Reference deployment: UAE policy pack KHDA · ADEK · SPEA · MoE Human approval on every write

AI agent for schools · K-12 · UAE / GCC

The AI agent for schools that never sends without you.

Alan reads your school’s own systems, finds what needs attention before the first bell, drafts the follow-up, and stops at a named human. No new SIS. No silent sends. No invented answers.

  • Read-only start
  • Offline mode available
  • Human approval on every write
ALAN / MORNING BRIEF MON 16 NOV · 06:30 GST
MSOE Academy, Dubai Synthetic demo · 50 students
  • 11 Unexcused absencesGrade 7 · 4 need a call today
  • 3 Clinic follow-ups overdue2 medication consents expire
  • 2 IEP reviews overdueInclusion · ADEK pack
  • 6 Grades awaiting moderationOldest 19h · 2 low confidence
4 Need your approvalAttendance · clinic · grading · cover

Every line is a row count from the school’s own tables.

Reference run · ~6 seconds · 0 side effects executed

In one sentence

Alan is an AI agent for schools. It sits beside your SIS, reads the records and policy you already own, and turns them into counted facts, drafted actions and a short morning brief — with a human approval gate on every write.

01 / The 6:40 problem

The first bell is at 8:00. The work started at 6:40.

A school morning is not short of effort. It is short of sorting. The right task reaches the right person late, or not at all, because it arrived in the wrong channel.

A parent texts in Arabic. A clinic consent expired. An IEP review lapsed in March. Two teachers need cover. A grade anomaly is sitting in a spreadsheet. None of it is hard. All of it is urgent. And every item lives in a different system, a different inbox, or a different person’s head.

This is not a staffing problem. It is a sorting problem.

  1. WhatsApp · غياب

    My son is sick today.

    Arabic reply, no student ID, no excuse document. The registrar starts the thread from scratch.

  2. Clinic

    Medication consent expired.

    Two students on today’s list. The nurse cannot administer without a signed, unexpired consent.

  3. Inclusion

    IEP review overdue by 9 days.

    The evidence exists in three places. Nobody has assembled it into one citable report.

  4. Cover

    Two teachers absent. One exam block.

    Twenty-two candidates on paper. None ranked by subject, capacity or the timetable in front of them.

  5. Grading

    Six submissions awaiting moderation.

    One is low-confidence. One has a blank scan. Both will reach a report card if nobody looks.

  6. First bell

    The day begins already behind.

    Eighty minutes of skilled attention spent sorting, not deciding.

02 / The product

Before the first bell, the day is already sorted.

Alan’s Overview is not a dashboard you have to feed. It is a brief built from row counts in your own systems: what happened, what is overdue, what needs a person, and what Alan has already drafted.

Alan
Overview Policy pack current · v2026.11

Good morning, Ms. Rahman.

Monday 16 November · 06:30 GST · MSOE Academy, Dubai

Interactive demo
Work today27agent runs since 00:00 GST
Needs you4pending human approval
Files12artifacts in workspace
Hours returned7.8engagement to date

Morning brief

Counted at 06:30
  • 11Unexcused absencesattendance_logs · 4 need a callsource
  • 3Clinic follow-ups overdueclinic_episodes · 2 consent gapssource
  • 2IEP reviews overduesen_profiles · ADEK packsource
  • 6Grades awaiting moderationgrading_submissions · oldest 19hsource
  • 1Timetable gap tomorrowtimetable_slots · exam blocksource

Needs you

4 pending
PENDING_HUMAN_APPROVAL
Send attendance follow-up

Grade 7 · 4 families · EN / AR

  • PII scrub passed
  • Curfew 07:00
  • Policy clause 4

  • Clinic

    Medication consent renewal

    1
  • Grading

    Publish 6 moderated results

    1
  • Cover

    Dispatch tomorrow’s plan

    2

Demo only. Nothing leaves this page.

01

Counted, not composed

Every brief line is one count against a table the school already owns. The source and as-of time travel with it.

02

Cited, not guessed

Every claim opens the clause, graph path or row behind it. Missing data returns NO_EVIDENCE, not a paragraph.

03

Staged, not sent

Every external write is PENDING_HUMAN_APPROVAL. Alan cannot approve its own work.

03 / How it works

Five steps. One human decision.

The loop is deliberately boring. That is what makes it auditable.

  1. 01

    Read

    Roster, attendance, clinic, grading, inclusion, timetable, fees and the policy pack. Read-only by default.

  2. 02

    Explain

    What changed, what is overdue, and which clause governs it. Every line carries a source and an as-of time.

  3. 03

    Propose

    Drafts the parent message, the cover plan, the moderation note or the document. Names the approval action.

  4. 04

    Approve

    A named human approves, edits or rejects. Alan has no approve tool and no send tool.

  5. 05

    Record

    Replay-safe outbox, side-effect audit, evidence file. One decision, one trace.

