Business Process
A document describing how the AI stack is run within an organization. AI isn't something that "just works once installed" — it only becomes useful when people feed it material and judge its output.
If the tutorial is about "getting your first answer from a single document," this document is about "keeping it useful over time."
Roles
| Role | Account | What they do |
|---|---|---|
| AI Operator | Admin console account (env) | Service status, document indexing, pipeline management |
| Field User | Platform account system | Ask questions in Ask, judge answers, approve/act on/record proposals |
| System Administrator | Server access | Installation, key rotation, backups |
The console account is one shared by the team → Changing the initial password.
Field user tasks — what's visible and approval authority
Copilot assigns each field user a job (operator, shift lead, maintenance, quality, plant manager). Everyone starts on the Ask home screen, and the job determines what's visible in the sidebar, which sources appear in proposals, and approval authority (based on the default bundle). The sidebar shows, from top to bottom, New Conversation · Proposals · Automation.
| Job | Visible | Additional tasks |
|---|---|---|
| Operator | New Conversation (Ask) · Proposals (plans, incidents) · Automation | Record results in proposals |
| Maintenance | + Anomalies to confirm in Proposals | Execute actions, approve maintenance request drafts, publish video work procedures, forward automation results to others |
| Shift Lead | + Anomalies to confirm | Execute actions, approve maintenance request drafts, publish video work procedures, forward automation results to others |
| Quality | Everything | — |
| Plant Manager | Everything | Execute actions, approve maintenance requests/order follow-up drafts, publish video work procedures, forward automation results to others |
If not set, the default is Operator. During rollout, if you at least assign jobs for shift lead, maintenance, and plant manager, each person's own work will show up in «Today's Tasks» on the home screen. Only administrators can change jobs and scope (Admin Console › Users › My Job) — because approval authority follows the job. Users themselves can only view this in Settings → Copilot Screen Map — Menus by Role · User Permissions and Job Roles · Access Control
Overall Flow
Where arrows cross lanes is the handoff point.
현장 피드백 → 문서 보강 → 재색인 If this loop doesn't run, the AI stays stuck at the level of the documents it was
first given.
Decide where feedback goes — this is usually the missing piece that breaks the loop.
1. Rollout — Getting to Usable
| Step | Task | Owner |
|---|---|---|
| 1 | Install → One-line install | System Administrator |
| 2 | Set up console account/keys → Changing the initial password | System Administrator |
| 3 | Check overall status is green on Home → Admin Console | AI Operator |
| 4 | Choose which documents to feed in | AI Operator + Field |
| 5 | Index documents + verify with Query Test | AI Operator |
| 6 | Release to the field | — |
Dumping in a lot of random documents doesn't help. Start with what the field actually asks most often — equipment manuals, work standards, then troubleshooting history usually have the biggest impact. Feeding in outdated documents makes the chatbot confidently give wrong answers.
2. Operation — Recurring Tasks
| Frequency | Task | Owner |
|---|---|---|
| Daily | Home status summary (confirm 0 outages) | AI Operator |
| Weekly | Pipeline status/run history | AI Operator |
| Weekly | GPU utilization/XID, LLM response speed (P95) | AI Operator |
| Monthly | Reindex updated documents, review cases of wrong answers | AI Operator + Field |
| Quarterly | Key rotation → Changing passwords/API keys | System Administrator |
All viewed from the Admin Console.
This indicates a GPU hardware/driver error. Even if it's still running now, the following symptom usually means the node is going down.
A Day in the Field — Ask, Proposals, Automation
A field user's day passes through three things in this order.
