The going rate for a human minute

An AI agent can price everything it runs on. Tokens are sold by the million, compute by the second, storage by the byte, and every one of these prices is published, moves with demand, and can be checked before the agent commits. The one input it cannot price is a person. The moment a workflow needs a human to check something, to agree on something, or to put a name to something, there is no posted price to consult. What exists instead is rate cards, retainers, or a message to a colleague who may be asleep.

Countersignatory publishes the missing price. On the index today, a screened human check is indicated at $0.80, three independent humans in agreement at $3.50, and the signature of a named, register-verified professional at $25. The same signature, needed within five minutes at two in the morning, their time, is $56. None of these prices come off a rate card, and what follows explains why.

How AdWords worked

Most people remember AdWords as the product that put ads next to search results, but that part was not Google's idea. GoTo.com had been selling search positions since 1998, to the highest bidder, who paid whatever they had bid. When Google redesigned AdWords in February 2002 it kept the auction and altered it in two ways. Ads were ranked by the bid multiplied by a machine's estimate of how often the ad would be clicked, and the winner paid only what was needed to stay ahead of the advertiser below them, rather than the full amount of their bid. Economists call this a second-price auction, and the clickthrough estimate grew into what Google later called the Quality Score.

Those two changes meant that an advertiser nobody had heard of, with five dollars to spend, could be served at a profit. Nobody at Google had to speak to them, check their credit, or judge whether their ad was any good. The machine did all of that, for everyone, all the time, at no extra cost per account. Selling advertising to small businesses had always cost more than they had to spend, which is why nobody did. Once a machine took over that work, the smallest advertiser on earth was worth serving, and the revenue from millions of them paid for a large share of the free web over the following twenty years.

The general lesson from this is usually stated too narrowly, as something about advertising. The broader version is that once a machine can judge quality well enough, a live market becomes safe to open to strangers, and a market that is open to strangers will reach customers that no broker or salesperson could ever track down and serve.

What an agent needs and cannot buy

Software agents now act in the world. They issue refunds, send contracts, place orders and change records about real people. The stacks they are built on give a developer ready-made components for compute, storage, payment and identity, but there is nothing equivalent for accountability. When an agent reaches a step it ought not to take on its own, the usual arrangement is a Slack message to a colleague in the hope that they are awake and paying attention.

What the agent actually needs at that moment is for somebody to have something at stake in the answer, and how much needs to be at stake depends on the decision. For a small one, a single screened person who stakes a quality score on confirming that an invoice is a duplicate may be enough. For a decision where one person could be lazy or bought, three people who have to agree are better. Some decisions need a person with a name and a professional registration, whose regulator can strike them off if they sign something false, and a few need a licence the state recognises, such as a notary's seal.

Those four levels are Countersignatory's four markets, called Check, Consensus, Countersign and Seal. They are not arranged by how much human effort goes into each one. Responders at every level are allowed, and expected, to use AI tools in their own work, since the buyer is paying for someone to stand behind an answer, rather than for the typing. What increases from one market to the next is what the responder stands to lose if the answer is wrong.

Why the price has to move

The conventional way to sell human work to software is a price list. Prolific, which is the best of the human-feedback platforms, charges a flat 42.8 percent platform fee on top of fixed hourly rates, and the newer gig-style services post fixed bounties per task. A price list is easy to explain and easy to budget for. The trouble is that a solicitor's sign-off at two in the morning, inside five minutes, is a different thing from the same solicitor's sign-off next Wednesday afternoon, in the way that a hotel room on a holiday weekend is a different thing from the same room on a wet night in November, and a single fixed price for both means either that nobody gets out of bed for the emergency or that every routine buyer overpays.

Countersignatory is therefore built as a market, and its auction is the AdWords one turned around. Buyers post a task with a deadline and the most they will pay. Verified responders keep standing asks, which are the least they will accept for a given kind of task. When a task arrives, the eligible responders are ranked by their ask divided by their quality score, so that a cheaper ask and a better record both improve your position, and the winner is paid the highest ask at which they would still have come first, adjusted for quality. In AdWords the advertiser pays just enough to beat the bid below them; here the responder is paid above their own ask, up to the level of the next ask in the book. The useful property, which carries over from the original, is that the sensible thing for a responder to do is simply to state the price they would accept, because a good reputation will earn them more than that without any negotiation.

