AI agents are beginning to place orders on behalf of companies, yet suppliers have no common way of verifying whom those agents represent or what they are authorised to purchase. Emerging technical standards address parts of the problem, but the greatest challenge may be organisational rather than technological.
Martti Asikainen | 16.11.2026 | Photo Created with AI
A purchase order placed by an AI agent can be genuine, correctly authenticated and entirely unauthorised. The supplier may know exactly which software sent it, and even which company operates that software, without knowing whether the agent has any right to commit the buyer to the transaction.
This presents an awkward problem for business-to-business commerce. Companies have spent decades developing systems to identify their trading partners, but considerably less thought has gone into verifying what autonomous software is entitled to do on their behalf.
In late 2025, Gartner forecast that AI agents would intermediate 90 per cent of B2B purchasing by 2028, channelling more than 15 trillion dollars of spending through machine-to-machine exchanges (Gartner, 2025). Whether or not such predictions materialise, the question of who authorises these transactions is becoming increasingly difficult to ignore.
When a supplier receives an order from an AI agent, it needs answers to three questions: Which piece of software sent it? Who does that software represent? And what has it been authorised to do?
The first concerns identity. Cryptographic signatures can establish that a request came from a particular agent rather than an impersonator. Yet a valid signature proves only where a request originated, not whether it should be accepted.
The second concerns representation. An agent may be developed by one company, operated by another and deployed by a third. Knowing who operates the software does not necessarily establish which business it represents in a particular transaction.
The third concerns authority. Even an agent genuinely acting for a customer may have limited purchasing rights. It might be permitted to reorder standard components from approved suppliers, but not to negotiate contracts or commit the company to purchases above a certain threshold.
Two orders from the same authenticated agent could therefore require entirely different decisions. One might fall within its purchasing mandate, while another exceeds its financial limits.
The supplier needs more than proof of identity. It needs reliable evidence of the authority granted for the transaction in question.
Several initiatives are attempting to address these challenges.
Web Bot Auth, now being standardised through the Internet Engineering Task Force (IETF), allows automated clients to sign requests cryptographically. The receiving website can then verify those signatures against a public key published by the agent’s operator (IETF, 2025).
Visa’s Trusted Agent Protocol and Mastercard’s Agent Pay build on this approach, using authentication and agent registries to establish which software is making a request (Lohe & Allen, 2025).
Google’s Agent Payments Protocol (AP2) takes an important step towards establishing authority. It introduces digitally signed mandates specifying what a principal has authorised an agent to purchase. Built using W3C Verifiable Credentials, these mandates can provide independently verifiable evidence of purchasing permissions and an audit trail when transactions are disputed (Google, 2025).
The difficulty is that much of this development has focused on consumer purchases through card-payment networks. Industrial components, for example, are typically purchased on invoice, with agreed payment terms, through enterprise resource planning systems. A substantial proportion of B2B commerce never passes through a card checkout.
Fragmentation presents another obstacle. Different networks and technology providers are developing their own approaches to identity and delegation. A supplier required to support numerous competing verification systems does not really have a common standard.
There is also a question of who controls this infrastructure. Whoever maintains a directory of trusted agents may acquire considerable influence over which agents, and therefore which organisations, are permitted to trade.
Even if these approaches converge, cryptography cannot establish whether the decision behind a correctly signed request was legitimate or sensible.
A genuine agent could be manipulated by malicious instructions embedded in a document or web page. Its resulting purchase order might carry a perfectly valid signature despite having been generated against its principal’s interests.
Nor are inappropriate purchases limited to compromised agents. Carlton (2026) cites research in which AI purchasing agents exceeded buyers’ budgets in approximately one in ten cases. They also disclosed confidential budget limits to sellers despite instructions not to do so.
The implication is straightforward: an agent cannot necessarily be trusted to enforce its own restrictions. Purchasing mandates need to be independently verifiable and checked by the supplier before a transaction is accepted.
Yet this leaves a more fundamental question unanswered. Who within the purchasing organisation has the authority to issue those mandates?
Someone must decide which AI agents can commit company funds, how much they may spend, for what purposes and for how long.
In Finland, the trade register provides information about people authorised to sign on behalf of a company. There is no comparable established mechanism for verifying which pieces of software have been delegated some of that authority.
The European Commission’s proposal for European Business Wallets, published in November 2025, includes provisions for digitally managing representation rights and mandates (European Commission, 2025). However, the framework has been developed primarily around people acting on behalf of legal entities.
Whether it can be extended to autonomous software remains an important question for Europe.
In the meantime, companies deploying purchasing agents should treat them much as they would employees entrusted with company payment cards. Each agent should have its own identity and clearly defined, time-limited purchasing authority. The person granting that authority must also be entitled to do so.
Suppliers should verify individual transactions against those mandates rather than relying on the agent’s claims about its permissions. Higher-value orders may require confirmation through a separate channel, while signatures, mandates and transaction records should be retained for audit purposes.
Finland and Europe have an opportunity to build on existing company registers and emerging digital identity infrastructure. Connecting those foundations to a reliable system of agent delegation could prove more valuable for trustworthy machine commerce than any individual vendor’s protocol.
Ultimately, the challenge is not simply to prove that an AI agent is genuine. It is to establish who stands behind it, what authority they have delegated and whether a particular transaction falls within that authority.
Until those questions can be answered reliably, machine commerce will have authentication without necessarily having trust. Identity without authority is merely a name badge.
Carlton, D. (2026, 3 February). When your AI agent goes shopping, how does the seller know it’s really you?
European Commission. (2025). Proposal for a Regulation of the European Parliament and of the Council on the establishment of European Business Wallets. COM(2025) 838 final.
Gartner. (2025). Top strategic predictions for 2026 and beyond. As reported in Digital Commerce 360 (28 November 2025). Article.
Google. (2025). Agent Payments Protocol (AP2). Specification.
IETF. (2025). Web Bot Auth working group charter. IETF.
Lohe, R. & Allen, W. (2025). Securing agentic commerce: helping AI agents transact with Visa and Mastercard. Cloudflare Blog.