| Step | Where | What happens |
|---|---|---|
| 1 | «Today's Tasks» on the Ask home screen | Initially shown collapsed into one line ("N tasks today"). Tapping it expands up to five lines in order of urgency (expanded/collapsed state is remembered per device); tapping a line opens that proposal. View All goes to Proposals. If there's nothing to do, this line isn't rendered at all |
| 2 | Proposals › To-do | Read the AI's observation, diagnosis, evidence, and proposed action, then approve. Tapping an evidence line opens the source document, drawing, or video in the right panel; Back returns to the detail view. Plans show «Approve and Start», anomalies to confirm show «Escalate to Incident», and incidents show «Approve Action Plan». For false alarms, choose a reason and dismiss |
| 3 | Proposals › In Progress | Execute the action and leave a result record (marked complete/not applicable plus a one-line note) |
| 4 | Same detail view | Verify no recurrence — watch for the same issue over a set period. If it recurs, it returns to «Recurred»; incidents are closed via «Confirm Verification and Close» |
| 5 | Proposals › Completed | View the history of actions taken and a one-line summary (number of actions, processing time, verification passed, recurrences) |
| 6 | Automation | Tasks that were checked daily/weekly, or that need to be flagged when a condition is met (formerly «Watch»), can now be delegated in a single sentence. Results come back as proposal cards/reports and feed into step 2, and are also delivered via app notifications, company email, or messenger channels |
- If something is unclear while working, use Continue Asking in the detail view to ask with that task as context.
- Tasks that affect other systems (maintenance requests, order follow-up) are only sent out as drafts approved by a person. Some actions require two-person approval.
- Step-by-step buttons and badges are in Proposals; how to build automation is in Automation.
Decide who opens «Today's Tasks» and when during the day (e.g., shift start), and who substitutes for someone with approval authority when they're unavailable → Access Control — Two-person approval/delegation
3. When an Answer Is Wrong — Where to Fix It
"The chatbot gives weird answers" can have causes at several layers. Narrow it down in the following order.
| Check | If so |
|---|---|
| 1. Is the source document indexed? | If not → add the document (Documents) |
| 2. Does search find it? | If not → try a different search method (hybrid/local/…) (Query Test) |
| 3. It's found, but the answer is wrong? | The source document may be outdated — fix the document |
| 4. Answer is slow or cuts off? | LLM/GPU issue → LLM · GPU screens |
Most of what gets reported in the field as "the AI is dumb" turns out to be missing or outdated documents.
4. Recording Whether Anomaly Judgments "Were Correct"
Anomaly detection has the same loop as document answers. When field users record whether an AI judgment was a false alarm, or whether a proposal actually helped, that feedback is used to measure per-equipment accuracy, and the AI adjusts its own judgment criteria per equipment. Since this is purely a record, it doesn't change the equipment or the order.
| Who | Where | What |
|---|---|---|
| Field user | Dismiss on anomalies to confirm in Proposals | Choosing a reason (setup/changeover in progress, equipment stopped/idle, under maintenance, sensor issue, other) marks it as a false alarm and feeds into AI accuracy |
| Field user | Accept · Reject on AI proposal cards in Proposals | Reason for rejection; after acceptance, mark effective/not effective/not yet known |
| Administrator | Admin Console › Accuracy Tuning | Review per-equipment threshold self-adjustments and promote or revert them |
The old «Awaiting Verification» screen (which marked shadow judgments as correct/wrong) no longer exists — findings recorded in shadow mode are now viewed by administrators under Admin Console › Field Notification Preview. The old «Report Missed Anomaly» and the «Reports › AI Accuracy» tab are also gone. Misses are now automatically flagged by the server against equipment downtime/failure records, and accuracy figures can be viewed at AI Findings and Accuracy — Where to View Accuracy.
→ AI Findings and Accuracy · Copilot Admin Console — Accuracy Tuning
Common Points of Friction
| Symptom | Usual cause |
|---|---|
| "We rolled it out but nobody uses it" | What the field asks doesn't match the documents fed in |
| "The analytics screen is empty" | The feature pipeline has stopped — check the pipeline, not the model |
| "Today's Tasks isn't showing" | By default it's a collapsed one-liner ("N tasks today") that must be tapped to expand. If the line itself is absent, there's nothing to do — that's normal. If it's always absent, check job/equipment scope |
| "It confidently gives wrong answers" | Outdated documents remain in the index |
| "It's gotten slower" | GPU contention or LLM queuing — check P95 |
Related Documents
- Tutorial — Getting the Chatbot to Answer from Our Documents
- AI Admin Console — the screen you check daily
- Copilot Screen Map · Proposals · Automation
- Anomaly Detection · Incident Response
- Database Model — what's stored where