OAIA AI assistant

Overview

Menu path: OAIA (/admin/oaia).

OAIA (the Omaya AI Assistant) watches the deployment, reports what changed, proposes improvements, and answers questions about the organisation's own data in plain language. With Autopilot on, it can also make a small, bounded set of changes itself.

The design principle comes first, because everything below follows from it:

The detectors detect. The model narrates.

Omaya's own code decides what is wrong and how serious it is. The model is asked to explain it in a sentence a person can act on. It never sets a severity, never sends a notification, and, outside the short list of changes that a person approves or allows Autopilot to make, never changes anything.

An organisation with no AI provider therefore still receives everything except the prose: findings, escalations and notifications all work, worded by the detector rather than the model.

NOTE

This chapter describes configuration and daily use. AI in Omaya explains the machinery in more depth: the full detector list, the action catalogue, and how "your data only" is enforced.

Enabling OAIA

Applicable versions / plans

SPECIFICATION

• OAIA must be offered to the organisation by the platform operator. Otherwise the page shows OAIA is not included for this tenant. Contact Omaya to have it enabled.
• Switching OAIA on, and choosing its intervals, is done on the AI card of System > Configuration. Saving that card needs manage-settings-general, which only the Admin role holds as shipped.
• Opening the OAIA page: view-oaia (Admin, Manager and Viewer as shipped).
• Changing Assistant settings, running Check now, and Undo / Approve on Autopilot rows: manage-oaia (Admin and Manager as shipped).
• An AI provider for the prose and the chat: the organisation's own key, or the platform's, chosen on the same card. See AI key selection.

Requirements

At Northwind Tower, Aisha Rahman wants OAIA to watch the building and tell the administrators when something changes, without being interrupted by routine noise. Ravi Kumar reviews its suggestions once a week.

• OAIA switched on for the organisation.
• A health check every two hours.
• A review of the whole deployment once a week.

Restrictions

• OAIA is off by default.
• Heartbeat every: 2, 4, 6, 8, 12 or 24 hours. Default 2 hours.
• Review every: 1 (daily), 2, 3 or 7 days. Default daily.
• The heartbeat interval sets how quickly news arrives, not how much OAIA does: a two-hour heartbeat on a quiet estate produces almost no messages, because it only reports when something changed.
• Only a role with manage-settings-general (Admin as shipped) can switch OAIA on or off. A role that holds manage-oaia but not manage-settings-general (Manager as shipped) uses OAIA and changes its other settings, but cannot switch it.

Configuration roadmap

(1) Open the AI card of System > Configuration.

(2) Switch OAIA on and choose the two intervals.

(3) Save the card and open the OAIA page.

Procedure

1. Choose System > Configuration and click the AI card.

2. In the OAIA — Omaya AI Assistant section, set the fields as follows.

   Configuration Item	Value	Note
   Enable OAIA for this tenant	On	Off by default. Disabled when the platform has not offered OAIA to the organisation.
   Heartbeat every	2 hours	2, 4, 6, 8, 12 or 24 hours. Default 2 hours.
   Review every	7 days	day (daily), 2, 3 or 7 days. Default daily.

3. Click Save.

Verification

• The card saves with AI settings saved.
• Choose OAIA. The page opens with the status line Watching · last checked … · next check ….
• While OAIA is off, the page instead says The assistant is switched off for this tenant. Nothing is scanned, no tokens are spent and no admin is notified until it is turned on from the A.I Model card. A user with manage-settings-general sees an Open the AI card button under it; any other user sees An administrator who can change the organisation settings can switch it on from Configuration → AI.

The OAIA page

Layout

The page reads top to bottom in the order of use: what happened, what to do about it, and a place to ask.

Part	Content
Status line	Watching · last checked … · next check … under the OAIA name, or Paused. Next check due now means a heartbeat is about to run.
Today's briefing	The latest digest as a headline and its points, numbers in bold, with a green or red dot on each point that went well or needs someone. The card's colour edge and pill give the state: red 2 critical open, amber 3 warnings open, grey Nothing new · 62 on hold (all snoozed or acknowledged) or green All clear. Below it, a button to the most urgent action (Review the 27 critical) and Ask about this briefing. On a phone it shows the headline and two points, with Read more.
Ask OAIA	The chat, and the main part of the page: the wide column on the left, as tall as the window and staying in view while the page scrolls.
Needs your attention	The column on the right. At its top, four tiles: Act now (critical findings), Look this week (warnings), Improvements (suggestions from the review) and Checks (the detectors, and whether the last run of each worked); clicking a tile opens its tab. Under them, a Filter by room, device or person box and five tabs.

