Commissioning Assistant · Schneider Electric
A setup flow that configures itself, and shows its work.
AI that reads a building's electrical drawings and network scans and proposes the whole monitoring setup. Installers review only the points the model isn't sure about.
- Role
- Digital Product Lead
- Company
- Schneider Electric
- Timeline
- 2024 to 2025
- Team
- 2 designers, 1 PM, 6 engineers, 1 ML engineer
- Platforms
- Web app, Tablet, on site
1Each suggestion points back to the drawing
2Review sorted by how sure the model is
3Bulk-accept the routine 89%
3 days → 4 h
Average time to commission a site
89%
AI-mapped points accepted without changes
−57%
Support tickets about setup
+19
Installer NPS
The problem
Before a building can be monitored, every meter, breaker and sensor has to be mapped into the software. Installers did it by hand from PDFs and spreadsheets, up to three days per site, and errors showed up months later as wrong energy reports.
What I did
I built the business case, led the design of a review-first setup flow on web and tablet, worked with our ML engineer on confidence thresholds, and tested on live sites with installers in the Netherlands and France.
What changed
Across the first 120 sites, commissioning dropped from three days to four hours. Installers accepted 89% of AI-mapped points as proposed, and setup tickets fell by 57%.
01 · Context
The least loved part of an installer's job.
Commissioning is the step between installing hardware in a building and seeing data in the software. An installer maps every physical point, such as a meter on floor 3 or a breaker in panel B, to its place in the system. A mid-size office has around 450 of them.
Installers are paid per site, not per hour. Every hour spent typing point names from a PDF is an hour they're not paid for, and every typo becomes a wrong number in a customer's energy report later.
- 01Install hardwareMeters, gateways, sensors
- 02Find drawingsOften outdated PDFs
- 03Scan networkList of devices found
- 04Map by hand450 points, one by one
- 05Check and fixWalk the building again
- 06Go liveCustomer sees data
Where two of the three days went
02 · Problem
Three days of typing, and errors found months later.
Errors were the hidden cost. A swapped meter would show the wrong floor using the most energy, and the customer would find out during an audit. Support spent a large part of its time fixing setups after the fact.
The question we worked on
How might we make setup feel like reviewing a proposal rather than filling in a form, while keeping the installer responsible for the result?
03 · Research
Watching eight commissioning jobs.
I observed eight commissioning jobs from start to finish, interviewed 15 installers and partner companies, and analysed a year of setup-related support tickets.
Installers trust drawings more than software
When software and drawing disagreed, installers believed the drawing, even when it was out of date. Any suggestion had to point back to the drawing.
“The drawing is the truth. Even when it's wrong.”
Most points are boring, a few are tricky
About 85% of points followed standard naming. All the risk sat in the other 15%, and that's where installers wanted to spend their attention.
“Ninety percent of this is copy and paste.”
Plant rooms are hostile places
Gloves, no signal and poor light. Anything that needed precise tapping or a connection failed on site.
“Down here, nothing loads.”
04 · Process
How proactive should the AI be?
Fully automatic go-live
Fastest on paper. Installers sign off on the site and refused to put their name on something they hadn't seen.
Suggestions inside the old form
Kept the friction of the form and saved about 20% of the time.
Proposal, then review by exception
The assistant proposes the whole setup. Installers review the uncertain points, check them against the drawing, and bulk-accept the rest.
05 · Design decisions
Review by exception.
Every suggestion points back to the drawing
The review screen puts the electrical drawing next to the proposed setup. Each suggested mapping highlights the part of the drawing it came from, with the model's confidence. Because installers trust the drawing, showing it next to the suggestion was what made them trust the suggestion.
Confidence decides which queue a point lands in
We worked with our ML engineer to set thresholds from the real error rate at each confidence level. Points sort into four queues, and each queue asks for a different action.
412 points
High confidence. Bulk accept in one click.
38 points
Naming differs from the drawing. Quick review.
3 points
Meter M-12 appears twice. Pick the right one.
1 point
No device found for breaker Q4. Skip or add by hand.
A tablet mode for the plant room
Some checks have to happen physically. The tablet mode works offline, uses large targets and high contrast, and only shows the points that need someone to look at the actual equipment.
Sign-off stays with the installer
The last screen summarises what the installer accepted, changed and added, and how long it took. The installer signs off with their name. That signature was non-negotiable for them, and keeping it made the whole feature acceptable.
06 · Outcome
Hours instead of days.
“I check the strange ones and accept the rest. I was home before dinner.”
“Fewer setup tickets meant we could finally work on the hard problems.”
07 · Reflection
What I learned.
Proactive AI works when it respects who is accountable. The installer signs off on the building, so our job was to make that signature quick and well-informed.
I'd collect correction data from day one next time. We only started feeding installer edits back to the model in month seven, and accuracy improved noticeably once we did.
Tools on this project
Figma, ProtoPie, Azure ML, Jira, Dovetail
Names of the people quoted have been changed to protect their privacy.
Quote-to-Order →