Procurement teams have used automation for decades: catalogues, reverse auctions, approval workflows. What is new in 2026 is software that can conduct the conversation itself, searching for suppliers, proposing terms and committing to a purchase. Before handing that authority to an agent, it is worth looking at what the evidence says about how such agents behave.
The infrastructure is arriving
Payments are one of the building blocks. In September 2025 Google announced the Agent Payments Protocol (AP2), an open protocol developed with more than 60 organisations from payments and technology (Google Cloud). Its central idea is the signed “mandate”:
- an Intent Mandate records what the user asked for and, for delegated tasks, the rules of engagement such as price limits and timing;
- a Cart Mandate records the exact items and price the user approved.
For procurement, this maps neatly onto existing practice: a mandate is a purchase authorisation with machine-readable limits and an audit trail.
What research shows about agents in markets
Microsoft Research has published one of the most detailed public tests of agents trading with other agents. Its open-source Magentic Marketplace simulated 100 customer agents and 300 business agents across several leading AI models (Microsoft Research; paper). Three findings matter for buyers:
- Good results depend on good search. Frontier models came close to optimal outcomes only when search conditions were ideal, and performance fell as the number of options grew.
- First-offer bias. All models showed a strong tendency to accept the first proposal they received. The authors report that this gives sellers a 10–30× advantage for responding quickly over offering better value.
- Manipulation. The researchers tested fake reviews, authority claims and prompt injection. Some models resisted; others could be manipulated.
The researchers’ own conclusion is blunt: “Agents should assist, not replace, human decision-making.”
The regulatory frame in the EU
Since 2 August 2026, Article 50 of the EU AI Act requires providers of AI systems that interact directly with people to design them so that those people are informed they are dealing with an AI, unless this is obvious (European Commission). A negotiating agent that writes to a human account manager at a supplier falls squarely in that category. Disclosure needs to be built into the agent’s messages, not added later.
A practical control framework
| Control | What it does | Why it matters given the evidence |
|---|---|---|
| Mandate limits | Hard caps on price, volume, term and supplier list | Contains the cost of a bad decision |
| Comparison rule | Agent must collect a minimum number of offers before accepting | Counters first-offer bias |
| Approval thresholds | Human sign-off above a set value or for new suppliers | Keeps judgement where stakes are high |
| Input hygiene | Treat supplier messages and documents as untrusted data | Reduces prompt-injection risk |
| Disclosure | Agent identifies itself as AI to human counterparts | Required under AI Act art. 50 in the EU |
| Logging | Every offer, counter-offer and decision recorded | Supports audit and dispute resolution |
Table: BusinessCraft framework based on the sources linked in this article.
Where to start
The lowest-risk entry point is high-volume, low-value spend with clear specifications, where an agent collects and compares offers and a human approves. Fully autonomous commitment should come later, once the organisation has its own data on how the agent behaves with real suppliers.
Key figures
| Figure | Value | Source |
|---|---|---|
| Simulated market size | 100 customer agents, 300 business agents | Microsoft Research |
| Speed advantage from first-offer bias | 10–30× | Microsoft Research |
| AP2 collaborating organisations | > 60 | Google Cloud |
| AI Act art. 50 transparency duties apply from | 2 August 2026 | European Commission |
Figures as published by the linked sources; table compiled by BusinessCraft.
Related reading
- AI Infrastructure in 2026: How Close Is Capital Spending to the Trillion-Dollar Mark?
- Three Playbooks for AI Power: How Washington, Beijing and Brussels Diverge in 2026
- Quantum Computing and AI in 2026: Separating Roadmaps From Hype
Sources
- Google Cloud: Powering AI commerce with the new Agent Payments Protocol (AP2), 16 Sep 2025
- Microsoft Research: Magentic Marketplace (blog, 5 Nov 2025)
- Bansal et al.: Magentic Marketplace: An Open-Source Environment for Studying Agentic Markets (arXiv 2510.25779)
- European Commission: Transparency obligations under Article 50 AI Act (FAQ)