The tiles count only what still needs someone. A finding that was acknowledged or snoozed is muted: it is shown beside the number (critical · 8 muted) rather than in it, and is listed on the Snoozed tab. The Act now and Look this week numbers therefore always add up to the count on the Problems tab.

Tab	Content
Problems	Open findings, banded Act now (critical), Look this week (warning) and For your information. When one check raises many findings they read as one line, 27 × Security incident unattended, with Show the 27 and Acknowledge all 27.
Improvements	Suggestions from the review: Ready to apply (those carrying a rule or an action) first, then Ideas to consider.
Snoozed	Findings that were acknowledged or snoozed, with Reopen.
Resolved	Findings that cleared in the last 7 days, proposals OAIA closed itself because the gap was fixed, and won't-fix ones.
Activity	The recent heartbeats, as a trace (+ new, − cleared) and a list, and Autopilot's record when Autopilot is offered, with Undo and Approve.

The selected tab is remembered while the browser tab stays open.

Buttons

Every action is a visible button before the pointer reaches it. The action OAIA recommends is solid blue (Acknowledge, Set capacity, Create rule); the alternatives are outlined (Snooze, Dismiss, Mark done); the lighter ones, Open, Ask about this, See the rule and Show the 27, have a pale blue fill. Read more, New chat and ← Back are plain links, and ⋯ holds the rarely used ones, such as Won't fix.

Open on a finding or a suggestion shows that one item in place of the list: OAIA's reading of it, the facts (where, when, the readings that raised it), the suggested next step, its buttons, a few questions that can be sent to OAIA with one click, and the raw evidence folded under Technical details. The chat is pinned to the item while it is open. ← Back to Needs your attention, or Esc, returns to the list.

Assistant settings

Settings opens Assistant settings (OAIA on for this organisation, Check every, Tenant review every) and, under About this assistant, the AI model and key OAIA runs on, how often it checks, its last run and its recent token use. Check now runs a heartbeat immediately instead of waiting. Both need manage-oaia.

CAUTION

OAIA on for this organisation is the same switch as Enable OAIA for this tenant on the AI card. Clearing it switches the whole assistant off, and the page then shows the switched-off message. It needs manage-settings-general: for any other role the switch is greyed out, with An administrator who can change the organisation settings can switch it on from Configuration → AI. under it.

The heartbeat

How it works

The expensive step is last and conditional. A quiet estate produces heartbeats that end at "nothing changed".

At every interval the detectors run over the organisation's own data. Each looks for one specific thing rather than general strangeness:

Detector	Raises a finding when
Gateway offline · Gateway signal	A gateway has stopped reporting (critical after 24 hours), or several gateways in one venue stopped within 15 minutes of each other, which points to power or the network rather than the gateways; a gateway's received signal over the last 30 days is falling steadily enough to forecast a failure; a gateway that keeps dropping: three or more outages of its own in 7 days, not counting outages shared with other gateways within 15 minutes (a site-wide power or network cut).
Ingestion silence	A venue's room decisions have fallen to a tenth or less of the same hours a week earlier, even though its gateways may still look online. A venue with too little history to compare is left out.
Device anomaly	A device has gone silent, is draining, or reports far less than its baseline.
Battery	A device's battery is at 20 % or less (critical at 10 % or less), from a reading taken in the last seven days; or a tag's battery voltage is below the low-voltage line set for its model, quoted in mV.
Environment baseline	A room's latest temperature is far from its own normal for that hour of the day over the last 14 days, by at least 2 °C (critical at 5 °C or more).
Outside	Only for organisations that track people or things with GPS. Someone has stopped away from every site and geofence for far longer than their own usual stops over the last 30 days; a tracker that was outside when last heard has been quiet for longer than its usual gap between fixes, and at least an hour; a tracker used outside has a battery at 35 % or less, or one expected to run out within 14 days; an asset or vehicle moved outside site hours, unless it does so on most days. A stop at a person's home is never a finding, and no finding carries map coordinates.
Connectors	An integration has stopped: its latest error is newer than its latest success, or a webhook's deliveries died today.
One per enabled app	Energy (something left on after hours in an empty room; a meter whose kWh counter stopped while it still drew power), Facility (a work order overdue or untaken; assets a week or more past their service date), Meeting Room (calendar sync stopped, bookings open past their end), Security Watch (an incident unacknowledged or open too long), Toilet Hygiene (a toilet in the red with nobody sent), Visitor (overstay; a visit waiting for approval past its start), Compliance (an assessment or calibration going stale), Roll Call (a drill overdue, a roll call left running), Data Centre (an alert nobody acknowledged; half or more of the plant no longer reporting) and Warehouse (an alert nobody acknowledged, including stock in a storage area outside its temperature or humidity range).

