A rate you cannot interrogate is a quote, not a price.
Insurance pricing is conventionally opaque, and the reason usually given — that the model is proprietary — is mostly true and entirely beside the point.
A buyer who cannot see why their premium is what it is has no way to lower it except by shopping. That is fine for the insurer and useless for the customer, and in this category it is actively harmful: the whole argument for insuring autonomous systems from telemetry is that the operator can change the number by changing the engineering. That argument only works if the operator can see which engineering.
So this document is the factor list. It is not the model — the elasticities are fitted and they change — but it is every input the model reads, with an honest label on how much we actually know about each one.
What the premium is charged on
The exposure base is the most consequential choice in any line, and it is usually made by accident. Get it wrong and the premium stops tracking the risk: charge on revenue and a profitable operator subsidises a reckless one; charge on headcount and automation looks like a discount on danger.
For embodied systems we charge on robot-hours — units multiplied by hours actually operating. It is the only base that moves the way the risk moves. A fleet parked for a quarter did not expose anyone that quarter and should not pay as though it did; a fleet running three shifts exposed three times as much. For software systems the equivalent is inference volume that reaches a person or takes an action, which is why internal evaluations and test traffic are excluded — they cannot hurt anybody.
For systems that act rather than advise, volume alone is the wrong measure and we say so plainly: a million advisory calls and a million executed payments are not the same risk at any price. Those are rated on maximum uncontrolled loss — action rate multiplied by value, by how long detection takes, by how little of it can be undone. The derivation is on the Intelligence page.
Every factor
Effects shown are directional multipliers on the base rate for a typical risk, not the fitted coefficients. They compound, and credits are floored — controls cannot take a risk below 55% of its base rate, because at some point the remaining loss is not something engineering can remove.
From data we hold. Telemetry we collect directly, or a quantity the operator's own systems report.
Inferred from an adjacent line. Industrial machinery, motor fleet, product liability, cyber — established experience that should transfer, and might not.
Engineering opinion. We believe this drives loss. We do not yet have the data to prove it, and some of these will turn out to be wrong.
Count the labels. On a document like this the ratio of Measured to Judgment is the honest measure of how much an insurer actually knows. Ours is currently an even split, which is better than it was and worse than it should be: a third of what sets your price is still an opinion we have not yet earned the right to hold. It improves in one way only, which is the next two sections.
What we don't know
We do not have a credible loss history for embodied AI, and neither does anyone else.
Credibility in the actuarial sense is a specific thing: enough observed claims that the observed frequency is more trustworthy than your prior. For humanoid fleet liability nobody has it, because the fleets are young and the claims have not happened yet. Any rate quoted for this class today — ours included — is mostly prior and a little bit data.
What we have instead is the adjacent experience marked Borrowed above: industrial machinery liability, motor fleet, product liability, and cyber. Those are real books with real history, and the assumption that a humanoid behaves like a blend of them is a reasonable starting point and an assumption.
The honest position for an insurer at this stage is not to claim precision. It is to say which parts of the rate are load-bearing guesses, price conservatively where the guess is doing the work, and replace guesses with observations as fast as the exposure allows.
Three specific things we expect to have wrong. Fleet-size discount: we credit larger fleets for process maturity, and it is possible the relationship inverts above some size as fleets standardise on a single failure mode. Teleoperation: we treat remote human control as risk-reducing, and the handover moment may turn out to be the most dangerous part of the operation rather than the safest. Update cadence: we load frequent releases, but frequent small releases may be materially safer than rare large ones — the software industry's own evidence points the other way from ours.
How the rate changes as data arrives
Every rate here is a blend of what we assumed and what we have since observed, and the weight shifts toward observation as observations accumulate. This is standard credibility weighting, stated openly because the alternative — a rate that changes without explanation at renewal — is what makes buyers distrust insurers.
Z near zero
Your price is the class prior. You are being charged for what we believe about machines like yours, because that is all anyone has.
Z rising
Your own record starts to matter. A clean operator separates from the class before a full loss history exists.
Z at one
You are priced on your own experience. The class prior stops mattering to you and becomes something you contribute to.
The practical consequence, and the reason this is worth publishing: an operator who instruments early reaches credibility earlier, and reaching credibility earlier is worth more than any control credit on the list. The fastest way to lower an autonomy premium is to make the risk legible sooner than your competitors do.
If you have to set a limit
More than thirty US states now have personal-delivery-device laws, and those that attach an insurance requirement have largely settled on $100,000 in general liability — frequently for the combined operations of an entire fleet. That number was not derived from loss experience. It could not have been; there was none when the bills were drafted.
We think that is a reasonable thing for a legislature to have done and a bad number to keep. If you are drafting a requirement — as a state, a municipality, a landlord, or a customer writing a vendor clause — the limit worth asking for is not a round figure but an answer to three questions:
What is the physical worst case?
Mass, speed, and proximity bound the injury. A 10 mph 45 kg device on a sidewalk and a 60 kg humanoid at arm's length are not the same requirement.
Is the limit per event or shared?
A fleet-wide aggregate is exhausted by the first serious claim. If the limit is shared, it is a compliance gesture rather than a remedy.
How many claimants can one failure produce?
A software defect reaches every unit at once. Correlated exposure is the property that separates a robot fleet from a fleet of vans.
We will answer those questions publicly as our data supports it, and this document is where the answers will appear. We would rather the reference number for machines in public space came from published loss experience than from whatever figure survived a committee — including when the published number is inconvenient for us.