promise was bigger than the available advice.
We removed the feed I had designed.
The week-six result did not last.
Households using at least one recommendation that week
At week six, 38% of the pilot households had used a recommendation during that week. By week twelve it was 14%. The preceding product’s matched billing-period comparison was 18%. The groups were not randomized, so the difference was not evidence that the new UI caused a decline. It was enough to reject our claim that the redesign had created a durable weekly habit.
In debriefs, households described returning around bills and unusual costs. They did not want an ongoing relationship with an energy coach. Better copy could make the advice understandable; it could not create a practical action every day.
Priya and I recommended retiring the daily feed. The team retained the meter waiting state, cost summary, and pause controls, then moved useful notifications into billing events. I had designed a large part of the interface we removed.
The recommendations asked for a different household.
The existing app opened with kilowatt-hours, tariff bands, and a projected annual cost. People could read the chart without knowing what to do differently. The team interpreted low engagement as an information-design problem.
Fifteen household interviews and a ten-day diary study complicated that explanation. People were busy, renting appliances they could not replace, or coordinating bedtime and caregiving. A cheaper time to run the washing machine was not necessarily an available time.
Meter readings could arrive 48 hours late, while the business wanted the product to feel useful every day. We designed an estimated-cost summary and optional advice, but left that tension unresolved.
We tested understanding and mistook it for value.
Nadia and I paired each diary entry with the recommendation that preceded it. We separated “did not understand,” “did not trust,” and “could not do.” The last group was larger than the team expected. I removed the efficiency score and added household boundaries for noise, comfort, and unavailable appliances.
We tested a coaching feed against a plain daily summary. Participants preferred the summary, but several said they would try a practical suggestion underneath it. I interpreted stated willingness as enough evidence to build the feed. That was the weak link in the decision.
The release prototype covered connecting a meter, waiting for readings, an estimate becoming a confirmed amount, and pausing a schedule. Tom checked the tariff language. Rosa and I made sure a missing reading could never appear as a zero-cost day.
The part worth keeping was the bill.
A cost range with a date
Estimated amounts used a range and showed the period they represented. Confirmed readings replaced the estimate without implying a live meter connection.
A recommendation people could reject
“Not practical for me” was available beside the suggestion. We saved the constraint and offered Undo, instead of treating dismissal as failure.
A schedule that could be paused
Comfort settings stayed intact when a household paused automation. Resuming did not require setting up the meter again.
Your meter is connected
Your first readings can take up to 48 hours. We’ll show an estimate only when there’s enough information.
The six-week readout was too early.
Ten moderated sessions tested whether people understood a tariff, a delayed reading, and an estimate. The subsequent 120-household pilot measured actual use over twelve weeks. Those two studies answered different questions, and the second one changed our decision.
Participants read an estimated daily amount as a final bill.
Revision. Used a range, the date, and an “Estimated” label beside the number.
Readout. Tariff comprehension rose from 58% to 82% across matched tasks.
H-06 understood the suggestion but could not run laundry after bedtime.
Revision. Added a practical-constraint dismissal and stopped recommending that task.
Readout. Recovery was clearer, but available suggestions soon became repetitive.
Early recommendation use declined after the sixth week.
Revision. Extended the readout to week twelve and interviewed both returning and lapsed households.
Readout. The coaching hypothesis failed; the bill summary remained useful.
I would require a repeated-use prototype before funding a coaching feed. We had evidence that people understood the advice and liked the idea. Neither demonstrated that the advice would remain useful once the novelty wore off.