An app that is not enabled contributes nothing and is never mentioned, so the model cannot be led into speculating about something the organisation does not have. A fall the guards were told about is a Security Watch incident, so it counts there; every suspected fall, including the ones nobody was told about, is in Reports > Fall & Man-down.

Findings, and what to do with them

A finding is one problem, tracked over time rather than re-announced. Its severity (info, warning or critical) is set by the detector. Each finding states what happened and, under it, the likely cause in plain words; its actions are buttons directly below.

• The suggested action, when OAIA chose one, is the first button. It is picked from a fixed list: Check power at the gateway, Check network / uplink, Inspect the gateway, Replace the battery, Review the device, Check the sensor placement, Review the rule, Draft a rule for this, Contact Omaya support, No action needed.
• A finding that stays open has its likely cause written again every few days from that day's evidence, so the explanation does not lag the facts. OAIA does not blame a battery when the device's last battery report was 30 % or more.
• Where there is a page for the action (Gateways, Devices, Rules), the action is a link to it.
• The raw evidence (the check, first and last seen, and the detector's own fields) is folded under Technical details.

Button	Effect
Acknowledge	Marks the finding as seen. It leaves the open list and stops counting as new. A grouped line (72 work orders nobody has taken) comes back if it grows by a quarter and by at least 3 since it was acknowledged.
Acknowledge all	On a grouped line: acknowledges, at once and after a confirmation, exactly the findings the line shows. With the filter box in use that is only the ones that match it, and the count on the button says how many.
Snooze	Hides the finding for 4 hours, 1, 3 or 7 days. If it is still true when the snooze ends, it comes back and notifies again.
Reopen	Undoes an acknowledge or a snooze, from the Snoozed tab.
Won't fix	Under ⋯. Keeps the finding out of notifications permanently, without pretending it went away.
Ask about this	Opens the chat with the finding already in front of it.

Who is notified. Every active administrator of the organisation, as an in-app notification (the bell, and the mobile app's Alerts tab under the OAIA chip), and only when the heartbeat produced something genuinely new: a new finding, an escalation, a snoozed finding that came back, or an acknowledged group that grew. An unchanged finding is the same news as last time and does not interrupt anyone twice. Info-level findings appear on the page and never interrupt.

Example

At 08:30 the page shows two findings: Gateway offline: Level 3 East (critical) and Battery: 4 tags below 15 % (warning). Ask about this on the gateway makes OAIA check its last packet and the tags it used to hear; it reports that the outage started at 02:10 with everything on that gateway at once, likely power. The administrator sends a technician, clicks Acknowledge on the gateway finding, and snoozes the battery finding for a day until the spares arrive.

The review

On the cadence set in Review every, OAIA looks at the deployment as a whole rather than at events. Omaya builds the facts (venues and their hours, zones and capacities, devices and where they are placed, what each app is set to do, what is measured and what is not) and a list of verifiable gaps, such as a venue with no hours or a counter zone with no capacity. The model turns that into up to five prioritised suggestions in five categories: Automation, Safety, Well-being, Efficiency and Configuration.

