← All products GravOS AI Autopilot · The intelligence layer

It doesn’t just watch your operation. It learns to run it.

AI Autopilot isn’t a separate product you buy — it’s the intelligence inside GravOS Charge, Electron and Edge. It observes how your sites actually behave, proves itself against your team’s decisions, and then takes over the work you’d rather not do by hand.

AI Autopilot — Control & Autonomy LIVE
Portfolio · 12 sites · Charge + Electron + Edge
Decisions automated94%6% escalated to a human
Recommendation accept rate91%trailing 90 days
Operator hours saved~38 / wkvs manual operation
Guardrail breaches0since deployment
Autonomy by decision type you set the level
Charger load balancingManage
97%
Peak-demand shavingManage
95%
Fleet departure sequencingManage
93%
Battery arbitrageManage
89%
Market & DR biddingRecommend
82%
V2G export windowsRecommend
76%
Session pricing changesObserve
61%
Proactive — before it becomes a problem predictive
  • Charger CP-07 likely to fault in ~6 daysConnector resistance drifting and 3 failed handshakes. Service ticket drafted — approve to dispatch.
  • Peak forming Thursday 17:10Forecast exceeds threshold by 140 kW. Storage pre-charged overnight; no action needed.
  • DR event probability 68% tonightHolding 18% extra SoC. Estimated value of reserving: +$310.
  • Bus 014 will miss target SoCRe-sequenced ahead of two lower-priority vehicles. Departure protected.
Autonomy Manage — acts within guardrails Recommend — proposes, you approve Observe — learning, no action

One console for how much autonomy you’ve granted, per decision type — and the confidence Autopilot has earned for each.

One brain, every product

Autopilot isn’t a module. It’s the layer underneath.

The same intelligence runs inside every GravOS product — which means it sees your chargers, your batteries and your sites as one operation instead of three disconnected tools.

Charge

Inside GravOS Charge

  • Balances load across stalls in real time
  • Sequences fleet charging to departure times
  • Predicts charger faults before drivers hit them
  • Flags pricing and utilization opportunities by site
Electron

Inside GravOS Electron

  • Forecasts load, generation and price
  • Stacks value streams without cannibalising them
  • Weighs every cycle against degradation cost
  • Bids and dispatches into programs and markets
Edge

Inside GravOS Edge

  • Runs the models on site, in under a second
  • Keeps deciding when the network is down
  • Enforces safety limits in deterministic code
  • Reconciles learning back to the cloud on reconnect

Why one brain beats three tools

Intelligence compounds when it isn’t siloed. What Autopilot learns in one product immediately improves the others.

Charge → ElectronCharging load shapes teach the optimizer exactly when your site peak forms, so storage is positioned before it lands.
Electron → ChargeTariff and market behaviour informs when charging is cheapest, so sessions are scheduled against real cost.
Edge → EverythingLocal fault signatures and asset quirks from one site improve prediction across the whole fleet.
Fleet → V2GDeparture patterns reveal genuinely idle windows, so export never threatens readiness.
The learning journey

Earning autonomy, not assuming it

Autopilot doesn’t arrive claiming to know your operation. It watches, proposes, gets scored against reality, and only takes the wheel once it has proven itself — typically within 90–120 days of operational data.

Day 0 – 30 Observe

Learn what normal looks like

Connects to your assets and builds a baseline: load shapes, session patterns, tariffs, asset behaviour, how your team actually operates. It takes no action.

You get: visibility, benchmarks and a measured baseline to judge everything else against.
Day 30 – 90 Shadow

Predict, then check its own work

Autopilot forecasts and decides in the background, then compares its choices to what actually happened and what your operators did. Every miss is a training signal.

You get: an accuracy record — evidence, not promises.
Day 90 – 120 Recommend

Propose actions worth taking

Once accuracy clears your threshold per decision type, Autopilot starts recommending — with the reasoning and expected value attached. Your team approves or rejects.

You get: fewer decisions to make from scratch, and a record of what you accepted.
Day 120 + Manage

Run it autonomously

You promote the decision types you trust to Manage. Autopilot executes continuously inside your guardrails and escalates only genuine exceptions.

You get: autonomous operations — your team sets policy instead of clicking buttons.

Timelines vary with data density, seasonality and site complexity. A high-throughput charging network reaches confidence faster than a site with a handful of monthly peak events.

Two modes

You decide how much to hand over

Autonomy isn’t all-or-nothing, and it isn’t permanent. You choose the mode per decision type, and you can change it at any time.

