This month I added a humanitarian layer to Date the war, my globe of sourced conflict events. Each country now carries a crisis severity rating from ACAPS, and a short guide to helping: which UN agencies and charities are working there, whether a gift is tax-deductible in the US, and six public checks on each group that anyone can repeat. I wrote it so that an AI assistant asked “where should I give?” can read it and hand the answer to the person who asked.
Building that left me with a bigger question. AI agents are starting to buy things for people, and to pay other agents to do it. Google, Coinbase, OpenAI and Stripe have all published open protocols for agent payments since 2025. If that becomes a large share of commerce, it is a new pipe that money runs through. Could a little of it be set aside, on purpose, for the people who need it most?
The idea
An agent’s owner says yes once: give 0.1% of what my agent spends, up to a monthly cap. From then on, each payment the agent makes carries one optional field. The processor settles the purchase in full and holds the small share apart. Once a month each processor sends one transfer to a global fund and publishes the total, along with a fingerprint of every line in it, so any owner can prove their gift is counted.
The fund would sit under UN responsibility, held by a trustee, and pay out by a published rule on public data: this year’s unmet UN appeals, and how severe each crisis is. The UN already runs a pool built for this. Its Central Emergency Response Fund has one window for disasters that just started and one for crises other donors have left underfunded, and it takes money from governments, companies and individuals.
It has been done with airline tickets
The UN cannot tax anyone. Countries can. In 2006, France and Chile put a small solidarity levy on airline tickets; France’s is one euro on an economy seat. The money funds Unitaid, which buys medicines for HIV, tuberculosis and malaria. From 2006 to 2019, 63% of its contributions came from that levy. Nobody stopped flying over a euro.
So there are two roads. The voluntary one starts now: an optional field that owners switch on. The mandatory one, if it ever comes, is the airline road: countries pass their own laws and point them at the same fund.
Where the money would go
The simulator runs the rule on this year’s real numbers. UN-coordinated appeals for the 35 rated countries that have one ask for $30.1 billion. They have received $12.6 billion. $17.6 billion is still missing.
Nobody knows how much agents will spend, so the simulator takes scenarios rather than forecasts. If agents handled $1 trillion a year, and 30% of owners opted in at 0.1%, the fund would raise $300 million. After a quarter is held back for new emergencies, it reaches 27 crises and closes 1.3% of the gap. At $10 trillion with half of owners in, it raises $5 billion and closes about 21%.
The rule is short. Weigh each severe crisis by its people in need, more for more severe ones. Fill unmet appeals in that proportion, never past what an appeal still needs, and pass anything left over to the rest. On the default settings, Sudan gets the most, then Nigeria and Yemen.
Running it on real data also showed me a hole I had not thought of. Ethiopia is rated at the top of the severity scale, with 16.6 million people in need, and has no UN country appeal on record this year. A rule that only fills appeals gives it nothing. The simulator has a switch to estimate a gap for crises like that. Which way to set it is exactly the kind of choice a fund should make in public.
What the tests found
A rule that moves money to crises should be checked against the people who already do it. So I ran it on seven years of real data, 2019 to 2025, giving it each year exactly what the UN’s fund for underfunded emergencies gave, and compared where the money went. About half of it lands in the same places, under every rule I tried. Much of the half that differs is large crises, such as Ukraine and Palestine in recent years, that the UN’s fund for underfunded emergencies did not cover that year. The data does not say why, and that fund knows things no formula sees: who can reach whom, which agencies can spend fast, and what its other windows already paid for.
I also tried to game it. Asking for more money in an appeal buys nothing at realistic sizes, because the rule weighs people in need, not dollars asked. Reporting 20% more people in need buys about 19% more money. That is the rule’s real weak point, and it is shared by every formula based on need. The defence is to check those counts against independent ones before paying on them.
The ledger is now built the way the web’s certificate logs are (RFC 6962), and a separate checker in Python has to agree with the page before anything counts. The field itself is written up as a draft extension against x402’s published spec, with the open questions for its maintainers. And because this is a set of choices rather than one design, the options page lays out each choice, its strongest alternative, and three paths: this voluntary one, a processor offering it the way Stripe Climate does for carbon removal, and countries passing a levy the way they did for Unitaid.
What it should never do
Nobody gets opted in silently. It is the owner’s money, and the field only rides on payments whose owner turned it on.
The fund learns totals, not purchases. Processors publish how much, never what.
Every step can be checked from outside. The ledger is a simulated month: twelve thousand made-up payments, three made-up processors, and a real allocation at the end. You can rebuild each processor’s fingerprint in your browser, check one owner’s receipt, and then change a single line by one cent and watch the check fail.
It is not a substitute for aid budgets. A fund like this should report what governments give alongside what it raises, so it cannot quietly become the reason they give less.
What would have to be true
A lot. Payment processors would have to carry the field and publish their batches. The UN would have to agree to hold and allocate the money, and a charity conduit would have to issue tax receipts where the law allows. Countries would have to decide whether to make any of it mandatory. None of that is mine to decide, and nothing here is affiliated with the UN or any company named.
What I can do is show that the pieces fit: one optional field in a payment, a public rule on real data, and a ledger anyone can check. If you work on humanitarian financing or agent payments, I would like to hear where this breaks.
I built all of it with Claude Code, from the data pipeline to the pages. The appeal data comes from OCHA’s Financial Tracking Service through HDX, and the severity ratings from ACAPS. Both are open.