READ → EXPLAIN → PROPOSE → HUMAN APPROVES → REPLAY-SAFE OUTBOX → AUDIT

Works offline. MSOE_LLM_EXTERNAL=off is a supported mode, not a fallback.

04 / What Alan does

Eight jobs that eat a school week.

Alan is not a chat window bolted onto a database. Every service is a native tool with a permission, an effect and an evidence contract.

What Alan does in each school workflow, and what remains with a human.
Domain Alan does A person keeps
01Attendance & family comms Classifies English and Arabic replies, applies excuse windows and staged warnings, drafts the message, stages the attendance change. The registrar approves the change and the send. The parent relationship stays human.
02Clinic & medication Closes episodes, checks consent and allergies, reconciles attendance, detects repeat-visit patterns and drafts the follow-up. The nurse and principal keep every clinical decision. Missing consent is a hard refusal.
03Grading & moderation Marks objective questions deterministically, scores structured answers with per-question evidence, flags anomalies and stages publication. The teacher moderates. A named human publishes. No grade reaches a report card alone.
04Inclusion & SEN Finds overdue IEP reviews and intervention gaps, builds citation-grounded reports, stages eSIS / PASS updates. The inclusion lead owns the judgment and the family conversation.
05Timetable & cover Finds the gap, ranks substitutes by subject and capacity, repairs the day and stages dispatch. Leadership approves the cover plan and the duty-of-care call.
06Documents & briefs Renders DOCX, XLSX, PPTX and PDF from persistent templates, with spec hash, evidence and a version. Staff edit, sign and own what leaves the school.
07Voice & capture Push-to-talk, hands-free and call-assist sessions turn a spoken thought into a structured note or draft. A transcript can stage an action. It can never approve one.
08Portal work Prepares forms in API-less portals, verifies the page before save, and stages the submit with a screenshot hash. A human submits. No silent portal write, ever.

No write tool exists in Alan. The approval gate is architectural, not a setting.

05 / The difference

An answer you can open.
Or no answer at all.

Ask a general AI what is happening with Grade 7 attendance and it will write a confident paragraph. Ask Alan and you get a path: Student → RiskFlag → Standard → Policy. If the graph returns nothing, Alan refuses.

GROUNDED25 evidence rows

“Which students are at risk of chronic absence, and what policy governs intervention?”

25 students are flagged under CHRONIC_ATTENDANCE and linked to governing clauses. Every claim below is backed by an explicit graph path and policy citation.

StudentRiskFlagStandardPolicy
25 evidence rows · 2 policies cited · MSOE-HS-2026-02
REFUSED0 evidence rows

“What is our carbon footprint in tonnes per year?”

I cannot answer that from verified sources. No matching risk-to-policy paths were found, and the question did not identify a grade whose roster could be counted. I will not generate an ungrounded answer.

REFUSED_INSUFFICIENT_DATA
Grounded-answer policy · hallucination is not permitted
74/74risk cases matched to a governing policy clause
19four-hop evidence paths in one agent run
0external actions executed without approval

Reference run on a 50-student synthetic school, 07–12 Nov 2026. Your pilot is measured on your own records.

The gate is the architecture

Four things Alan cannot do.

  • No write tool. Alan’s tools are read or propose. A write tool fails the manifest.
  • No direct send. Mail, WhatsApp and portal submissions go through the approved dispatch path.
  • No invented fact. A missing clause, count or graph path returns NO_EVIDENCE.
  • No cross-tenant read. school_id comes from the signed session, never the request body.
approval.jsonread-only view
{
  "status": "PENDING_HUMAN_APPROVAL",
  "alan_can_write": false,
  "approval_action": "SEND_PARENT_NOTIFICATION",
  "evidence": {
    "source": "attendance_logs",
    "as_of": "2026-11-16T06:30:04+04:00",
    "query_hash": "…",
    "row_count": 4
  }
}

06:34:07 approved by S. Rahman · 06:34:09 outbox row created · 06:34:11 delivered · evidence sha256 …

“If an inspector asks ‘on what basis?’, you open the path.”

06 / Fits your stack

Keep the systems you already pay for.

Alan sits beside your SIS. Keep the roster, gradebook, portal and parent app. Connect what you have, start with one workflow, and expand only when the numbers earn it.

Scope an integration

SIS / MIS

GibbonPowerSchooliSAMSOpenEduCatCanvasOneRosterCSV drop

Communication

GmailMicrosoft 365 / TeamsWhatsApp Business CloudCalendarDrive

Policy & evidence

KHDAADEKSPEAMoE / SFEAyour own policy pack

Deployment

Private cloudon-premUAE residency checksoffline mode

If it has a roster export, we can start. One manifest drives the host name, role map, branding and policy pack.

07 / The pilot

Thirty days. One workflow.
A number you can defend.

We don’t ask you to bet a term on a platform. Start read-only, pick the workflow that costs the most hours, and measure it against your baseline.

  1. Week 0

    Scope & connect

    One SIS export, one policy pack, one workflow. Read-only.

