Atlas Copilot · Schneider Electric
An AI copilot for building operators that shows where every answer comes from.
An AI assistant inside a building-management app. Facility managers ask questions about their buildings' energy data and get answers with sources, a confidence level and drafted actions that a person approves.
- Role
- Digital Product Lead, design lead
- Company
- Schneider Electric
- Timeline
- 2023 to 2024
- Team
- 2 designers, 1 PM, 7 engineers, 2 data scientists
- Platforms
- Web app, Microsoft Teams
1Every claim links to the sensor and time range behind it
2Confidence changes what the copilot is allowed to do
3Actions are drafts until a person approves them
26 → 16 min
Median time to diagnose an alarm
71%
Pilot facility managers using it weekly
4.6 / 5
Operator trust rating after 90 days
0
Actions applied without a person approving
The problem
Answering a simple question like 'why is Building C using more energy than last week?' took an expert, three dashboards and an afternoon. Our first AI prototype answered fast, but cited nothing and sometimes invented equipment.
What I did
I led discovery with 22 operators, defined the 'answer object' every response had to follow, designed the citation, confidence and approval patterns, and set up the evaluation set that gated every release.
What changed
In a 90-day pilot at 14 sites, diagnosis time fell from 26 to 16 minutes and 71% of facility managers used Atlas weekly. The citation-first pattern became the reference for AI features in three other product teams.
01 · Context
The data was there. Getting to it took an expert.
Facility managers look after hundreds of assets across several buildings: chillers, air handlers, meters, switchgear. Our building-management platform collected data from all of them, about 40,000 data points per large site.
Most facility managers aren't energy engineers. When consumption went up or an alarm fired, they needed to compare charts across dashboards, read equipment logs and often call an expert. In 2023, leadership asked for AI in the product within a year. Operators asked for fewer alarms.
02 · Problem
Our first AI prototype was confident and wrong.
The data science team built a first chatbot prototype in a few weeks. In demos it was impressive. In the field it answered confidently, cited nothing, and twice referred to a chiller that didn't exist. In a domain where a wrong setpoint costs thousands of euros or a tenant's comfort, that's a dealbreaker.
First prototype
- ✕Fluent paragraphs with no sources
- ✕Same tone whether it knew or guessed
- ✕Could suggest actions in free text only
- ✕No way to measure if a new version was better
Atlas Copilot at launch
- ✓Every claim linked to a sensor and time range
- ✓Three confidence states that change the UI
- ✓Actions as drafts with impact preview
- ✓300-question test set gating every release
The question we worked on
How might we give non-expert operators expert-level answers about their buildings, without ever asking them to trust a model blindly?
03 · Research
What operators trust.
I ran 22 contextual interviews at six sites in France, the Netherlands and Germany, visited three control rooms, analysed 1.2 million alarm events with a data scientist, and ran a Wizard-of-Oz test where an energy engineer answered questions behind a chat interface.
Operators trust what they can check
In every Wizard-of-Oz session, people asked to see the trend. A chart they could inspect beat a well-written paragraph every time.
“Don't tell me it's the chiller. Show me the chiller.”
Most questions are comparisons
'Compared with last week', 'versus the other building', 'normal for this time of year'. 68% of questions in the test were relative.
“Is this normal? That's all I want to know.”
Alarm fatigue is the real job
Sites averaged 340 alarms a day and about 80% were noise. Operators wanted help deciding what to ignore.
“I need fewer things to look at, not a new thing to talk to.”
| Type of question | Share of questions | First prototype correct | Atlas at pilot correct |
|---|---|---|---|
| Compare with an earlier period | 41% | 52% | 94% |
| Compare buildings or assets | 27% | 38% | 91% |
| Is this alarm important? | 19% | 21% | 88% |
| How do I fix it? | 13% | 9% | Hands over to an expert |
04 · Process
Designing the answer before the conversation.
Instead of starting with chat screens, I started with what a good answer contains. With the data science lead we defined an 'answer object': a claim, the data behind it, a confidence level and the next safe action. Every design and every model change was judged against it.
- 01QuestionNatural language, any building
- 02RetrieveOnly the customer's own data
- 03ClaimOne or two sentences
- 04EvidenceLinked sensors and time ranges
- 05ConfidenceHigh, medium or low
- 06Next actionA draft, never automatic
The parts the first prototype was missing
Autonomous assistant
Could change setpoints directly. Operators refused to switch it on for live buildings. They are accountable, and they wanted to stay that way.
Text-only suggestions
Safe, but people had to re-enter everything elsewhere. Usage dropped after the first week.
Side panel with drafted actions
Lives next to the building page. Answers link into the data, and actions are drafts with an impact preview and an approval step.
05 · Design decisions
Sources, confidence and drafts.
Every answer shows its sources
Each claim links to the exact sensor, time range and chart it came from. Tapping a citation opens the trend right inside the answer, so checking it takes one click instead of a hunt through dashboards.
Confidence changes what the copilot is allowed to do
Instead of a general warning that AI can make mistakes, every answer has a confidence state, and it changes the interface. With low confidence, Atlas asks a clarifying question or hands over to an expert instead of guessing.
Answer and draft actions
Claim with citations. Actions available as drafts.
Answer with caveats
Shows which data is missing or out of date.
Asks first
A clarifying question before any answer.
Says so plainly
Routes the question to an energy expert.
Actions are drafts until a person approves them
Atlas can draft a schedule change or a work order, but it never applies one. Drafts show the expected impact and land in the same audit log as manual changes, so the operator stays the person who made the decision.
Evaluations decide what ships
Together with operators we wrote 300 real questions with correct answers. Every new model, prompt or retrieval change ran against that set. If accuracy fell below 90%, citation coverage dropped, or Atlas mentioned an asset that didn't exist, the release was blocked.
This was the most valuable thing I designed on the project, and it isn't a screen anyone outside the team sees.
06 · Outcome
Faster diagnosis, and a pattern other teams adopted.
After 90 days, 71% of pilot facility managers used Atlas weekly and rated their trust in its answers 4.6 out of 5. The citation and confidence patterns were adopted by three other product teams building AI features.
“I click the number and it shows me the meter. That's why I believe it.”
“It told me it didn't know. Honestly, that's when I started trusting it.”
07 · Reflection
What I learned.
The most valuable design work on this project was the answer object and the evaluation set. Once the team agreed what a good answer looked like, every discussion about prompts, models and screens got shorter.
If I did it again, I'd spend less time on the chat input and more on alarm triage. Operators told us early that alarm fatigue was their biggest problem, and we only started on it after release.
Tools on this project
Figma, FigJam, Python notebooks, Azure OpenAI, Dovetail, Amplitude
Names of the people quoted have been changed to protect their privacy.
Alarm Inbox →