Most school software is a place to record what already happened. Someone marks the register, someone enters the receipt, someone types the circular. The software stores it faithfully. The following-up — the second reminder, the third phone call, the noticing that a child has been absent all week — stays with a human being who is already doing four other jobs.
Autopilot is the part of EdunodeX that does the following-up. This page explains what it actually is, mechanically, so you can judge it rather than take an adjective on trust.
What “Autopilot” Means Inside EdunodeX
Autopilot is a catalogue of packs. A pack is a named, self-contained automation with a fixed purpose: chase overdue fees, flag absence streaks, remind the accountant about a statutory deadline. Enabling a pack for your school writes one or more rules into your school’s own database schema. From then on, a background worker evaluates those rules on their schedule and dispatches whatever they are configured to dispatch.
Nobody opens a screen for this to happen. That is the whole point, and it is also the reason the guardrails below matter more than the feature list.
Packs are grouped into categories that map to the departments that own them — fees, attendance, academics, engagement, compliance, operations, and payroll. You can see the current catalogue, with each pack’s settings and its estimated cost per firing, on the Automations screen.
The Anatomy of a Pack: Condition, Action, Schedule
Every rule a pack builds has exactly three parts.
Condition — a typed query against your data. fee_ledger_overdue with a days-overdue value and a minimum amount. ml_score_threshold reading a risk score. attendance_pending_after_period checking whether a register was marked. Conditions are compiled server-side from a fixed vocabulary; no language model writes SQL against your school’s records.
Action — what happens when the condition matches. The supported action types are an approved WhatsApp template, an SMS, an email, an in-app notification, a push notification, or the invocation of a named internal intent that assembles a summary. Each action carries its variable map, its deduplication key, and its deduplication window.
Schedule — cron, interval, one-off, or event-driven. Cron schedules are expressed in your school’s own timezone, so “10:00” means ten in the morning where the school is, not on a server somewhere.
Because all three parts are data rather than code, a pack’s behaviour is inspectable and adjustable. When you change the fee reminder from seven days to ten, you are editing a threshold, not asking anyone to redeploy anything.
Rule Mode and Agent Mode: Two Ways a Pack Decides
Most packs run in rule mode. The condition matches, the action fires, the message uses a pre-approved template with variables substituted. Deterministic, cheap, and predictable — a seven-day reminder is a seven-day reminder.
Some packs also implement agent mode. Here, deterministic SQL still selects and enriches the candidate rows — the model never queries your database — and a single structured-output call decides something judgement-shaped: which tone this particular guardian should get, which absence streak looks chronic rather than incidental, which three defaulters actually warrant Monday’s phone calls. The output is validated against a strict schema and anything unrecognised is discarded.
Agent mode is opt-in per pack, and each pack declares a narrow allow-list of actions its agent may execute directly. Anything outside that list is forced into a review queue instead of being carried out. Tenant data handed to the model is wrapped in explicit delimiters with instructions to treat it as untrusted content, so a note typed into a student record cannot act as an instruction.
Guardrails: Deduplication, Circuit Breakers, Send Windows
Automation that can spam parents is worse than no automation. Three mechanisms exist specifically to prevent that.
Deduplication. Before any message is dispatched, the action’s deduplication key is claimed atomically in Redis with a time-to-live equal to the configured window. Two workers evaluating the same rule in the same tick will see exactly one successful claim. Keys are namespaced per school. Critically, this check fails closed: if Redis is unavailable, the action is skipped rather than sent, because a skipped tick recovers on the next evaluation while a duplicate message cannot be unsent.
Circuit breakers. Each rule tracks consecutive errors. At five, the rule is moved to an error state and stops firing entirely. Admin and principal users receive an in-app notification, deduplicated so a broken rule cannot itself become a source of noise, and WhatsApp actions that tripped the breaker are recorded in a dead-letter table for inspection.
Send windows. Packs schedule their sends at civilised, configurable hours — fee reminders default to 10:00 local, the collection briefing to 08:00, the attendance wrap to 17:00 after roll-call. Several packs expose skip-weekends and skip-holidays settings. Separately, broadcasts on intrusive channels — push, SMS, WhatsApp — are held during a school-default quiet window of 21:00 to 07:00 IST, with emergency-category messages exempt.
Where to Start: A Three-Pack Rollout Order
Enabling everything on day one is how schools end up distrusting automation. A sequence that works:
- Something read-only, addressed to staff. The daily collection briefing or the attendance summary. Nothing leaves the building. You learn whether the numbers match what you already believe.
- Something transactional and welcome. Payment confirmations. Parents are pleased to receive them, and it exercises the WhatsApp path end to end.
- Something that chases. Fee follow-up or absence streaks. By this point you have a week of firing history and you know the system’s numbers are right.
Watch the firing log after each step. Every evaluation writes a row recording what matched, what was dispatched, what was skipped as a duplicate, and what failed.
What Autopilot Deliberately Does Not Do
It does not submit anything to a government portal on your behalf. Compliance packs surface deadlines and prepare the relevant file for review; the filing remains a human action.
It does not make financial decisions. Waivers, concessions, reversals and adjustments route to the human approvals queue regardless of what any model suggests.
It is not included in EdunodeX Lite. The pack runner is explicitly gated out of that tier; Lite schools get view-only AI surfaces instead.
And it does not run without an audit trail. Every firing, every deduplication skip, every circuit-breaker trip is a row you can read. Autonomy without a record is just software doing things you cannot check.