Skip to content
plugboard.appRequest a demo

Running the desk

Ask and AI features

One connector supplies the language model for everything on this page. The connector you choose decides where your ticket text goes. Ask also works with no model at all, from the desk’s own records.

  • Search or ask, at the top of the menu, is the one way to ask the desk a question. It opens the Ask panel. See asking a question.
  • Suggested values on a ticket, inside the Type, Category, Priority and Assignee fields they would change.
  • Summary, one paragraph at the top of a ticket’s Activity.
  • A suggested reply waiting in the reply box, and Draft to write one.
  • Describe a filter on Tickets, Repairs, Devices and Loans.
  • Describe a widget when you customise Home or a report board.
  • Describe a flow in the flow editor, and Describe it on New flow. See building a flow.
  • The MCP server, the same tool catalogue, for an AI client you connect.

AI output says in words what wrote it: Suggested, Draft, Summary, Read by the AI model, Set automatically (with Undo) and Proposed by: Plugboard. Point at the words to see why. Replies in the Ask panel are attributed to Plugboard.

Settings, AI and Settings, AI activity are where a school decides what each feature may do and reviews every model call. See AI settings and AI activity.

When the school has switched AI off, or no model is connected, nothing that only AI can do is drawn: no Describe a filter, no Describe a widget, no Draft or Improve in the reply box, no Refresh on a ticket’s summary, no Describe a flow. They are left out, not greyed.

What runs on rules stays: Search or ask, exact answers from the desk’s records, related records, and fields filled in from rules. A question in the Ask panel that needs a model gets a one-line answer saying so. Somebody who can manage connectors also sees the reason and a link to the settings page that fixes it.

Where it runsWho paysWhere ticket text goes
A local modelYour configured Ollama serverYour school supplies the hardwareTo that server
Your own keyAnthropic, or any OpenAI-compatible endpointYou, to themTo that provider, or to your own GPU box
Vendor-hostedOur endpointUs, within your plan’s allowancePseudonymised, to Cloudflare: embeddings (BGE-M3) and typed decisions (Cloudflare Clef) on Workers AI with no region pinning, and chat on Anthropic’s Claude Haiku 5.5, which Cloudflare relays to Anthropic in the United States

A local model is available through the container deployment’s Ollama option. Native and source installations need a separately configured Ollama server and model. Local inference sends prompts to that configured server; other connectors can still make outbound requests. CPU-only operation is possible, with response time depending on the model and hardware.

Your own key is faster and billed to you. It also decides where student and staff ticket text is allowed to go, so each option is a separate connector.

Vendor-hosted is managed hosting only. Each school’s managed deployment has its own server. Hosted inference avoids reserving that server’s memory for a local model. Eligible provisioning can configure vendor inference before the first sign-in. See plans for the allowance and going over.

Self-hosted deployments never receive vendor credentials.

Ask appears for any role that can read part of the desk: it needs at least one of client.view, device.view, loan.view, repair.view, ticket.view, monitor.view, operations.view, stock.view, report.view, connector.manage, branding.manage or backup.view. Owner, Technician and Read-only all qualify. Asking grants nothing: Ask reads and acts only with your own permissions. A custom role with none of these does not see it. See permissions.

  1. Press Search or ask at the top of the menu, press Ctrl K (Cmd K on a Mac), or press / anywhere outside a text field.
  2. Type your question. Words that read as a question, ending in ”?” or starting with what, which, who, how, when, why, is, are, do or does, put Ask: ”…” first, above the pages and records that match. Anything else offers it last, after them.
  3. Choose Ask. The question is asked once, when you choose it, never while you type.

An exact answer from the desk’s own records, a count or a list, shows in the search panel itself. A count leads with its number. A list reads like the list it came from, five rows at a time with Show all. When the words were read as filters, a line says Read as with each one, and names any word that was not applied. From desk records closes the answer, and Ask a follow-up carries it into the Ask panel.

Anything else opens the Ask panel with the answer already in it.

Asking from a person’s, ticket’s, repair’s or device’s page sends that record with the question, so “who raised this” needs no number. The panel says About and the record’s name under the question box. Ask re-reads the record with your permissions and ignores one you cannot open.

See search for everything else the field finds.

The panel opens on the right over the dimmed page.

  • Answers are attributed to Plugboard. A count leads with its number, and records come back as rows you can open. From … links name the articles and records an answer used, and its last line says From desk records or Read by the AI model. Check a reading by the AI model against the records it links.
  • Follow-ups under an answer suggest a next question.
  • Suggestions. With nothing asked yet, the panel offers questions On this page and Across the desk, three of each, with More questions. Recent lists your last few conversations above them.
  • History lists Changes to review first, while a conversation has prepared a change that waits for your approval, then your saved conversations by their opening question.
  • New conversation starts again.
  • Escape, Close, or a press on the dimmed page closes the panel, and focus returns to where you were.