Button	Effect
Set capacity	On a suggestion to give a room a capacity: the box names the room, contains the number OAIA proposes, and shows the room's capacity now beside it if someone has set one since. The number can be changed (a whole number, 1 or more; a typed number is kept while the page refreshes). The capacity is set at once, and a note confirms it (Boardroom 12 capacity is now 8.). Needs edit-spaces.
Apply	On a suggestion that carries another ready-made change: makes that change, with the same checks as applying one from the chat. Apply will: above it states what it does. A change is applied once: if someone else has just applied or closed the suggestion, the page says it is already closed and the list refreshes.
Do it with OAIA	On a suggestion with no ready-made change: hands it to Ask OAIA, which works out the exact change (asking for any number it needs) and offers it to Apply in the chat. It waits if OAIA is still answering something else.
Create rule (inactive)	For a suggestion that carries a rule: saves it switched off, for review and activation. Draft saved — switched off until you enable it. Needs create-rules. See the rule shows the draft first.
Mark done	The gap was closed another way.
Dismiss	Not wanted. A dismissed idea is not suggested again for 30 days.

• A suggestion that carries a rule carries a valid rule. The draft goes through the same checks as the rule builder, so it can never offer something the builder would refuse.
• Suggestions close themselves once the gap is fixed; they move to the Resolved tab with the reason (Closed by OAIA: capacity now set — Boardroom 12 8). Most are checked at each heartbeat; a capacity suggestion is checked every time the page loads, so a room given a capacity elsewhere stops asking immediately. Until the last room in it has one, the card lists those left: Still without a capacity: Huddle Room 3, Pantry L2 · done: Boardroom 12. A card that names several rooms stays open when one of them is set, and closes itself when the last one has a capacity. A suggestion to change a capacity a room already has never closes on its own: that is a decision for a person.
• Older suggestions are looked at again. Once a day, OAIA rechecks up to six open suggestions that are more than three days old against the current data: whether the work orders it named are still open, whether the room is still without a sensor, whether the rule is still missing. A suggestion whose problem is clearly gone moves to Resolved with the evidence (Rechecked by OAIA: work order 1042 was cancelled on 9 Oct). One that still holds, or that the data cannot settle, stays open and is looked at again three days later. A sensor or meter that has gone quiet never counts as fixed: OAIA leaves that card open.

Asking OAIA questions

How it works

The chat cannot write anything by itself: its only way to change something is to put a proposal in front of you. Proposals made from Telegram are never applied there; they wait on the OAIA page under "Proposed from the app or Telegram — apply it here".

Type a question in Ask OAIA and click Ask, or pick one of the starter questions shown while the conversation is empty. After each answer, You could ask next offers up to three follow-up questions that fit what was just said, naming the gateway, room or number; clicking one asks it. They are suggestions only: any other question can be typed as usual, and they disappear once a question is sent. The grey hint in the box changes as the conversation goes on and is never sent.

Ask about this on any finding or suggestion pins it above the box (About: …), and it stays pinned for follow-ups (cancel both as stale still identifies the two work orders meant) until ×, New chat, or Ask about this on another item. Questions are phrased as they would be to a colleague, for example:

Why is the Level 3 East gateway offline?

Which zones stopped reporting yesterday?

How many people were on Level 2 at 3pm on Tuesday?

Is anyone in the Server Room right now?

While OAIA works, a line shows the current step (Looking at what data exists, Checking the data, Reading the Omaya documentation, Writing the answer), and the answer appears as it is written. A simple question takes a few seconds; one that needs several queries can take half a minute. OAIA has a limit on the number of lookups for one answer; when a question needs more, it answers from what it found, states what is still unknown, and suggests how to ask for the rest. If it guesses a name in the data wrongly, it is given the correct names and tries again, so a slip costs one lookup, not the answer.

Data available to OAIA

• Ready-made lookups for gateway status, finding a device, who and what is in a zone now, and the open issues; one read-only query at a time over the organisation's own views; and the Omaya documentation.
• The views cover, among others, state changes (doors, leaks, lights and switches, air-conditioning), outside weather, gateway outages, tag health (battery voltage, and signal for LoRaWAN devices), distance sensors, noise, custom sensor fields and each meeting room's day, so How long was the cold room door open yesterday? or Which gateway dropped most this week? are answered from the data.
• Answers use the organisation's data, in the organisation's time zone.

"How many were here?" For a venue, a location or the whole organisation, OAIA reads Footfall's own figure: the registered tags, each counted once per place. It is instructed never to add up the zones, because someone who walked through four rooms would be counted four times. It answers report questions like this one, and states that the reporting assistant is built for them. For which people figure answers which question, see How many were here.

