What Arbiter actually does

Written for someone who does not work with computers for a living. No code, no jargon that is not explained.


The one-sentence version

AI assistants are starting to spend money on people's behalf, and they cannot tell a normal payment from a trap. Arbiter is the second opinion they buy before they act — about a fifth of a penny per question.


The problem, with an analogy

Imagine you hire an assistant who is extraordinarily capable. They read any document instantly, work all night, never get bored, and do exactly what you ask.

But they have no street smarts. They have never been warned about anything.

One day a letter arrives. It says "Sign here to confirm your address — no money will change hands." Your assistant reads it, sees that no money changes hands, and signs it.

The letter was not about your address. Buried in the wording, it signed over control of your bank account to a stranger. Nothing happened that day. Two weeks later the account is empty.

Your assistant did nothing wrong by their own reasoning. They read what was in front of them, checked that no money was moving, and acted. They just did not know what they were looking at.

That is exactly the situation with AI agents today. They are being given the ability to pay invoices, move funds, and sign agreements — and they have no instinct for danger.


What Arbiter is

Arbiter is a service the assistant can consult before it acts.

It reads the thing the assistant is about to sign, works out what would actually happen, and answers in plain terms:

Answer Means
Allow Nothing wrong here. Go ahead.
Warn This is risky. Only proceed if a person has accepted the risk.
Block Do not do this. It would very likely cost you money.
Escalate I could not tell. This is not permission — ask a person.

And it always says why, in sentences a human can read.

The whole exchange takes under a second and costs a fraction of a penny.


Four things it catches

These are real, and all four are things a capable AI would otherwise walk into.

1. The document that signs away the account

A payment that says it moves zero money. Most software shows it as harmless — nothing is moving, so what is the risk?

Hidden in it is an instruction that permanently hands control of the account to someone else. Every future payment would be authorised by them, not you.

Arbiter reads that instruction and says: Block. This transfers control of your account to a stranger.

2. The permission slip

Signing something that lets a company take money from your account — not once, but any amount, any time, forever.

The dangerous part is that nothing happens when you sign it. Your balance does not change. There is nothing to notice. The money leaves next week, or next month, and by then nobody connects it to the thing you signed.

Arbiter flags the permission itself, at the moment it is granted.

3. The invoice with the wrong bank details

Your supplier sends an invoice. The company name is right, the amount is right, the logo is right. Someone has changed the account number.

You pay. The money is gone, and your supplier still wants paying.

This is one of the most common frauds in business, and an AI reading the invoice has no way to spot it — everything on the page is correct.

Arbiter checks the account number against who the supplier actually is, and says: Block. This account does not belong to them.

4. The payment that silently never arrives

On the payment network Arbiter uses, an account has to switch on the ability to receive a particular kind of money before it can receive it. If it has not, the payment is refused.

It does not bounce back with an error. It does not appear anywhere. It simply never happened — and neither side is told.

Arbiter checks first, so you find out before you send rather than three weeks later when someone chases the invoice.


The part where real people get paid

Some questions no computer can answer.

Does this photograph show a parcel actually left at a front door? Does this shop exist at this address? Is this receipt genuine?

When an AI hits a question like that, Arbiter sends it to real people. They look, they answer, and they say what they saw. Several people answer independently, and the answers are compared before a result is returned.

Those reviewers get paid, in digital dollars, straight to their own account. There is nothing to install — it works in a normal web browser, and there is no account to open with us.

Two deliberate choices:

  • Reviewers are paid for answering, not for agreeing. If you paid people for matching everyone else, they would stop looking at the evidence and start guessing what the crowd would say. That would quietly destroy the whole point.
  • Their answers include what they actually saw, not just yes or no. "Brown box on the doormat, label facing up" is checkable. "Yes" is not.

Getting paid, and keeping the books

The invoice side of this does two more things worth explaining plainly.

It shows you what the money is worth where you live. Invoice for $250 and it tells you that is roughly ₦340,000 today. That figure is a guide, not a promise — see below.

It keeps your books from the public ledger. Every payment you have ever received is recorded on a public network, so the app can rebuild your complete income history from your address alone. Change laptop, clear your browser, come back in two years — the record is still there, because it was never ours to lose. You can export it as a spreadsheet, and every line carries the transaction reference, so an accountant or a tax office can check it independently instead of taking your word for it.

What this does not do

It does not convert your digital dollars into naira, rupees or pesos and put them in your bank account. Doing that legally means being a licensed money transmitter in each country, with banking partners and identity checks. That is a regulated business, not a feature.

So the exchange rates shown are for orientation. In several countries the rate you actually get when converting to cash differs noticeably from the published one. Treat the local-currency figures as "roughly what this is worth", never as "this is what will land in my account".

What you receive is digital dollars. Where and how you turn those into local cash is your choice, and happens outside this app.

How the money works

There is no subscription, no sign-up, no sales call, no invoice.

The AI asks a question. Attached to the question is a payment of a fraction of a penny, handled automatically. The answer comes back. That is the entire transaction.

Question Cost
Is this safe to sign? about a fifth of a penny
Is this really who I am paying? about one penny
What would a person say? about twenty pence

The prices are deliberately tiny for the first two. A safety check that costs almost nothing is one an AI can afford to run on every single action — and running it on everything is the point. The moment it becomes expensive enough to think about, someone starts guessing which actions are risky, and guessing is exactly the thing that fails.

The digital dollars are called USDC. One USDC is one US dollar. Payments settle in a few seconds on a payment network called Algorand.


Who uses it

Businesses running AI assistants that touch money. Anyone whose AI pays suppliers, moves funds, or signs agreements. Arbiter is the check between the AI deciding to do something and it actually happening.

Developers building AI agents. It plugs into the common tools for building agents, so adding the safety check is a couple of lines rather than a project.

People who want to earn. Anyone with a phone or a browser can answer questions and get paid.


Why it needs to exist now

Until recently, an AI could only suggest. A person still pressed the button, and that person was the safety check.

Agents now press the button themselves. The safety check that used to be a human glancing at the screen is gone, and nothing replaced it.

Arbiter is the replacement. Not a person watching every action — that would defeat the purpose of automating it — but a service that reads every action in under a second, costs almost nothing, and knows what danger looks like.


Seeing it work

The website runs the checks live, on real examples, every time the page is opened. It is not a video or a screenshot — the results shown were worked out at the moment you loaded the page.


Honest status

It is real and it works. Money has genuinely moved — AI agents have paid for answers, and a reviewer has been paid for their work, both settled on a public payment network where anyone can verify it.

It is currently running on a test network, which behaves exactly like the real one but uses practice money. Moving to real money is a configuration change, not a rebuild.