We have a further reason for wanting a market rather than a price list, which is that nobody knows which of these tasks will still need a human in three years. Routine judgment is getting cheaper every quarter as the models improve, and any company that has fixed a price for it has, in effect, bet that this will stop. A market does not need to take a view. As models absorb the easier work, the asks in those markets fall and the index falls with them, and what is left on the book is whatever people still cannot get from a model, which at the moment looks like accountability, physical presence, work that the law reserves to licensed people, and taste. Letting the price find its own level is a way of not having to guess what stays human.

The law, and the insurers

There are two dates to know about. Colorado's Automated Decision-Making Technology Act comes into force on 1 January 2027, and from then a consumer in Colorado can demand a meaningful human review of an adverse decision that automated technology materially influenced, in employment, housing, lending, insurance, health care, education and public services.¹ The EU AI Act's obligations for stand-alone high-risk systems apply from 2 December 2027, sixteen months later than originally planned after July's Digital Omnibus. Among them is Article 14, which requires that such systems be designed for effective oversight by natural persons.²

The insurers are ahead of both. The syndicates writing cover for AI deployments at Lloyd's have been explicit that they want to see meaningful human oversight and the telemetry to show it happened, and AIUC-1, the audit standard for agents that some of them have begun to ask for, tests for the same controls.

None of these rules requires anyone to use a market. What they require is that an accountable person can be reached at the moment the agent needs them and that a secure record is kept of what they decided; and for most companies it will be cheaper to find that person when needed than to keep one on a retainer.

Where we are today

We should be clear about where we have got to so far. The market itself is not open. What is running is the part that comes before a market, namely the index, the quote engine and registration for both sides.

The Spot Index has been seeded from prices anyone can check for themselves: Prolific's published rates plus its fee, the wholesale cost of remote online notarisation, observed costs for three-person consensus tasks, and expert-network hourly rates. Every figure is marked indicative and the derivation, including the things the model does not yet know, is at countersignatory.com/v1/methodology. The index will start to move when real transactions replace the seed, and until then it will not move at all.

The quote engine is the code that will eventually price the live book. You can ask it now what a verified human would cost for a particular task at a particular deadline, and it replies with a price, a band around it, the multipliers it applied, the inputs it ignored, and whether the most you offered would have been enough. No task is being fulfilled yet and nobody has been paid. One commitment to responders is already in force, which is a published floor of $0.25 net per active minute. We will not adjust the floor to make the numbers look better. It is the test a market has to pass before it opens, and a market whose clearing price cannot cover it will stay closed.

We are finding out whether anyone wants this by watching what people do with the endpoint rather than by asking. An API key is free, immediate and needs no approval, so that the count of distinct callers reflects demand, and every quote is logged with the market and the tier it was for. We have given ourselves six weeks and a bar of 200 distinct callers or 2,000 priced tasks, and we will be paying particular attention to whether anyone asks for the Countersign market, because that is the one the business depends on. We will publish the result whatever it is, including if this turns out not to be a thing.

The ask

If you build agents, consider whether there is a point in your workflow where a person ought to be involved and is not: a transaction size above which a human look would reduce your exposure, a class of decision where three independent people agreeing would settle an ambiguity your model keeps tripping over, or a step that cannot go ahead until a named or professionally registered person has signed it off because a regulator, an insurer or a customer says so.

If there is, price it. Take a key with a POST to /v1/keys and send the task to /v1/quotes, or add https://countersignatory.com/mcp to an MCP client and let the agent ask on its own behalf. Either way it takes a couple of minutes and costs nothing. If what you need is a name on the record, say so on the interest form, because that is the market we most need to hear about.

If your name is on a professional register, whether as a solicitor, an accountant, a notary, an engineer or a nurse, or you are simply a careful person with the occasional spare half hour, the other side of the book is open. Credentials are verified against the register, the floor is published, payouts will be immediate when the first market opens, and a good quality score earns you more than your own asking price. You will not be charged to join, and there are no fees to responders at any point; the exchange takes its fee from the buyer's side of the transaction.

The Electric Monk

In Dirk Gently's Holistic Detective Agency, Douglas Adams writes about a device called the Electric Monk, a labour-saving appliance along the lines of a dishwasher, except that what it saves you is the effort of believing things.³ There is so much to believe these days that a Monk will take it on for you, and it will believe things that are not true with exactly the same conviction as things that are.

Every agent in production is an Electric Monk in that sense. It acts on what it believes, tirelessly, and it is as confident when it is wrong as when it is right. What we are building is the appliance that goes with it: rather than a machine that believes things on your behalf, a machine that finds the person who is prepared to put their name on the line. That person is paid the same whether they sign or decline, because what the buyer is purchasing is the examination rather than the approval, and a signature that could not have been withheld would be worth nothing.

The going rate for that person is on the front page. We await your countersign.