  2. Week 1–2

    Shadow mode

    Alan produces briefs and drafts. Staff approve or reject. Nothing auto-sends.

  3. Week 3

    Measure

    Hours returned, approval SLA, exception rate, staff confidence.

  4. Week 4

    Decide

    Expand, adjust or stop. You keep the evidence file either way.

Request a pilot scope 20 minutes. Your workflow, not a canned tour.

08 / Questions

What school leaders ask first.

Short answers, no slideware. Ask us anything else in the walkthrough.

Prefer to write? hello@alan.school

What is an AI agent for schools?

An AI agent for schools is software that reads a school’s operational records, understands the policies that govern them, and does the repetitive work of sorting, drafting and following up — while a named staff member keeps every decision. Alan is that agent for K-12 schools. It works beside the SIS rather than replacing it.

How is Alan different from ChatGPT or another AI tool for educational institutions?

General AI tools answer from whatever is in the prompt. Alan answers from the school’s roster, attendance, clinic, grading and policy records, returns the source path behind each claim, and refuses when the data is missing. It also cannot send, submit or approve anything on its own.

What should a school look for in an AI tool for educational institutions?

Four things: answers grounded in the school’s own records, a citation for every claim, an explicit refusal when data is missing, and a human approval gate on every external action. Ask vendors to demonstrate all four on your data — not on a canned demo. Alan is built around those four requirements.

Does Alan replace our SIS?

No. The SIS remains the system of record. Alan reads a canonical roster through an adapter — Gibbon, PowerSchool, iSAMS, OpenEduCat, Canvas, OneRoster or a CSV drop — and writes only through the approved path. Schools keep their portal, gradebook and parent app.

Will Alan send messages to parents on its own?

No. Alan can classify a parent reply, draft the response and stage the send, but the outbox dispatches only after a named human approves the action. There is no send tool in Alan. PII scrubbing and a messaging curfew run before anything leaves the system.

How does Alan keep student data private?

Alan enforces tenant isolation on every read and write, scrubs PII before outbound messages, and supports a deterministic offline mode where no external model is called. Data residency checks and retention sweeps are part of the deployment. A pilot can run fully offline; in that mode, no student data leaves the deployment for inference or model training.

Which systems does Alan integrate with?

Alan reads from Gibbon, PowerSchool, iSAMS, OpenEduCat, Canvas, OneRoster and CSV exports, and connects to Gmail, Microsoft 365 and Teams, WhatsApp Business Cloud, Calendar and Drive. Policy packs cover KHDA, ADEK, SPEA and MoE references in the UAE reference deployment, and can be replaced with a school’s own pack.

What does “grounded in policy” mean?

Every claim is tied to a clause in a versioned, effective-dated policy pack and, where relevant, a path in the school knowledge graph. If the graph returns nothing, the answer is NO_EVIDENCE, not a guess. The citation travels with the answer so a principal or inspector can open it.

How long does a pilot take?

Thirty days, one workflow, read-only to start. Week 0 scopes and connects; weeks 1–2 run in shadow mode; week 3 measures hours returned, approval SLA and exception rate; week 4 is a decision. The school keeps the evidence file either way.

What does Alan cost?

Pilots are scoped to one campus and one workflow. Pricing depends on the workflow, the SIS and the policy pack, so we quote after a 20-minute scoping call — not from a generic per-seat page. The pilot is designed to produce a measured hours-returned number before you commit to a larger rollout.

Who approves Alan’s work?

A named staff member with the right role — registrar, nurse, inclusion lead, head of department or principal. Alan cannot approve its own proposals, and every decision is written to the audit log.

Is Alan available in Arabic?

The interface and family communications support English and Arabic, including right-to-left rendering. The reference engine also renders French and Hindi family digests. Policy packs and message templates are localised per deployment.

09 / At a glance

Alan, on one page.

Product
Alan
Category
AI agent for schools / school operations
Built by
Alchemist
For
K-12 school leaders, registrars, inclusion leads, IT
Deployment
Beside your SIS; private cloud or on-prem
Start
30-day read-only pilot, one workflow
Guardrail
Named human approval on every write
Grounding
Versioned policy pack + knowledge graph; NO_EVIDENCE on gaps
Languages
English / Arabic (RTL); French & Hindi family digests
Interfaces
Embedded SIS panel, web app, push-to-talk voice

Last updated: · Reference deployment: UAE policy pack

10 / Walkthrough

Bring your messiest Monday. We’ll show you Alan’s Tuesday.

A 20-minute working session, not a generic demo. Tell us the task that eats the most time and we’ll scope it live.

  • 20 minutes, no slides
  • Your workflow, not a canned tour
  • Read-only pilot scope, if it fits

Prefer email? hello@alan.school

Request a walkthrough 20 min

We’ll reply with two times. No newsletter, no cold sequence. This form opens your email app; nothing is stored on this page.

Book a 20-minute walkthrough