Skip to content
plugboard.appRequest a demo

Running the desk

AI settings

Settings, AI (/admin/ai), gated by connector.manage. Where a school decides what Ask and the background AI features are allowed to do, and how many background AI calls may run in a day. AI activity in the header opens the record of every call for anyone who holds audit.view. Every change is recorded in the audit log.

The page runs top to bottom: Where it runs, Features, Models, Usage and limits, Drafting style and School tags. Switches apply at once. Typed fields (the daily limit, the drafting style and a tag being edited) are saved together with Save changes at the foot of the page, or put back with Discard. Leaving with unsaved changes asks first.

AI for this school, the first row of Where it runs, has three settings:

SettingWhat it does
OffNo model is called for this school, for anything. Turning it off asks you to confirm first
On request onlyStaff can ask the AI when they press a button. Background features wait until you choose On
OnStaff can ask the AI, and background features run in the modes set under Features

Left unset, a school on managed AI is On and any other school is On request only. Managed AI has no On request only setting. The setting is stored as ai.enabled. Under the switch, AI diagnosis says, a sentence each, whether AI is on, whether a healthy connection answers chat and search embeddings, and why recent requests got no answer.

With no AI connector set up, or none turned on, the page says so and links to Connectors. Nothing is sent anywhere, and the desk works without AI.

Below the switch:

RowWhat it says
ProviderManaged by Plugboard: Anthropic’s Claude models, served by Cloudflare, or the names of your own AI connectors that are on, with a link to Connectors
Where text is processedFor managed AI, the United States, through Cloudflare. For your own provider, wherever it processes text under your agreement with it
Names and IDsWhether names and IDs are masked before sending, and the switch for your own chat connector (see pseudonymisation)

What we process names the sub-processors for managed AI.

On by default: names on the school’s roster, email addresses, phone numbers, IDs, device serials and street addresses are replaced with tokens before a call leaves and restored in the answer. It is always on for managed AI. For a school’s own chat connector, an administrator can switch it off on the Names and IDs row, which needs a step-up confirmation. Typed decisions always mask, whatever that switch says. What masking misses folds out the limits: a description that identifies a student without naming them, or a name that is not on your roster.

Features counts in its heading how many are on, for example “8 of 32 on”, and lists every AI feature in seven groups:

GroupFeatures
Sort and route new workTriage, School tags, Incoming message review, Ticket routing, Fill in new tickets and repairs, Flow decisions
Write first drafts for staffTicket brief, Draft replies, Wrap-up note, Knowledge drafts, Polish wording, Notice translations, Flow writing, Morning briefing
Keep tickets tidyMerge duplicates, Link similar tickets, Propose incidents, Ticket tidy-up
Look out for peopleWelfare concern review, Requester tone
Answer questionsAsk panel commands, Ask about this record, Portal article suggestions, Describe a widget
Prepare changes for approvalJoiner and leaver plans, Access requests, Routine requests, Device fixes, Likely repair outcome, Repair cover
Set up the deskDesk setup plan, Set up from a document, Flow drafting

A few of them, in the words the page uses:

  • Ask panel commands recognises a supported command typed into Search or ask. Complete direct commands stay instant, and permission checks and action reviews always apply.
  • Ask about this record answers staff questions about the record on screen.
  • Morning briefing fills Needs you and Spotted on Home each morning.
  • Describe a widget builds dashboard and report widgets from a description, using only data the person can see. Nothing is added until the person accepts it.
  • Requester tone and Welfare concern review never run on a ticket from a student, a parent or an unknown sender. Safeguarding words on a student’s ticket are always flagged to staff. See sentiment & staff experience.
  • Repair cover is rules with no model, so it shows Rules, always on and has no switch.

Each row has its name and a line saying what it does, its state in a word or two (“Suggest”, or “Shadow · until” a date while it is learning), and a switch. The switch turns a feature on in Shadow, or off. Its settings sit behind the fold beside the switch; a feature with a minimum answer probability set opens on arrival.

The fold’s Mode control offers every mode:

ModeWhat it does
OffNever runs. Where a feature also has rules, its note says what Off does to them
ShadowWorks quietly and is scored against what people do. Nobody sees it
SuggestShown as a suggestion. A person accepts, edits or dismisses it
AutoApplied on its own, labelled as AI, and can be undone

A feature with no mode set may be learning in Shadow first; its fold says until when, and the mode it moves to then. The fold also shows how many suggestions have been decided and how accurate they were, and folds out Results by model: each model earns automatic actions separately, and changing provider or model version starts a new approval history.

Auto unlocks only after a proven accuracy record for that feature at that school: 50 decided suggestions at 90% accuracy within 60 days, said once above the list. Some features are never automatic, whatever their record: Welfare concern review, Propose incidents, Device fixes, Flow writing, Flow drafting, Describe a widget, Set up from a document and Desk setup plan. Identity work never runs on Auto either: people flows for joiners, movers and leavers always stop at a person’s go-ahead before any account changes.

With a decision provider on, a feature that uses it takes a Minimum answer probability: answers below it stay in shadow. Chat answers have no calibrated probability, and automatic actions still need the model’s review history.

Some features name the flow that runs them, with its state and Open flow. Those features start through a system flow: the flow decides when the feature fires, and the mode set here still decides what it does. While the flow is not on, nothing starts the feature, whatever its mode.