Ratings, long conversations and history

• Rating an answer. Under each answer, Helpful (thumbs up) or Not right (thumbs down), with an optional What was wrong? (optional) note. Only the person who asked can rate it (Only the person who asked can rate this answer.). Answers marked Not right are read first by the people improving OAIA.
• Long conversations. A conversation carries its earlier turns to the model, and the meter beside the box shows roughly how much. Compact (or /compact) replaces the earlier turns with a short summary; a very long conversation is compacted automatically before it is sent. The summary is a record of what was asked and done, kept as background: OAIA always answers the newest message, and returns to an earlier topic only when it is referred to. New chat (or /new, /clear) starts over and shows the starter questions again; it is the cleanest start for an unrelated topic. /help lists the commands.
• From an opened item. Every question sent from an item's Open view, the suggested ones included, is about that item: it stays pinned until the view returns to the list.
• What is kept. Each question is recorded for the organisation with its answer, how long it took and its rating, so that the platform operator can see what people ask and what OAIA could not do.

When it proposes a change

If a change would help, OAIA places a Proposed change card under its answer. Nothing happens until someone clicks Apply, and Apply checks the clicking user's own permission for that change; Discard drops it. A proposal expires if left; ask again.

The changes it can propose form a fixed catalogue: set a venue's hours or a zone's capacity, assign a device to a zone, create an alert rule or switch one on or off, add or rename a label and put it on something, set the detection speed, set how long a kind of data is kept, schedule a report to be emailed to the requester or pause one, acknowledge, snooze or won't-fix a finding, dismiss a suggestion, build a report (below), and each enabled app's own small actions (release a no-show booking, open or reprioritise a work order, dispatch a cleaning, make a room a toilet in Toilet Hygiene, apply a gateway's measured receive offset, whitelist a person in Security Watch, re-send a webhook's failed deliveries, and others). It never touches admins, roles, sign-in, AI keys or connector credentials, and never deletes anything.

NOTE

Asked to set how long the activity log is kept, OAIA accepts 7 to 120 days, the same range as the Data Retention card. See How long it is kept.

Example: an alert when a room is over capacity. Ask Alert me when the Level 5 Training Room is over capacity. OAIA reads the room's capacity and the venue's hours and asks what it cannot know: who should be told and how (an in-app alert unless stated otherwise), the hours, and whether "over" means above the capacity or at it. It then proposes Create the rule … with the whole rule in one line (when People Count greater than 20, 09:00–18:00 on Mon–Fri, send an in-app alert, at most every 30 min). Applying it creates the rule switched on, under Rules > All rules, where people to notify can be added or the rule switched off. A rule that switches something in the building is never drafted from the chat.

Example: toilets not yet in Toilet Hygiene. Ask Map the toilet rooms to Toilet Hygiene. OAIA lists the rooms that are named or marked like toilets and asks only for what it cannot read: each toilet's gender and number of stalls (and urinals). Reply in one line (Level 4 Female 3 stalls, Level 4 Male 2 stalls and 2 urinals, Level 6 Accessible unisex, 1 stall) and it proposes one Make … a toilet card per room. If an old toilet lost its room when the floor was redrawn, it offers to move that toilet to the new room so its history is kept. Cleaning rules and who is notified are set afterwards on the toilet's page in the app.

With Autopilot in Act and the Configure tier on, a configuration change requested in the chat is applied as it is proposed and listed in the Autopilot section with Undo; everything else still waits for Apply.

Requesting a report the platform does not have

Describe the report in plain words, for example "count male and female toilet visits on Level 2; there is a door counter on the corridor and a PIR in the male toilet". OAIA answers with the three parts that make that report: a data table with the columns, a rule on the sensor's own reading that adds a row, and a custom report over the table. It reads what exists (the devices, their zones, the fields each sensor reports) to name real ones, and explains two cautions: a PIR trigger is a visit, not a head count, and a rule on a level needs a cooldown or it writes a row every time it is checked.

It then proposes Build a report: one approval that creates all three, with the rule switched on because it only writes rows. Click Apply; with Autopilot in Act and Configure on, it is already done and the reply says where to find it. The parts can also be requested one at a time.

