Wastehero
WasteOS · Autopilot
The AI-native way to run tickets

AI suggests. You decide. It proves it.

Autopilot is the AI inside WasteOS that watches your ticket stream, learns which routine work is worth automating, and prices every proposal from your own history. Say yes once — and it keeps a ledger of promise next to measured outcome. Never an estimate dressed as a result.

wasteos · Autopilot · Pulse
Hours given back · this month
862.4 h
Needs your approval
3
Watching
14 flows
Handled zero-touch · this month
8,921
08:14Extra collection · 41 Harbour RoadCreatedHandled · zero-touch
08:14Food waste bags delivered · 12 Mill LaneCreatedHandled · zero-touch
08:15Textile bag dropped off · 3 Station SquareCreatedHandled · zero-touch
08:15Invoice question · DKK 240 waive requestedCreatedPaused · waiting for you
Faithful recreation of the live Pulse screen, running on a demonstration tenant. The fourth ticket touches money — so the AI waits for a human. That is the point.
0
real tickets across customers — the history Autopilot mines for patterns
0
historical events replayed to price one parked automation — before it ever ran
0
automated tests on the engine that executes, pauses, and audits every run
01 · Pulse

One glance, every morning.

Pulse opens on four honest numbers: hours measurably given back, what needs your approval, what the AI is watching, and the top things still worth automating — each priced from your own history, not a benchmark.

Priced, not pitched. Every suggestion carries the evidence line it was mined from and the sample it was priced on.
Built means gone. The moment you build a suggestion, it stops pitching itself. It moves to your automations — with the trail kept.
Zero is a real number. If nothing qualifies, the feed is honestly empty. Nothing is fabricated to fill it.
Pulse · Top things to automate
Auto-handle “Food waste bags delivered” tickets
3,127 completed tickets since March · 96% resolved the same way
551.3 h/moBuild
Auto-handle “Textile bag dropped off” tickets
940 completed tickets · uniform quick work
165.0 h/moBuild
Auto-handle “Hazardous-waste box collected” tickets
871 completed tickets · uniform quick work
153.3 h/moBuild
Suggestions from the demonstration tenant — mined and priced from its own completed tickets.
Builder · Autopilot’s proposal
Auto-handle “Food waste bags delivered” tickets
Grounded in your own rule for this ticket type — the AI proposes, your history decides.
WhenA ticket of this type is created
DoAssign to the team that already handles it · from your rule
DoClose it out (set status to Completed) · the step its 551 h/mo is priced on
See what it would have doneTurn it on
The draft is rebuilt server-side from your rules every time — a client can’t sneak an edited version past review.
02 · The mining AI

It doesn’t guess. It mines.

Suggestions come from deterministic pattern mining over your completed tickets — routine volume, response-time outliers. The numbers are computed; the AI phrases the sentence, and the screen says which one wrote it.

Grounded drafts. A suggestion becomes a flow built from the rules you already trust for that ticket type — reviewed, never invented.
The price includes the work. If a suggestion is priced on closing tickets out, the draft actually closes them out. No pitch-versus-product gap.
Sample floors. Thin history prices nothing. A number the AI can’t defend is a number it won’t show.
03 · Backtest

See what it would have done. Then decide.

Before a flow touches anything, Autopilot replays it against your real recent history: which tickets it would have fired on, what each step would have changed, and what that was worth. A parked draft on the demonstration tenant priced out at 256 hours per 30 days — over 10,071 real events.

Fleet check. One click simulates a flow together with everything already live and lists real collisions — it caught two flows firing on the same 2,198 events before they ever double-handled a ticket.
Checked zero ≠ unmeasured. “0 fired” comes with coverage: whether that’s a verified quiet period or simply data the backtest can’t see.
Dry by construction. A backtest never sends, routes, or changes anything. It reads and predicts. That’s all.
Flows · Check against the fleet
Would have firedTogether withOverlapVerdict
2,198 eventsBin delivery (live)2,198 sharedCollision
2,198 eventsConfirm extra pickup (live)0 sharedClear
Escalate missed collection (draft)different triggerNot compared
A real collision, found in preview — two automations claiming the same tickets, surfaced before either double-fired.
Ask Hero · the what-if is parsed by AI; the numbers come from the same deterministic backtest engine as everything else.
04 · Ask Hero

Ask in plain language. Get a backtest, not a vibe.

Type a what-if the way you’d say it in the morning meeting. Hero replays it against your real history and answers with what it would have done — fired-on counts, hours, collisions — then offers to arm it through the same reviewed builder path as everything else.

