Main site

Education AI agent · Safety monitoring

Student Safety-Monitoring AI Agent

Match text on school-managed accounts against the district's approved terms, assemble the alert around the words that triggered it, and route it to the on-call counsellor, who assesses risk.

Typical delivery
4–6 weeks
Deployment
Your stack
Counsellor read
Alert only
After launch
Agent Care

What this agent does

Delivers the alert, not the judgement

In

  1. 01

    Ingest text from school-managed accounts and devices inside the monitoring window the district configured.

  2. 02

    Normalise the source surface — document, message, search or draft — and carry the timestamp forward.

Reason

  1. 03

    Match against the district-approved term list, and against nothing outside it.

  2. 04

    Apply the scope, hours and exclusions configured for the district.

  3. 05

    Attach the matched words with the text around them, so the trigger can be read rather than inferred.

Decide

  1. 06

    Label the context class — coursework, creative writing, journalism, message — or mark it unresolved.

  2. 07

    Route the alert to the named on-call member of the counselling or crisis team.

Out

  1. 08

    Retain the trigger, the routing, the acknowledgement and the assessment against the alert.

  2. 09

    Execute write actions only inside the approval boundaries agreed during implementation.

Product statement

The agent delivers an alert; a counsellor, psychologist or crisis clinician assesses risk, and the district stays the decision-maker.

Example workflow

One alert, trigger to assessment

  1. 1Text matched in scopeSchool document, school mail, managed search or classroom message
  2. 2Signals assembledThe matched words, the text around them, the surface and the time, each with its named source
  3. 3Alert packet draftedMatched text, context class, source surface and confidence
  4. 4Suppression checks runScope and hours checks, excluded-term checks, context labelling and confidence threshold

No human action required

The first four stages run unaided, and nobody is contacted at any of them — the agent is assembling a packet, and the counsellor's lane opens at the confidence gate.

5DecisionSplits at the triage threshold
Low concern

Goes to the on-call counsellor to read.

Elevated concern

Adds a crisis-team triage read first.

Counsellor assessment

The alert is held with its matched text, its context class and the time it fired.

Acknowledge · Close · Refer to the crisis team

Assessed — recorded against the alert

  1. 6Case record updatedOnly where write access and routing policy allow it
  2. 7Alert evaluatedClosure rate, cohort alert rates, context-class accuracy and time to acknowledge

Downgrades

Each alert a counsellor closes is counted in the evaluation.

What should not run autonomously

Human approval stays in control

Outside the boundary — human approval required

8 items

  • Contacting police, a resource officer or emergency services.
    • Notifying a parent or guardian.
    • Deciding that a welfare check should happen.
    • Assessing risk, or deciding a threat is credible.

  • Automation boundary

    Agent acts unaided

    • ✓Match text against the district-approved term list for the named owner.
    • ✓Assemble the packet.
    • ✓Label the context class, or mark it unresolved.
    • ✓Deliver to the named on-call recipient and log the acknowledgement.

    Any write happens inside the boundaries agreed at implementation, never ahead of assessment.

    • Disciplining, excluding or removing a student from class.
    • Disclosing under the health-and-safety-emergency exception.
    • Extending the scan to personal accounts, devices or off-hours.
  • Adding, removing or reweighting a monitored term.

Example output

One alert, annotated

Everything the agent sends is attached to the text it was drawn from.

Alert output · single trigger

Illustrative example
Alert
Term-list match
Matched text
The narrator says he wants all of it to stop
Fired at
09:14 local
Source surface
Shared class document
Confidence
76%
Context class
Creative-writing draft

As received

Taken from the school document and the district's term list — nothing on this side is decided by the agent.

Context supplied

  • The paragraph around it
  • The assignment it sits in
  • Surface, device and time

Why this context class

The line sits inside a submitted assignment, not a message to anyone.

What the score decides

Below the configured threshold the alert picks up a crisis-team triage read before.

Action

AcknowledgeCloseRefer to the crisis team

Value

Where AI adds value

The same four claims, placed at the point in the workflow where each one applies.

Where the value lands

Value 01 – 04

Every alert

From school-managed accounts

03Matching

Match the district's list

Work from the approved term list and the scope, hours and exclusions the district configured.

01Approved path

Route it to a person fast

The packet reaches an on-call counsellor without anyone watching a queue.

02Human review

Give the counsellor the words

The matched text and its context class travel with the alert, so the read starts at the sentence.

04

Build an evidence trail

Every alert keeps its trigger, its routing and the person who assessed it.

Integrations

Typical integrations

Five system groups connect to the same agent. Which of them are in scope is decided in discovery.

  • Safety alerting

    • Gaggle
    • Bark for Schools
    • Lightspeed Alert

  • Filtering and activity

    • GoGuardian
    • Securly
    • ManagedMethods

  • Productivity and LMS

    • Google Workspace for Education
    • Canvas
    • Microsoft

Agent

Student safety monitoring

  • Reads the surface
  • Assembles the alert
  • Routes to a person

  • Rostering and cases

    • PowerSchool
    • Clever
    • Counselling case systems

  • Observability & evaluation

    • OpenTelemetry
    • Langfuse
    • Supported monitoring/evaluation sources

Integration availability depends on the client's existing systems and API access.

