CE CoPilot: the service manual, answering back

At a global semiconductor equipment maker, customer engineers keep chip-fab tools running, and when a tool faults, the fix is somewhere in thousands of pages of approved documentation. CE CoPilot lets an engineer describe the problem and get a step-by-step answer from those documents only, with its sources and a confidence score. When an answer is wrong, the admin console turns the complaint into a fix.

Client
Customer engineering at a global semiconductor equipment maker
Role
End-to-end design: mobile app, web app and admin console
Built at
LeewayHertz
Year
2026
CE CoPilot across its surfaces: the admin console overview behind three phone screens, a new chat, a step-by-step answer with sources and confidence, and an answer drawn as a flowchart.

The context

The fix was in the documents. Finding it was the problem.

When a tool faults, a customer engineer needs the right procedure, fast, and it sits somewhere in thousands of pages of approved documentation. CE CoPilot meets them in three places.

Mobile app: for the engineer standing at the tool.

Web app: the same chat at a desk, with the sidebar open.

Admin console: usage, answer quality, the knowledge base and feedback.

A wrong torque value on a fab tool is not a typo.

The brief

Three rules the product could not break

Everything in CE CoPilot follows from them.

  • 01Only approved documentsAnswers come from approved technical documents and structured datasets, nothing else. Off-topic questions are declined.
  • 02Always show the sourceEvery answer names its documents and carries a confidence score an engineer can question.
  • 03Safety before stepsLockout, venting and preconditions come before any diagnostic step, every time.

Design decisions

Small calls that kept the product honest

Each one protects the engineer’s trust, or the truth of a number in the console.

  • Decline, but show the way inAn out-of-scope reply lists what to add, so a refusal becomes a better question.
  • One chat, one issueResolving closes the chat, so every resolution rate in the console stays true.
  • Adoption on an average dayNot “ever used”, which only climbs. The daily share is the number that moves.
  • Each number appears onceA tile gives the headline, a chart the trend, a table the detail.
  • Fixes are scoped and versionedA fix for Singapore stays in Singapore, and none is ever deleted.

In practice

An engineer at the tool, from question to resolved

Five steps on the mobile app.

01

Ask the way you’d ask a colleague

Type the symptom in plain words, paste an error code or attach a script. If a question is out of scope or too thin, CE CoPilot asks for the platform, part or code.

New chat: a greeting, one field, and the scope one tap away.
Describing the issue in plain language.
A script attached to the question.
An out-of-scope reply asking for the platform, part or problem code.

While it works, it says what it’s doing

Five honest states instead of one spinner. A seven-second wait feels shorter when you can see the work.

Status line: Thinking
Status line: Working on it
Status line: Reading context
Status line: Searching knowledge
Status line: Reading documents

02

Read a procedure, not a paragraph

Every troubleshooting answer uses the same five parts, so an engineer always knows where the safety notes and the fix will be.

  1. 1Issue description
  2. 2Safety and preconditions
  3. 3Root cause
  4. 4Diagnostic steps
  5. 5Corrective actions
An answer opening with the issue description, safety and preconditions, and root cause.
Diagnostic steps as short, checkable bullets in order.
Corrective actions, then the sources, the confidence and the option to make a diagram.

03

Question any answer

The footer is where trust is earned: the sources, a confidence score explained in plain words, and a way to say what was wrong. Closing a chat asks for a category.

The confidence score explained: how closely the documents match, lowered by details still missing.
Share feedback: status, category and what went wrong.
Mark as resolved: closing the chat asks for a category.

04

Follow it at the tool

Long procedures with branches read better as a flowchart. One tap turns an answer into one, with a legend for every shape.

An answer drawn as a flowchart, to pinch, zoom and pan in place.
The flowchart legend.

05

Sign in once, find any chat again

Microsoft sign-in with the company’s own branding, an error that says how to fix it, and a history where every chat carries its category and status.

Welcome screen with one button: Sign in with Microsoft.
A sign-in error that says what failed and how to fix it.
Recent chats grouped by day, each with its category and status.

The feedback loop

Wrong answers become better ones

This is the part of the product I’m proudest of. A thumbs-down in the app isn’t the end of the story: it starts a fix that every later answer benefits from.

  1. 01An engineer flags itStatus, category and what was wrong, from the answer footer.
  2. 02It lands in the inboxSorted by type: incorrect, outdated, incomplete or a safety concern.
  3. 03An expert writes the fixScoped to one site or one tool family, with its source document.
  4. 04Future answers use itVersioned, never deleted, counted every time it’s applied.

02

It lands in the inbox

Flags arrive sorted by type (incorrect, outdated, incomplete or a safety concern), each in the engineer’s own words.

Feedback inbox: pending, in review, refined and median time to refine, then a table of flagged answers with type, category, site, submitter, age and status.

03

An expert writes the fix

The answer, the complaint and the fix sit side by side. Here, Copilot quoted 18 N·m from a superseded revision; the expert stores the right value, scoped to one site.

A feedback record: on the left the expert writes the refinement, chooses whether it applies to one site or a tool family, and names the source document; on the right the original thread and what the engineer said was wrong.

04

Future answers use it

Every fix traces back to the feedback that created it, and counts each time it’s applied.

Stored refinements: each with the feedback it came from, where it applies, its owner, how many times it has been applied and its status.

Is the loop working?

Insights track feedback volume, refinement coverage and repeat feedback. Repeat feedback is the number to watch.

Feedback insights: feedback volume, refinement coverage, answers using refinements and repeat feedback rate, with weekly volume and breakdowns by category and type.

Admin console

Four questions an admin asks every week

Is it being used? Are the answers good? Where is the knowledge failing? Who can see what? The console answers them in that order.

The overview

Activity by day, then the five numbers that matter: adoption, conversations, resolution rate, feedback coverage and response time.

CE CoPilot admin overview: an activity heatmap by day, adoption 23.6%, 26,022 conversations, 64% resolution rate, 68% feedback coverage and 6.9s average response time.

Behind each number

Usage, the conversations themselves, and the documents CE CoPilot is allowed to read.

Messages and people on one daily trend.

Usage page: conversations, messages, resolution rate and adoption, with a daily trend of messages and people.

What I learned

Trust is something you design

Show your work: engineers don’t trust an answer because it sounds sure. Showing the source and an honest confidence score let them check it for themselves.

A no can still help: declining an off-topic question is easy. Telling the engineer exactly what to add turns a dead end into a better question.

Plan for being wrong: some answers will be wrong. Designing the way back, from a thumbs-down to a scoped fix, mattered more than any single screen.