Recommend mode
  • Autopilot proposes the action and the expected value
  • Your operator approves, edits or rejects
  • Every decision carries its reasoning
  • Ideal for market bidding, pricing and anything commercially sensitive
  • Ideal while your team is building trust
Manage mode
  • Autopilot acts continuously inside your guardrails
  • Humans set policy and review outcomes, not individual actions
  • Only genuine exceptions escalate to a person
  • Ideal for load balancing, peak shaving and departure sequencing
  • Revert to Recommend at any time, instantly

Mixed by design

01

Per decision type

Run load balancing autonomously while market bidding still needs a human. One operation, different levels of trust.

02

Per site

Promote your best-understood sites to Manage first, and keep a new or unusual site in Recommend until it settles.

03

Per threshold

Autonomous below a value or risk threshold, human approval above it. You define where the line sits.

04

Always reversible

Drop any decision type back to Recommend or Observe instantly. Autonomy is a setting, not a commitment.

Proactive, not reactive

Most systems tell you what already went wrong

An alert after a charger failed, a peak after it set your bill, a missed event after the window closed. Autopilot is built to act before those become facts.

Predicts failures

Drifting connector resistance, repeated handshake errors, comms degradation — surfaced days before the hardware actually fails a customer.

Pre-positions for peaks

Sees the peak forming hours ahead and charges storage in advance, instead of reacting once the threshold is already breached.

Reserves for opportunity

When a demand-response event looks likely, it holds capacity back — trading small certain value for larger expected value.

Protects commitments

Detects that a vehicle won’t reach target charge in time and re-sequences the depot before anyone notices a problem.

Spots quiet degradation

Utilization drifting at one site, a battery aging faster than its peers — patterns too slow for a human to catch in a dashboard.

Escalates properly

When something genuinely needs a decision, it arrives with context and a recommendation — not a raw alarm at 3am.

Trust & control

Autonomy with a seatbelt

Handing operations to software is a serious decision. Autopilot is built so that decision stays reversible, auditable and bounded.

How we keep it safe

The AI never gets the final word on safety

Intelligence and authority are deliberately separated. Autopilot decides what should happen; deterministic, safety-coded logic decides what is allowed to happen.

  • Hard constraints — warranty limits, reserve floors, readiness thresholds and electrical limits are enforced in code, not learned behaviour.
  • Explainability by default — every action carries its reasoning and expected value, so an operator can audit any decision.
  • Full audit trail — what was decided, why, what it delivered, and who approved it if anyone did.
  • Instant revert — drop to Recommend or Observe at any moment, per decision type or across the portfolio.
  • Fail safe — on sensor fault, comms loss or model uncertainty, the system falls back to safe deterministic behaviour.

Autopilot has to earn every increment of autonomy — and you can take it back in one click.

ObserveLearns · takes no action
ShadowDecides · scored, not applied
RecommendProposes · you approve
ManageActs · inside guardrails
Move up when it earns it. Move back the moment you want to.
Questions we get

Autonomy, data and control

Do we have to let the AI control anything?

No. Autopilot is useful from day one in Observe and Recommend modes — forecasting, benchmarking and proposing actions. Manage mode is optional, granted per decision type, and reversible at any time.

Why 90–120 days?

That's typically how long it takes to see enough operational variety — weekday and weekend patterns, weather swings, tariff periods, real peak events — for confidence to be meaningful rather than a small-sample illusion. Data-dense operations get there faster; sites with rare events take longer.

What happens in the meantime?

You get the platform value immediately: unified visibility, rule-based automation, alerts and reporting. Autopilot's learning runs alongside that from the first day of data.

Does it need our data to leave the site?

Edge runs models locally, so operational control doesn't depend on sending data anywhere. Cloud learning is used for cross-site improvement and can be scoped to your tenant.

How do we know it's actually working?

Every decision is logged with its reasoning and expected value, and measured against the baseline established in the first phase. Accept rate, forecast accuracy and realized value are all reported.

What if it makes a bad call in Manage mode?

It can't breach a hard constraint — those are enforced outside the model. For anything softer, exceptions are logged and escalated, thresholds can be tightened, and the decision type can be dropped back to Recommend immediately.

Does Autopilot cost extra?

It's the intelligence layer inside GravOS Charge, Electron and Edge, not a separate SKU you bolt on. Talk to us about how it's packaged for your deployment.

Start in Observe. Graduate to autonomous.

Connect your assets, let Autopilot learn your operation, and hand over only what it earns. No leap of faith required.