Agent controls

Six layers between the model and a welfare check

Layers wrap inward, and the map below names what each one misses.

L6 · Outermost — last line of defence

Inward → L1 · closest to the model

L6Rollback / safe modeStop triaging and pass each match to a person when evaluation or production signals degrade.Roll back
L5Version monitoringTrack model, prompt, term-list and scope-configuration changes.Track
L4TraceabilityRecord the trigger, the packet, the routing, the acknowledgement and the assessment.Record
L3Counsellor assessmentHold alerts for the on-call counsellor; it governs who reads one, not whether the match meant anything.Gate
L2Policy guardrailsTest alerts against the district's scope, hours and excluded terms; a failure returns the alert.Restrict
L1Confidence thresholdsRoute low-confidence alerts to a crisis-team triage read before the counsellor sees them.Require review
Model coreAlert produced — matched text, context class, source surface and confidence
L1 – L2
Test whether an alert may stand
L3
Puts the assessment in a counsellor's hands
L4 – L5
Retain the trigger and the assessment beside it
L6
Routes everything to a person rather than triaging

How Nestack evaluates it

Evaluate the whole alert path — not only the words that matched.

Coverage runs the whole depth of the workflow, and every layer is cut by slice.

Surface — the alert the counsellor sees

  1. E1Final-output evaluationDid the packet carry the matched text and enough context to read it?
  2. E2Step-level evaluationDid the agent use the right scope, term list and monitoring window?
  3. E3Tool evaluationDid it read the correct surface and route to the correct on-call recipient?
  4. E4Confidence calibrationDo low-confidence alerts actually attract more closures?
  5. E5Slice evaluationHow far apart do alert rates sit across specific student cohorts?
  6. E6Business outcomeHow many alerts closed with no risk, and how many were escalated?

Floor — the contact a student experiences

Failure modes

Where each failure originates in the agent

Seven modes, from signal to referral.

Agent lifecycle

Direction of processing →

  1. 01 · Retrieval

    1 mode

    • SM-03

      Snippet without its document

      A quoted line arrives stripped of the text it belongs to.

    Stage gathers

    School documents, messages, with the source each came from

  2. 02 · Reasoning

    2 modes

    • SM-04

      Fiction read as intention

      A first-person line in a draft is treated as a statement of intent.

    • SM-06

      Identity term read as risk

      Vocabulary about identity is scored as though it predicted harm.

    Stage proposes

    The matched text, its context class and confidence

  3. 03 · Tool / write

    2 modes

    • SM-02

      Alert routed past the counsellor

      The packet reaches a security list before a counsellor reads it.

    • SM-05

      Student text in a shared record

      The writing behind an alert lands where it can be disclosed unredacted.

    Stage writes

    Only where write access and approval policy allow it

  4. 04 · Output

    1 mode

    • SM-01

      Label without the sentence

      A category and a severity arrive with no text to read.

    Stage returns

    The alert the counsellor reads and acts on

  5. 05 · Change / Version

    1 mode

    • SM-07

      Silent term-list regression

      A dictionary update moves alert volume mid-year with no re-baseline.

    Stage tracks

    Model, prompt, term list and scope config

  • Sev-1 · contact made outside the boundary
  • Sev-2 · a wrong alert reaches a counsellor
  • Sev-3 · context degrades, alert routes to triage
See the complete Education failure-mode catalogue

Affected slices

Overall alert rates can hide one cohort

Break the alert rate down and the cohort names itself Nestack reports it by slice rather than in aggregate The cohorts that carry it are named, not averaged away Nestack reports it by slice rather than in aggregate..

Slice performance — reported separately, not only in aggregate

Illustrative example
SliceFailure rateLiftLift vs. thresholdStatus
Students writing about identity7.8%2.7×Review
Creative writing and journalism6.1%2.1×Review
Students in mental-health support5.2%1.8×Watch
Students in none of these cohorts2.6%0.9×Normal

Bar: alert-rate lift vs. the all-alert baseline · scale 0–4.0× · tick marks the 2.0× review threshold

2 of 4 slices over threshold

Evidence-linked improvement

A cycle ends in the suite

The cycle closes on a test, not a write-up — one the next release has to pass That suite is what the following detection is measured against..

Improvement cycle · five stages

Switchback — the path turns at Improve and returns at Learn

  1. 01

    Detect

    Alert rate rises in one student cohort.

  2. 02

    Diagnose

    Which moved — the term list, the context labelling or the scope? The alerts are read back until one answer survives.

  3. 03

    Improve

    Version, reviewer and the alerts behind the change stay together.

  4. 04

    Verify

    Re-run the affected cases; a fail holds the release.

