Intelligence
AI triage
AI triage reads a ticket together with what Frontline PSA knows about the customer, then proposes an assessment, a plan and a list of suggested actions. Every claim points at the evidence it came from. Triage never changes anything by itself: each suggestion waits for you to click Apply.
Before you start
- Triage is available to administrators, service managers, dispatchers and technicians.
- It needs an AI gateway, which your administrator configures. See Deployment.
- If no gateway is configured, a trial installation runs triage on a stand-in and the panel shows a MOCK MODEL tag. The stand-in matches keywords so you can try the workflow. It is not a model's judgement, so do not act on its advice.
- In production with no gateway, the panel says that no AI model is configured and Run triage is disabled.
Run triage
- Open a ticket. See Tickets.
- Find the AI triage panel.
- Click Run triage.
The panel shows each step as it happens. To stop the run, click Cancel.
When the run ends, the tag in the panel header shows the result:
| Tag | Meaning |
|---|---|
| Queued, Running | The run is in progress. |
| Ready | The run finished and every source was checked. |
| Ready · gaps | The run finished, but something could not be checked. Read What could not be checked or needs care. |
| Failed | The run stopped. The panel shows the reason and how far it got. Nothing is guessed or filled in. |
| Cancelled | You stopped the run. |
If the ticket changes after a run started, the panel warns that new activity is not included. Click Re-run to triage again. Earlier runs stay available in the list in the panel header, and suggestions from an older run that you did not act on are marked superseded.
What triage reads
Triage gathers the following and gives each item an evidence number (E1, E2, and so on):
- The ticket and its most recent notes.
- The contact's and the company's ticket history.
- The company's alert, contract and SLA.
- Resolved tickets that look similar, with how they were resolved.
- The devices linked to the ticket or the contact.
- Published knowledge base articles. See Knowledge base.
- Customer memory for the company and contact. See Customer memory (DPMem).
Read the result
Summary and classification
The top of the result has a short summary and four facts:
- Classification: the issue type triage would give the ticket. If it differs from the current one, the old value is struck through.
- Priority: the same, for priority.
- Caller: how the customer comes across in the conversation.
- Confidence: the model's own confidence in its diagnosis and plan.
Below these, Frontline PSA lists the results of its own consistency checks on the answer. It warns you when:
- Confidence is low (under 45%).
- The model says the ticket can be resolved now and also says it needs escalation.
- The company has an alert that the plan does not take into account.
- No knowledge base article, memory or similar ticket was found, so the plan is general.
Findings
Findings are what triage concluded from the evidence. Each one ends with the evidence numbers it relies on. A finding with no evidence is tagged uncited: treat it as the model's opinion.
Plan
Plan is a numbered list of steps. A tag tells you whether triage thinks the front line can resolve the ticket or it needs more than front line, with a time estimate where it has one. The plan can also include:
- Say to the customer: wording you can use on the call.
- An Escalate banner naming the queue and the reason.
- Risks to be careful about.
- Ask the customer: information that is missing.
Citations and evidence
Click any citation, such as E3, to jump to that item under Evidence. Each evidence item shows where it came from, when, and the excerpt the model was shown. Where the source is a record, click open to go to it.
If the model cites an evidence number that does not exist, Frontline PSA removes the citation and reports it, so every citation you see points at something real.
Apply, edit or dismiss suggestions
Suggestions are actions triage proposes. Each has a kind, a risk level and a short reason.
| Kind | What applying it does |
|---|---|
| Update ticket | Changes the listed fields, such as issue type or priority. You see the before and after values first. |
| Escalate | Moves the ticket to the suggested queue and adds the reason as a note. |
| Internal note | Adds an internal note to the ticket. You can edit the text first. |
| Customer reply | Adds a customer-visible reply to the ticket. You can edit the text first. |
| Tenant action | Runs an action on the customer's tenant. A read runs at once and the result is noted on the ticket. A change is sent for a second person's approval. See Customer tenants. |
| Scout diagnostic | Starts a Scout session on the named device. You can edit the objective first. See Scout. |
| Save to memory | Saves a note to customer memory. You can edit what is remembered. |
| Link KB article | Links a knowledge base article to the ticket. |
For each suggestion:
- Read the reason and check its citations.
- Edit the text or fields if you want to change them. The button changes to Apply with my edits.
- Click Apply, or click Dismiss…, type why not, and click Dismiss.
Applying a suggestion does the work as you. It is recorded on the ticket with your name and in the audit log, and the suggestion shows what happened with a link to the result, for example the note or the Scout session. If applying fails, the suggestion is marked failed and says why.
Some suggestions are blocked, for example a tenant action you do not have permission to run. The suggestion shows the reason next to a lock icon.
Note: Dismissal reasons are kept. Recent ones are listed on Intelligence → Overview, which helps show where triage is not useful.
Limit: Applying a Customer reply records the reply on the ticket. Frontline PSA does not email it to the customer. See Tickets.
Check what the run did
Two sections in the panel let you inspect a run.
- Run trace lists every step with its duration, grouped into Gather, DPMem, Reason, Validate and Suggest. Click a step to see what it did and which evidence it produced. Steps that need attention are counted in the heading.
- What the model was sent shows the size of each part of the request. Text written by customers and text from memory is marked as untrusted data, and the model is told never to follow instructions inside it.
Review runs across the desk
Open Intelligence to look at triage across all tickets.
- Overview shows the last 30 days: the number of runs, the share of suggestions applied, average run time, outcomes by suggestion kind, why runs failed, and recent dismissal reasons.
- Triage runs lists every run with its ticket, status, confidence, the number of suggestions applied, dismissed and open, the model, who started it and how long it took. Filter by status, or click a row to open the ticket.
The side panel in Intelligence also shows whether AI, DPMem and the Scout agent are live or running as stand-ins.