When Machines Pay Machines: The Economy Between the Clicks

A small business owner says, “Keep enough packaging in stock for the next two weeks. Stay inside this budget. Ask me before changing suppliers.”

What follows could become one of the most economically important sequences of the AI era. Her agent compares offers. A supplier's agent confirms availability. Another service estimates delivery. An authorised payment settles. Stock arrives before anyone loses a day to an avoidable shortage.

This is an imagined transaction, not a claim that every element already works reliably across Nairobi's businesses. But the pieces are moving into place. I believe agent-to-agent commerce could become a major economic layer as digital labour gets cheaper, physical production improves and routine coordination becomes easier to automate.

The slogan is attractive: cheap labour plus cheap goods creates an A2A economy. The missing term is trusted authority.

A payment is not a conversation

Two agents can exchange convincing messages without either being authorised to spend money. They can agree on a price while misunderstanding the product. They can be manipulated by a webpage, a forged invoice or an instruction hidden inside a supplier's description.

The economic problem is therefore more demanding than connecting chatbots. Commerce requires identity, permission, exact terms, settlement and recourse. A useful system must answer who instructed the purchase, what limits applied, what was actually bought, and who is responsible if something goes wrong.

Google announced the Agent Payments Protocol in 2025. Its current AP2 specification describes signed checkout and payment mandates, receipts, and both direct and autonomous modes. The specification separates probabilistic agents from deterministic validation responsibilities. It is a developing technical standard, not proof of universal merchant adoption or a guarantee against fraud.

A2A communication and AP2 payment authorisation solve related but distinct problems. A protocol for finding and coordinating agents does not, by itself, give them permission to debit an account.

Why cheaper intelligence changes transaction size

Human attention has a minimum practical price. A manager cannot spend twenty minutes negotiating every small restock. A developer cannot manually bargain for every tiny computational service. Many potentially useful transactions do not occur because coordination costs exceed their value.

If agents reduce those costs, smaller exchanges become viable. A business could buy a few seconds of specialised computation, a narrow data-validation service or a transport quote bundled into a larger workflow. Physical goods could be replenished in smaller, better-timed batches where logistics costs permit.

This does not mean all payments become microscopic or that existing firms disappear. Networks charge fees; sellers set minimum orders; delivery still costs money. Agents change the coordination frontier. They do not abolish the physical economy.

The agent economy will grow where the cost of deciding falls faster than the cost of trusting.

Africa has a head start—and unfinished work

Mobile money has already made digitally mediated exchange familiar across much of the continent. That experience matters, but consumer payment adoption does not automatically create safe autonomous spending. Machine authority needs its own controls and its own relationship with regulated institutions.

A Kenyan business should be able to express limits in terms it understands: maximum amount, approved suppliers, permitted product categories, timing, required confirmation and revocation. An agent should receive a narrowly scoped ability to act rather than unrestricted access to a wallet or banking credential.

Receipts must remain understandable to the human principal. A person should not need to reconstruct a hundred agent messages to discover why money left an account. The commercial advantage comes partly from removing administrative burden; an opaque autonomous system can add a different burden in its place.

Security is the growth strategy

A signed payment can still reflect a bad decision made before signing. Cryptography protects particular representations and authority relationships; it cannot make a false supplier claim true or guarantee that an agent interpreted a user's intention correctly.

Recent research has begun testing that boundary. A 2026 AP2 red-teaming preprint reports prompt-injection failures in a particular shopping-agent implementation. Such studies should be read as evidence about tested systems and conditions, not universal failure rates for every AP2 deployment. The lesson is valuable: formal payment authorisation and robust upstream decision-making both matter.

Practical systems need constrained tools, verified supplier records, explicit approvals for exceptional spending, anomaly detection, transaction logs and a way to halt activity quickly. Deterministic checks should enforce hard limits even when the model proposes something fluent and persuasive.

The business model may be quieter than the hype

The biggest early opportunity may be routine business purchasing rather than autonomous consumer shopping. Companies already have budgets, approval chains, suppliers and recurring needs. An agent can reduce reconciliation work, discover inconsistencies and prepare decisions before receiving permission to transact.

A sensible adoption path starts with recommendation, moves to supervised execution and only later permits narrow autonomous transactions. Progress should depend on measured accuracy, fraud exposure and administrative savings. A business does not need to surrender all control to benefit from automation.

Interoperability also matters. If every agent operates only inside its provider's marketplace, the new economy may become a collection of expensive gates. Open standards can reduce switching costs, but contracts and implementation choices determine whether that promise reaches customers.

More commerce is not always more welfare

An economy full of agents could generate excessive bidding, spam, manipulative pricing and machine-speed disputes. Transaction counts alone would be a poor measure of success. The useful measures are lower total costs, reliable fulfilment, reduced fraud, better access for smaller firms and time returned to people.

I expect A2A commerce to grow because coordination is such a large hidden cost. Cheap digital labour and increasingly affordable physical goods create more opportunities to exchange. Trusted, bounded payments let those opportunities become transactions.

The future owner in our opening scene should not be impressed because an agent spent money. She should be satisfied because the right stock arrived, the accounts reconcile, the limits held—and she remained in charge.

Editorial note: This essay expresses the founder’s ideas, developed with AI-assisted research and writing and reviewed against the linked sources on 29 October 2026. Convictions, proposals, and forecasts are identified in the text; scenes framed as imagined are illustrative.

Cover photograph / illustration: Fiona Graham. CC BY-SA 2.0. Original source. Illustrative image: M-PESA mobile money and Equity agent, Nairobi, Kenya. Photograph: Fiona Graham. CC BY-SA 2.0.