Amazon just gave AI agents their own wallet. That's the purchasing-authority question headed for your procurement agent

Amazon's cloud division just shipped the plumbing for AI agents to pay for things on their own. Bedrock AgentCore Payments, built with Coinbase and Stripe, went generally available on August 18. An agent can now hit a paywalled API, data feed, or another agent's service, get an HTTP "402 Payment Required" response, sign a stablecoin payment, and get what it asked for — no human clicking "buy," no card on file, no sales call. Construction hasn't started using this yet. But it's the standard other vendors will build procurement and estimating agents on top of, and the industry doesn't have a version of "who approved this purchase" for a system where the buyer isn't a person.

What did AWS actually launch?

AgentCore Payments runs on x402, an open payment standard Coinbase released in 2025 that repurposes the long-dormant HTTP 402 status code. When an agent requests something that costs money — an API call, a licensed data feed, a paid tool another agent offers — the server answers with payment terms instead of the content. The agent signs a stablecoin payment, attaches proof, and the server hands over the resource, machine to machine, in seconds, no checkout page. By AWS's account, x402 was already processing well over 100 million transactions a month across blockchain networks before this launch made it a built-in AWS capability instead of something a developer had to wire up by hand.

Who's actually in control of the money?

AWS's own documentation is explicit that the guardrails live outside the AI model, not inside it. A human provisions the wallet, approves which merchants it can pay, and sets a budget through a console the agent can't reach; spending limits are enforced at the infrastructure layer independent of anything the agent reasons its way into deciding, and every transaction lands in AWS's logging tools alongside the rest of the agent's activity.

That's a real improvement over letting a model hold its own credentials. It's also built for a world where "the merchant" is another API endpoint charging fractions of a cent, not a building-material distributor invoicing net-30. Nothing about it touches a construction ERP, a lien waiver, or a schedule of values.

Why should a GC's ops director care about a feature that doesn't touch materials yet?

Because it's the pattern construction software vendors keep building agents on top of. Estimating platforms are shipping agents that pull live pricing off distributor sites. Procurement tools are piloting agents that renegotiate subscription data feeds — plan-room access, code-compliance databases, pricing services. Right now those agents work through APIs a human already contracted for. AgentCore Payments is the piece that lets the next version discover a new paid data source mid-task and pay for it itself, the same way it currently pulls a web page or runs a calculation.

Construction runs on the assumption that a person authorizes every dollar leaving the company: a purchase order, a three-way match against the receipt and invoice, a signer card at the bank. None of that has an equivalent for "an agent decided this API was worth $0.003 and paid for it without asking." At a fraction of a cent, nobody notices. At a hundred agents each making that call a thousand times a day across a portfolio of projects, it adds up the same way an unsupervised token-spend problem did for one CEO's AI coding agent last week — except this time the agent isn't burning compute, it's transacting with outside vendors.

What should a contractor do about it now?

Nobody needs a wallet-controls policy for a feature they aren't using yet. But the moment a vendor pitches an agent that can "source its own data" or "negotiate its own subscriptions," ask directly:

• Who is the named human owner of any payment credential the agent holds — the way a person, not a department, holds signing authority on a company card?
• Is there a hard dollar cap per transaction and per period, enforced somewhere the agent can't override, not just a soft budget the vendor promises to respect?
• Is the merchant list an allowlist, so the agent can pay only pre-approved vendors rather than any endpoint that answers with a 402?
• Does every agent-initiated payment land in the same AP review — matched against a PO or an approved subscription — a human purchase already goes through?

The technology to let an agent pay for things without asking is now a standard AWS feature, not a research demo. Construction's paperwork trail hasn't caught up to a buyer that isn't a person — worth closing before an agent that can pay for things gets pointed at your job cost codes.

Before approving an AI agent for procurement, estimating data, or subscription management, ask the vendor directly whether it holds its own payment credentials — and if so, who set the cap and who reviews the transaction log.

Construction AI Brief publishes three times a week. Subscribe at constructionaibrief.com.