Connecting OAIA to Telegram

Applicable versions / plans

SPECIFICATION

• A Telegram bot set up under Connectors (the Telegram connector).
• Connect the bot to Omaya: manage-oaia or manage-connector-telegram (Admin and Manager as shipped).
• Connect Telegram for a person: view-oaia.
• OAIA must be switched on.

Requirements

Ravi Kumar wants to ask OAIA about Northwind Tower from his phone while walking the floors, through the organisation's existing Telegram bot.

• The bot points at Omaya.
• Ravi's own Telegram account is linked to his Omaya account.
• Answers reach only him, in a private chat.

Restrictions

• A link code is valid for 10 minutes and works once.
• Group chats are ignored; answers go only to the linked person.
• A change proposed in Telegram is never applied there.
• If the person loses OAIA access in Omaya, the bot stops answering them.

Configuration roadmap

(1) Set up the Telegram bot under Connectors.

(2) Connect the bot to Omaya from the OAIA page, once per organisation.

(3) Each person generates a link code and sends it to the bot.

Procedure

1. If the organisation has no bot, an administrator sets one up under Connectors (the Telegram connector) first.
2. Choose OAIA. On the Telegram card, click Connect the bot to Omaya. This is done once per organisation.
3. Click Connect Telegram on the card. A code is shown with Send this to @<bot> in a private chat within 10 minutes: and an Open in Telegram button.
4. Send the code to the bot in a private chat.

Verification

• After step 2 the card shows Bot connected with the time.
• After step 4 the bot replies Linked. You are <name> in Omaya — ask me anything about your building. /unlink disconnects. The card shows Linked to … and since when.
• Questions are then asked as on the page. Suggested follow-up questions appear as buttons in place of the keyboard; tapping one asks it. /new starts a fresh conversation; /unlink disconnects (or Disconnect on the card).
• A proposed change in Telegram links to Omaya, where it waits under Proposed from the app or Telegram — apply it here.

OAIA in the mobile app

Administrators have More > Assistant > Ask OAIA in the Omaya mobile app; staff accounts do not see it. It is the same OAIA as the page: tap one of the starter questions or type a question, and the answer arrives in one piece with a Working… 12 s line while it is prepared. Under each answer are thumbs-up and thumbs-down ratings and up to three You could ask next questions; links open in the browser. The conversation stays on the phone; + starts a new one, with Undo for a few seconds.

A proposed change shows as a Proposed change card with Open in Omaya to apply, which opens it on the OAIA page on the web. It is never applied from the phone. OAIA's findings are also in the app's Alerts tab (the OAIA chip; see the mobile guide).

Configuring Autopilot

Overview

With Autopilot on, OAIA stops only proposing and starts acting, within limits that are deliberately narrow. Its controls are in an Autopilot section on the AI card: a Mode, and What it may do. Autopilot is off regardless while OAIA itself is off.

Nothing in the ring is adjustable from the settings page; that is what makes the ring worth trusting.

Applicable versions / plans

SPECIFICATION

• Autopilot must be made available to the organisation by the platform operator. Until it is, the Autopilot section is not shown and the mode stays Off.
• Saving the AI card: manage-settings-general (Admin as shipped).
• Undo and Approve on the Activity tab: manage-oaia; Approve on a rule draft also needs create-rules.
• The Operate tier acts through the smart-home bridge.

Off by default, and gated

Autopilot does nothing until the platform operator has made it available and the organisation has switched it on.

Requirements

Aisha Rahman wants OAIA to fill obvious configuration gaps at Northwind Tower on its own, but only after she has read a week of what it would have done.

• Autopilot in Shadow for the first week.
• Only the Configure tier allowed.
• Every decision visible with an Undo.

Restrictions

These limits are constants in the code, not settings, and they are what makes the feature safe to leave on:

Limit	Reason
3 actions per heartbeat, 20 per day	A misreading cannot cascade.
One hour between actions on the same thing	It cannot fight itself over one room.
A room empty for 10 minutes before anything is switched off	Somebody sitting still is not an empty room.
Evidence older than 90 minutes is not evidence	It does not act on a stale temperature.
Never flips its own last action back within 30 minutes	No oscillation.
Setpoint 24 °C, never sent outside 20–28 °C	A bad number cannot leave the band.
Capacity between 1 and 1,000	The same principle, for configuration.