One question is capped at 2,000 characters, as is each earlier turn the conversation carries forward. Enter asks; Shift Enter starts a new line. Everything you send is read again by the model on each step it takes, so a pasted log file costs several times what it looks like.

Ask remembers who you are talking about within a conversation, so “issue them a loan” works after looking somebody up. It stays on the service desk: trivia, arithmetic and general knowledge are not its to answer.

  • How many laptops are on loan?
  • Who has laptop loan #1?
  • Available iPads
  • Overdue repairs
  • Unassigned repairs
  • Show me open tickets
  • Unassigned tickets
  • How much did we spend this year?
  • Purchases this week
  • Summarise ticket 412
  • Search knowledge for Wi-Fi setup
  • Issue a loan to Riley Brooks
  • Create a damaged display repair for Riley Brooks with a loan

Exact wording is not required with a model connected. Without one, a built-in command engine answers the recognised phrasings above from the desk’s records, offline and free.

Tickets and repairs are two queues, and Ask keeps them apart. Tickets means the service desk: access requests, accounts, software, network. Repairs means work on a device. On a school without the Tickets module the word still means repairs, because there is no other queue to read.

Before anything that changes data, Ask shows the change for you to confirm. A change prepared in a conversation waits as a review: View proposed change opens it, and it waits under Changes to review in History, and on Home’s Needs you, until you decide. Reviews expire and cannot be reused for a different action; ask again if the record or your permissions have changed.

Actions respect your permissions, and every one is audited. The model chooses among commands; the permission checks run regardless, so there is no path through Ask to something your role forbids.

Describe a filter turns a sentence into the list’s own filters. It is first in the Filter menu on Tickets, Repairs and Loans, and beside the search on Devices. Type something like unassigned repairs ready for collection and press Apply. The example in the box is one that list can read.

The list reads its own words first, and the everyday ways of saying them: assigned to me and my are Mine, with no owner and assigned to nobody are Unassigned, and Repairs reads ready for collection and with the vendor. Small words such as show me, the and please are passed over. Most sentences people type apply at once this way, with no call to the model.

Anything else goes to the school’s AI, which reads it for this list’s filters only: it runs no other tool and prepares no action. While it reads, the box says Reading it with the school’s AI. This takes up to 8 seconds. After 8 seconds it stops and sets no filter.

The result is the list’s ordinary filter chips, removed the ordinary way. Nothing else on the page changes, and the list never shows more than you could already see. When no filter on the list matches, or the reading took too long, it says so and names the words the list does read, such as Try unassigned, overdue, mine, a status or a priority.

Describe a filter needs the same permissions as Ask and is drawn only while the school’s AI can answer. A question asked in the Ask panel that names the filters of the list you are on still applies them, with or without a model.

New tickets are sorted in two passes, in this order.

Rules first. Deterministic, and they run with no model configured. You can edit them.

Then the model, where one is connected, asked to improve on the rules rather than to start from nothing. Its answer is accepted only where it picks from lists your school uses. Anything else is dropped, so an invented category never reaches your reports.

What it proposes appears inside the field it would change, under Type, Category, Priority and Assignee: Suggested: and the value, with a tick to accept it and a cross to dismiss it. Point at it to see why. Where a school lets a feature apply its answer on its own, the field reads Set automatically with Undo.

Everything it decides is overwritable. The moment somebody edits one of those fields the ticket is marked as theirs and is never re-sorted.

A plain paragraph at the top of the ticket’s Activity says what is wrong and where it stands. Its footnote reads Summary · From N messages. While the model answers, Refresh prepares it again. Suggested steps, a likely cause and similar tickets fold away under it.

Nothing is drawn until a summary exists.

A reply prepared for the ticket waits greyed in an empty reply box, marked Suggested reply · Tab to use. Tab or Use draft takes it, typing ignores it, and Dismiss hides it. Draft in the reply box writes one on demand, and Improve rewrites a reply you have typed. Both are there only while the model answers.

Nothing reaches the person who raised the ticket until somebody sends it. See the composer.

A local model costs nothing beyond the hardware.

Your own key is billed by your provider. The work is small: a triage or a summary is a few thousand tokens in and a few hundred out. A desk triaging around three hundred tickets a month lands between five hundred and fifteen hundred calls.

Vendor-hosted is included up to your plan’s monthly allowance of AI calls. Only calls a person asks for count; background AI is included. Over the allowance AI keeps working, and overage is billed in blocks of 1,000 calls at the price on your licence. Plugboard never bills calls on your own key or a local model.

AI settings shows AI calls used this month and sets the daily limit on background AI.

Background work, such as model triage, runs only when the school switch allows it: by default only with vendor-hosted (managed) AI. A school on a local model or its own key turns it on there. With more than one model connected, a connector that names a model for the job is tried first, and the next one answers if it fails.