AI parses, the engine proves. The language model never invents a number — every figure in the answer is a deterministic replay of your own events.
Arming is never a shortcut. “Arm it” drops into the normal review → test → approve path. No side door to live.
Honest when it can’t. A question the engine can’t backtest gets “I can’t measure that yet” — not a made-up answer.
Activity · one run, step by step
Ticket created — Invoice question
trigger matched · run started
Assign to Customer Service
executed · verified on the ticket afterwards
Waive service fee — DKK 240
money step · run paused, waiting for a named human
Approved by Hussam — “credit note already agreed”
resumed from the approved step · earlier steps never re-run
Confirmation sent · run completed
outcome re-read from the ticket — not assumed from the log
Money and mass-communication steps always pause for approval. Everything else is one click to switch off — with the reason on record.
05 · Guardrails

Live — with a leash.

Autopilot executes through the same domain code your team uses by hand, and treats “live” as a privilege it has to keep earning. AI suggests, you decide — and the system checks its own work.

Approvals that actually pause. A money or mass-comms step freezes the run mid-flight and waits. Approve, and it resumes exactly there — nothing re-fires.
Kill switch, with a reason. Any flow, off in one click; who and why go in the audit log. A circuit breaker demotes a misbehaving flow on its own.
Trust, then verify. After every action, Autopilot re-reads the ticket to confirm the change actually landed — and says “held” when it didn’t, never a hopeful green.
06 · Impact Ledger

The promise is frozen. So is the truth.

The moment a flow goes live, its promised saving is locked into the Impact Ledger. From then on the page shows the promise next to what was actually measured — and neither is ever edited. Misses show up as plainly as wins.

Measured means measured. Hours come from zero-touch completions your records can prove. A run a human had to approve is never priced as zero-touch.
Unavailable beats estimate. Deflected calls, fuel from smarter routing, service-level gains — real impact, not counted yet, and the page says so in plain words.
Roads not taken, priced. Eligible automations you haven’t switched on are priced from their own backtests — the value still on the table, in the ledger’s own unit.
Impact · promised next to measured
AutomationWent livePromisedMeasuredZero-touch
Auto-handle food-waste bagsProven in simulation551.3 h/mo86.2 h1,034
Confirm extra pickupManual overrideno forecast on record12.4 h149
Escalate missed collectionProven in simulation31.0 h/monot measured yet0
An override flow honestly has no forecast — the ledger says so instead of inventing one.
Automations · Rules · your existing estate
RuleRuns today onConvertibility
Food waste bags → assign + closeClassic engineSafe · 1:1Convert to draft
Extra collection → route + notifyClassic engineNeeds reviewPreview
RFID block-untilClassic engineStays classichonestly unsupported
Converting never edits or deletes the original. Activating the Autopilot version disables the source rule in the same step — nothing ever fires twice.
07 · Zero-loss migration

Your rules come along. Nothing fires twice.

Autopilot sits on top of the rules engine you already run — one product, one vocabulary. Every legacy rule is visible, with an honest label: converts faithfully, needs review, or stays on the classic engine.

Non-destructive by rule. Conversion creates a draft; your rule keeps running byte-for-byte until a human reviews, tests, and switches over.
No double execution. Going live and retiring the source rule are one atomic step — a correctness gate with its own tests, not a nicety.
Honest coverage. A rule Autopilot can’t reproduce yet says so — it is never “converted” into something quietly different.
08 · Capabilities

What it can do today

Thirty-three registered actions — each executing through the same domain code your team uses by hand, each labelled honestly in the builder. If an action isn’t connected yet, the palette says “Not connected yet” instead of pretending.

Handle the ticket

Set status and priority, assign the right person or team, add the internal note, create the follow-up ticket, link related cases.

Talk to the citizen

Send the confirmation or the apology — templated or custom — through the same messaging pipeline your team already uses.

Fix it on the route

Re-dispatch a missed or extra emptying into the next available route, honouring your route schemes and areas.

Touch money — carefully

Waive a service fee on a grounded ticket. Always approval-gated, always audited, never an amount the AI computed itself.

For municipalities

The municipal waste-scheme world is the home turf: food-waste bags, extra collections, hazardous-waste boxes — the routine 80% handled the moment they arrive, the sensitive 20% paused for a named human. Fully bilingual: every screen ships in English and Danish.

For private operators

Same engine on your commercial estate: confirmation flows, missed-collection escalation, fee decisions with an approval trail your auditors will actually enjoy reading.

How it keeps itself honest

Honesty is the architecture, not the tagline.

Every number Autopilot shows traces to your own records: suggestions are priced from completed tickets, backtests replay real events, and the ledger only counts what it can prove. The AI phrases; the engine computes; the screen says which is which. Where the truth is “not measured yet” or “no history for this type”, that is exactly what it says.

Bring your noisiest ticket queue.

The one that eats your mornings. Watch Autopilot mine it, price what it finds, and show you — before anything goes live — exactly what it would have done last month.