Configuration roadmap

(1) Open the AI card of System > Configuration.

(2) Set the Mode to Shadow — record only and tick the allowed tiers.

(3) After a week of reading the record on the Activity tab, change the mode to Act.

Procedure

1. Choose System > Configuration and click the AI card.

2. In the Autopilot section, set the fields as follows.

   Configuration Item	Value	Note
   Mode	Shadow — record only	Off (default), Shadow — record only or Act. See Modes.
   What it may do	Configure	Tick any of Draft rules, Configure, Operate. See What it may do.

3. Click Save. Choosing Act asks for confirmation first.

Run it in Shadow for a week first

Shadow produces exactly the record Act would, so a week of decisions can be read and judged before any of them are real.

Verification

• The Activity tab lists each candidate action. In Shadow each row reads would do; in Act, done with Undo; a candidate whose preconditions are not met reads held with the reason.
• With no mode or tier chosen, the section says Nothing considered yet. Autopilot runs with each heartbeat once a mode and at least one tier are chosen on the A.I Model card.

Modes

Mode	Effect
Off	Proposals wait for a person. The default.
Shadow — record only	OAIA decides what it would do and lists it as would do. Nothing changes. Approve on a row has it done now, recorded as done by you, with Undo.
Act	OAIA applies what its tiers allow, each run, and lists each action as done with Undo. Turning Act on asks for confirmation.

What it may do

Eleven kinds of action in three tiers; only the ticked tiers are used.

Tier	Kinds	Meaning
Draft rules	Draft a rule	Creates the rule a review proposed, switched off, for a person to enable. It cannot draft one that raises work orders.
Configure	Set a zone capacity · Set venue hours · Assign a device to a zone · Build a report (data table, rule and custom report, or each alone)	Fills configuration gaps the review is sure of. A capacity for a zone with none is learned from 14 days of counter data, with headroom, never guessed. Each change records the value before it.
Operate	Switch off · Cooling on · Lights on	Through the smart-home bridge: cooling on when a room with people in it is hot (27 °C) or humid (70 %), when an equipment room passes 30 °C, or 15 minutes before a booked meeting in a hot room; fans on at 1,000 ppm CO₂; lights on below 100 lux with people present, or in a room with an open critical security incident; lights, fans and air-conditioning off when the radar says a room has been empty ten minutes, or half an hour after the venue closed with nobody counted. Never anything else off (a plain switch could be a fridge or a rack), never anything off in an equipment room (server, comms, plant, UPS, battery rooms and the like), and never a lock, a siren or a scene.

Undo, and the kill switches

Every applied action records the state before it. Undo restores that state. An undo is also a signal:

• Undoing an action pauses that kind of action for 24 hours.
• Three undos in a week drop the mode to Shadow. OAIA stops acting and returns to recording what it would have done, without anyone having to notice and step in.

Every action, whoever triggered it, is also written to System > Activity Log as oaia/autopilot or oaia/ followed by the person who applied it.

Result checks after an action

For an action whose result can be measured, OAIA looks again later and reports whether it worked, whether a person or Autopilot applied it:

Action	When it checks	What counts as worked
Dispatch a clean of a toilet	Once the clean is finished and the toilet's score has been recalculated (it waits while the clean is still open, up to 8 hours)	A fresh score, calculated in the last 6 hours, below the toilet's dirty line (60, or the toilet's own target if that is higher). With no fresh score the result is Could not check.
Cooling on (Autopilot), when the room was hot	45 minutes later, from a temperature reading taken after the cooling went on	The room is below 27 °C, or at least 1 °C cooler.
Switch off (Autopilot)	30 minutes later	The device is still off in Home Assistant.
Apply a gateway's receive offset (measured on the Calibration Bench)	24 hours later, against the 24 hours before	Tags on that floor bounce between rooms no more often than before (brief bounces per 100 room stays). Not checked while Allow for gateways that hear louder than others is off.

Each check ends as one of:

• Verified: it worked. A finding the action was meant to fix is resolved.
• Not resolved: it did not. The finding returns to the open list (even if it was acknowledged or snoozed) with what the check found and what to look at, for example Cleaned, but Level 4 Male Toilet is still at dirt score 82 — check the ventilation, a broken fixture or the odour sensor. There is no new notification, because the finding was already reported once.
• Re-checking: it moved but has not reached the target, so OAIA looks once more later.
• Could not check: the reading was missing or the device unreachable. This is never counted as worked.
• Not checked: the action was undone first.