Three features cover AI inside flows:

FeatureKeyWhat it coversHighest mode
Flow decisionsflow.decideConditions in words and Decide stepsAuto, earned per question
Flow writingflow.writeSummaries and reply drafts written by a flowSuggest
Flow draftingflow.buildDescribe a flow in the flow editor, and Describe it on New flowSuggest

Flow decisions caps every question in the school’s question library: at Shadow, no question goes beyond watching; at Suggest, a person confirms every answer; at Auto, a question acts alone only once it has earned it with the model that answers. Until it is set, questions only watch. Flow writing and flow drafting never act on their own, because a person always uses what they write. See decision models.

One set of rows per chat connector: Writing (the fast model, and the capable model for longer work when it differs) and Similar-ticket search (the embedding model), each with the connector’s state. Managed AI writes with Claude Haiku 5.5 and searches with BGE-M3, through Cloudflare. On your own connector, Models for that connector sets the Fast model, Capable model and optional Embedding model names; blank uses the connector default. A line under similar-ticket search says how many past tickets are indexed so far.

Decisions lists the decision providers in the order they answer (see decision providers). Add a provider in Connectors.

AI usage is shown as a number of calls. Calls staff ask for count towards the allowance; background work is included, up to 8 writing calls and 80 decisions per ticket a month. The section shows:

  • This month: AI calls a person asked for on managed AI, against the calls your plan includes, for example “1,240 of 5,000 AI calls used this month”, with how many are left and the date the count starts again (the first of next month, UTC). Under it, background calls on managed AI this month, which are included, and calls on your own AI account or model, which your AI provider bills (a local model bills nothing). Neither counts towards the allowance.
  • Background today, against the daily limit when one is set.
  • Background calls a day, the daily limit.
  • What it saved, once the AI has done something this month: suggestions kept and undone, as shares of the recorded outcomes, and an estimate of the time saved, with Suggestion breakdown by feature folded under it.

Going over the allowance never stops AI. The page says by how many calls you are over. Money appears only when your licence prices overage: the overage this month, billed in whole blocks of 1,000 calls at the licence’s price per block. See plans.

Background calls a day (ai.dailyBackgroundCalls) is a whole number from 0 to 100,000:

  • Leave it empty for no daily limit. This is the default.
  • 0 stops all background AI.
  • Any other number counts background writing and decisions that a model answered, and resets at midnight UTC. Search indexing is not counted; only a limit of 0 stops it.

Calls a person asks for never meet the daily limit.

Plugboard also keeps its own daily safety limit on each school’s background AI. You cannot see or set it. If it is reached, the section says “Background AI has paused until midnight UTC because Plugboard’s daily safety limit was reached”. Writing that staff ask for still runs.

The allowance also shows on Settings, Licence & billing.

A Style note shapes drafts a person edits and sends, for example “Warm and brief. Australian spelling. Sign off as ICT Service Desk.” Families do not see it.

Optional decision providers handle fixed classifications separately from the chat provider used for drafts. Decision requests always mask identifying data, and each model earns its own approval history. Laya ships without a trained Plugboard checkpoint; Jev starts disabled. Review suggestions on your own data before considering automatic use.

The order of Decisions, under Models on this page, sets which enabled provider answers first and which answers only when the ones before it cannot. Every typed decision follows it, including those in flows; a flow never picks its own provider. With no decision provider on, decisions use the chat model, which gives no confidence, so its answers stay suggestions until that model earns its own record. See decision models.

Optional decision providers can recognise a supported command typed into Search or ask before conversational reasoning. Exact commands keep their fast path. Permissions, record access and action reviews still apply; a classification never supplies an arbitrary tool or target. Each model starts in shadow. Shown interpretations can be corrected without running another command.

Decision providers can match an existing help article or school service while someone describes a problem. Choosing a service fills its form; the person still completes and submits it. Existing selections are retained. Portal feedback is linked to the person’s own suggestion and fails safely if content changes.

Access and routine request matching have separate model records. Identity requests retain staff approvals and step-up. Model-based routine fulfilment needs both the school’s pre-approval and the answering model’s earned auto record; switching providers cannot inherit the old model’s accuracy.

In Settings, Statuses & fields, Sorting criteria, describe when each category, subcategory, priority, department, repair type or catalogue item fits. Use up to 200 characters and five example phrases of up to 120 characters each. Try it runs one metered sample decision against what is typed, saved or not, and changes nothing. The criteria follow your selected decision provider. Priorities still follow the impact/urgency matrix. The tab is there whether or not AI is on.

School tags, at the end of this page, holds up to ten yes/no questions, asked alongside triage. Each starts in shadow. Suggestions offer Add label and Not relevant; applied tags are normal labels staff can remove. Every tag and actual model needs its own 50 human decisions at 90% accuracy over 60 days before auto can apply it. Changing the question or description starts a fresh record. Manually changing one label does not label all the other AI questions as false.

Triage with all ten tags on a local decision provider is background work. The Laya connector waits up to 75 seconds for an answer by default (timeoutMs). Ask panel commands and incoming message review keep their 2.5 second limits.

Triage can use up to ten resolved tickets that staff classified in the same school, campus and department. The review links each supporting record. If history disagrees with the decision model, staff must review the choice. History-only suggestions always require a person. Support from history is shown separately from the model’s answer probability. Changed or reopened source tickets invalidate the review. An embedder is required; no training or decision provider is required for this history path.