  5. 05

    Learn

    A standing test, plus a change to the escalation protocol.

  6. Learn → Detect

    The return edge. Each later detection is measured against a longer suite.

Typical build scope

Twelve workstreams across six weeks

The build scope read against the delivery timeline. Week structure follows the six-week plan — discovery, surfaces, alert workflow, evaluation, integration, then production validation and handover.

Workstream

Week 1Week 2Week 3Week 4Week 5Week 6

  1. 01Escalation protocol discovery and boundary definition.Week 1
  2. 02Account, device and surface-scope assessment.Week 1, Week 2
  3. 03Term-list, scope and monitoring-window mapping.Week 1, Week 2
  4. 04Surface ingestion and text normalisation.Week 2
  5. 05Matching logic and context labelling.Week 2, Week 3
  6. 06Confidence scoring and on-call routing.Week 3
  7. 07Counsellor assessment workflow.Week 3, Week 4
  8. 08Case-system and rostering integration.Week 5
  9. 09Alert regression cases.Week 4, Week 5
  10. 10Guardrails and routing controls.Week 4
  11. 11Alert-trail instrumentation.Week 5, Week 6
  12. 12Deployment, documentation and Agent Care handover.Week 6

12 workstreams · 6 weeks · bar shows the weeks a workstream is active — several run in parallel

Final scope and sequence confirmed in discovery

Engagement tiers

What each tier includes

Rows are the capabilities named in each tier's scope. Higher tiers include everything below them.

Capability✓ in scope · — not at this tierPilotOne district, one surfaceProductionProduction case systemsAdvancedMultiple districts / surfaces
Introduced at Pilot
Matching to your list and scope✓✓✓
Counsellor assessment✓✓✓
Alert-quality baseline✓✓✓
Introduced at Production
Reporting by trigger class—✓✓
Routing workflow in your systems—✓✓
Approved write-back—✓✓
Case-system integration—✓✓
Introduced at Advanced
Multi-district and multi-state rules——✓
Multi-stage crisis-team routing——✓
High alert volume——✓
Enterprise escalation controls——✓
Build priceFrom $5,000From $8,000Custom quote
Final build price
Confirmed after discovery based on integrations, workflow complexity, alert volume, routing controls and deployment requirements.
Separate from build
Build pricing is separate from recurring Agent Care, which covers managed monitoring, evaluations, incidents and verified improvements after launch.

What we need from you

What you bring, and what we build with it

Each input maps to a piece of build scope and a week in the delivery timeline.

  1. 01Your escalation protocol and on-call roster

    Escalation mapping and on-call routing

    Week 1

  2. 02Representative alerts from your surfaces

    Matching baseline, context labelling and packet assembly

    Week 2

  3. 03Your approved term list and its exclusions

    Term-list, scope and monitoring-window mapping

    Week 1

  4. 04Access to relevant APIs, feeds or exports

    Account, device and surface assessment, then integration setup

    Week 2

  5. 05Alerts that were not a risk

    Alert cases and failure-mode testing

    Week 4

  6. 06What an alert must reach before anyone is contacted

    Confidence scoring, on-call routing, guardrails and escalation controls

    Week 3

  7. 07Named counsellors on the on-call roster

    Counsellor assessment workflow, then pilot and production validation

    Weeks 5–6

Nothing else is required

Deployment, documentation and Agent Care handover are ours.

Delivery timeline

Four phases across six weeks

Each band spans the weeks its work is named in, and week 5 carries two because both genuinely run.

Phase

W1

W2

W3

W4

W5

W6

  1. DiscoveryW1
  2. BuildW2 – W3
  3. EvaluateW4 – W5
  4. Pilot & LaunchW5 – W6

Week focus

W1Escalation discovery, term-list mapping and the automation boundary

W2Surface integration and the matching baseline

W3Alert workflow, confidence logic and routing controls

W4Evaluation suite, cohort alert-rate checks and failure-mode testing

W5Case-system integration, a pilot alert window and targeted corrections

W6A month of alerts triaged by the counselling team, then handover

Reading the band
Bands sit on the weeks the work is named in and nowhere else. The doubled fifth week is real.
At the end of W6
Hand-over happens against a verified alert window, not a demo.
Duration
Six-week plan shown · typical delivery 4–6 weeks depending on scope confirmed in discovery.

Next step · Education AI agent

Build a safety-monitoring agent around your district's escalation protocol.

Show us your surfaces, your term list and your on-call roster. Your counselling lead tells us where an alert has to land, and we map the routing, set the automation boundary and name what stays with a person.

Typical delivery
4–6 weeks
Deployment
Your district stack
Counsellor read
Alert only
After launch
Agent Care available

Nestack Agents · Student safety monitoring

AGT-ED-07 · Agent Care available after launch

Build a safety-monitoring agent

Four questions so the call starts from your actual setup — then pick a time that suits you.

About you
Your surfaces

No obligation. We’ll map the workflow, name the automation boundary and tell you if an agent is the wrong answer.