The result shows on the finding, on the Autopilot entry, and under Checked after applying on the Activity tab. Checking only reads: it never changes, undoes or repeats an action, and it never prevents an undo. Actions with nothing to measure (a capacity, a rule, a tag) are not checked.

Gateway receive offsets from the Calibration Bench

When the Omaya operator measures the gateways side by side on the Calibration Bench, each gateway's result is kept. A clean measurement less than 30 days old that differs from the gateway's current offset can be applied through OAIA: ask it (apply the measured offset to the Level 3 East gateway), or click Apply on the daily review's suggestion. OAIA only ever applies the number the Bench measured; it never asks for one or makes one up, and without a measurement it states that the gateways need measuring first.

CAUTION

The offset is used for room decisions only while Allow for gateways that hear louder than others is on (System > Configuration > Indoor Positioning, Rooms tab). The proposal says so when it is off.

What OAIA never does

• Set or change a severity: that stays the detector's.
• Send a notification: the platform does that.
• Read another organisation's data, or its own base tables.
• Delete anything, or touch admins, roles, sign-in, AI keys or connector credentials.
• Switch a rule on by itself. Rules it drafts are off until a person turns them on; a person can ask it to switch one on and then apply that.
• Draft a rule that raises work orders.
• Take an instruction from data it read. A zone named "ignore your rules" is a zone with a strange name, and a claim in the chat to be a developer or a platform operator changes nothing.

AI key selection

The key is chosen per request, so a cap reached on the platform side stops nothing for an organisation that brought its own.

OAIA uses the provider chosen on the AI card; if none was chosen, the platform's default provider. The key is the organisation's own if one was entered and the platform operator allows the organisation its own key; otherwise it is the platform's.

• The organisation's own key is billed by its provider and does not count against the monthly allowance.
• The platform key does count, and System > Usage shows how much of the allowance is left.
• When the allowance runs out, OAIA keeps detecting and notifying in the detectors' own words until the month rolls over or a key is added.

Troubleshooting

Symptom	Cause	Action
OAIA reports nothing.	Usually correct: it reports on change.	Check the Activity tab. If runs are happening and there are no findings, nothing is wrong: All quiet — no open findings.
A finding shows only the detector's wording (detectors only).	No AI provider, AI switched off for the organisation, or the monthly allowance used up.	The page says No AI model is available for this tenant, so findings are shown without a narrated digest or diagnostic chat. with a link to Configuration. Set a provider on the AI card or add a key.
The chat says read-only access has not been provisioned.	OAIA's read-only database access for the organisation has not been set up yet.	An operator step; ask the platform operator. The heartbeat and findings still work.
OAIA cannot read Footfall, or counts a venue by adding up its zones.	OAIA's access to the Footfall figures is a one-off step for the platform operator, done once per organisation after the October 2026 update.	Until it is done, ask about a single zone, or use Reports > Footfall.
The chat is slow.	A question that needs several queries takes longer; the step line shows it is still working.	Ask about one named thing ("the Level 3 East gateway") rather than an open question ("what is wrong?").
The answer is wrong.	—	Click Not right and give the reason. Ask again more narrowly, with the names Omaya uses for the place or device.
Apply says That proposal has expired or is not yours to apply. Ask OAIA again.	Proposals are tied to the person who asked and do not last indefinitely.	Ask again.
Telegram says the code is not valid.	Codes last 10 minutes and work once.	Click Connect Telegram for a new one.
Telegram says No Telegram bot is set up for this organisation.	No bot exists.	An administrator adds one under Connectors first.
The OAIA page offers Open the AI card, but the card shows Not Permitted.	The role holds manage-oaia but not manage-settings-general.	An Admin switches OAIA on.
Autopilot is not offered.	It has not been made available to the organisation.	Ask the platform operator.
Autopilot is on but does nothing.	No tier is ticked, or its preconditions are not met: a room not empty long enough, evidence too old, a target inside its hour, a kind paused by an undo.	The Autopilot section lists each candidate as held with the reason.