Glossary

Human-equivalent hours

A billing unit denominated in the work replaced, not the compute consumed.

Definition

Human-equivalent hours are a unit of measurement that expresses delivered work in terms of how long it would have taken a competent person, rather than how long a machine ran or how many tokens it consumed. The unit exists because machine runtime is meaningless to a buyer comparing an agent against hiring, and because faster hardware should not make the same delivered work cheaper to describe.

Measures: Work delivered, not compute used
Polaris rate: ~$2 per human-hour
Clamp: 5 minutes to 8 hours per job
Estimated from: Observable effort

Why the unit exists

Every other billing unit for AI work is denominated in something the buyer does not care about. Tokens measure text volume. GPU-seconds measure hardware. Seats measure headcount. None of them answer the question a manager is actually asking, which is whether this is cheaper than the alternative way of getting the work done.

Human-equivalent hours answer that question directly, and they have an uncomfortable property that makes them honest: if the model gets faster, the bill does not fall, because the work replaced has not changed. Nobody would design this unit to flatter a vendor.

The Polaris formula, in full

Polaris estimates each job from observable effort. The formula is published so it can be checked.

Base
15 minutes for a delivery job, 10 minutes for a review job. The fixed cost of picking up a piece of work and understanding it.
Research
12 minutes per web search performed during the run.
Writing
Finished prose counted at 90 characters per minute, which is a normal composing-and-editing rate rather than a typing speed.
Checklist and conversation
8 minutes per acceptance-criteria item ticked, 10 minutes per comment posted.
Artefacts
10 minutes per file produced, plus 5 minutes for each generated PDF.
Clamp
The total is bounded to a minimum of 5 minutes and a maximum of 8 hours per job, so no single run can produce an absurd line on the bill.

Commonly confused with

Unit	What it measures	Who it favours
Human-equivalent hours	Work delivered, in human time	Buyer: faster models do not raise the bill
Compute time	Wall-clock machine runtime	Neither, reliably: it tracks hardware, not value
Tokens	Text processed in and out	Vendor: verbose work costs more regardless of usefulness
Full-time equivalent (FTE)	Headcount capacity over a period	Different purpose entirely: workforce planning, not billing a job

The honest limitations

This is an estimate, not a measurement of a real person doing the same task. The counters are objective, but the conversion rates are judgements: twelve minutes per search and ninety characters per minute are defensible averages, not laws. Work heavy on thinking and light on output is undercounted by a formula that counts output.

The mitigation is transparency rather than precision. The formula is published, every job's estimate is itemised on the worker's work log next to the effort it came from, and any line can be challenged from that entry.

Work log

The record that makes an unwatched run reviewable afterwards.

Usage-based pricing

The bill follows consumption, which cuts both ways for buyer and vendor.

Per-seat pricing

The bill tracks headcount, which is only a proxy for value received.

What an AI worker costs, and how the hours are counted

The whole formula is on this page, including the parts that make it an estimate rather than a measurement.

Per-seat pricing and usage pricing charge for different things

One model prices access. The other prices output. Almost every argument about software cost is really about which of those you are buying.

AI worker

The difference between an agent you prompt and an agent you assign work to.

Questions people ask

+Is a human-equivalent hour the same as an hour of machine time?

No, and they are usually far apart. A job that a machine finishes in minutes may represent an hour or more of human-equivalent work, because the unit describes the work delivered rather than the time the hardware was busy.

+How can an estimate be fair if nobody did the work by hand?

It cannot be exact, and Polaris does not claim it is. What makes it checkable is that the inputs are counted rather than asserted, the conversion rates are published, and each job's estimate appears next to the effort that produced it so a customer can recompute and dispute it.

+What does an hour cost?

Polaris bills roughly two dollars per human-equivalent hour delivered, with no charge for the software itself and no charge when nothing is delivered. Unlimited humans, tasks, workstreams and docs are included at no cost.