What AI agents actually do today – and what’s still the future
The term “agentic commerce” describes purchases that are prepared or completed fully or partly by AI agents. A few announcements from major providers show what’s already real today:
| Date | Provider | What was announced |
|---|---|---|
| Nov. 25, 2024 | Anthropic | Model Context Protocol (MCP): an open standard for connecting AI assistants to the systems where data lives – such as databases or business applications [1] |
| Sept. 16, 2025 | Agent Payments Protocol (AP2): an open protocol for payments triggered by agents. At its core are “mandates” – cryptographically signed proof of what the user has authorized. According to Google, supported by more than 60 organizations [2] | |
| Sept. 29, 2025 | OpenAI and Stripe | Agentic Commerce Protocol (ACP) and Instant Checkout: users in the US can buy from Etsy sellers directly in ChatGPT; the order is passed via the protocol to the merchant’s system, which can accept or decline it [3] |
These examples come mainly from retail with consumers. In B2B business with consulting-intensive services, a purchase completed by an agent is still the exception today.
Much more common is an earlier phase: agents do research. They visit websites on a user’s behalf, summarize offerings, and create comparisons. OpenAI, for example, names a separate user agent called “ChatGPT-User” for retrievals triggered by a user [4].
What agents look for on a B2B website
An agent researching for a buyer needs the same information as a person – only more strictly: what isn’t clearly stated on the page, it can’t use or has to guess.
- Scope of services: what exactly is offered, in which packages or variants?
- Prices or pricing logic: specific prices, price ranges, or at least what the price depends on.
- Terms: contract periods, cancellation, service levels, data protection, data location.
- Fit: which company sizes, industries, and requirements is the offering designed for – and which not?
- Integration: which systems are supported, are there interfaces and documentation?
- Next step: how can someone request a quote, start a trial, or book a meeting?
For products, Google offers structured data that marks up price, availability, and other details in a machine-readable way. Google distinguishes between pages where you can buy directly and pure product pages [5]. These formats don’t always fit B2B services that need explaining – but the underlying principle still applies: unambiguous information in a fixed place.
Five readiness questions
| # | Question | How you know you’re “ready” |
|---|---|---|
| 1 | Are your offerings described unambiguously? | Every service or package has its own page with scope, target audience, and boundaries |
| 2 | Can prices or pricing logic be found? | Prices, ranges, or pricing factors are in the text – not just in a PDF or behind a form |
| 3 | Is the information consistent? | Website, data sheets, marketplaces, and profiles state the same facts |
| 4 | Is the content machine-readable? | Key information is in the HTML, not only in images or scripts; structured data where appropriate |
| 5 | Are there rules for what agents may do? | A deliberate decision in robots.txt; clear processes for inquiries that arrive automatically |
Risks and governance
If you make access easier for agents, you should know the risks:
- Misrepresentation: an agent may pick up outdated or contradictory information and pass on a false picture of your offering. Consistent, current data is the best protection.
- Manipulated content: the OWASP initiative ranks “prompt injection” first on its list of security risks for applications using language models. This includes indirect attacks via external content such as websites. According to OWASP, no reliable protection is known yet [6]. Anyone using agents themselves – in procurement, for example – should keep this in mind.
- Accountability: when agents trigger inquiries or orders, who is responsible, and how do you verify that the order was really intended? Google’s AP2 addresses exactly this with signed “mandates” [2].
- Data protection and confidentiality: which information should be public and machine-readable, and which only after contact?
A sensible first pilot
Getting started requires neither new protocols nor a big project. A pragmatic pilot in four steps:
- Pick one offering – ideally one with a clear structure, such as an entry package or a standard product.
- Answer the five readiness questions and close the gaps: scope, prices or pricing logic, terms, fit.
- Run the agent test: ask an AI assistant with web search to compare your offering with two competitors. What’s represented correctly, what’s missing, what’s wrong? Do several runs, because the answers vary.
- Document the results and repeat after three months.
Only once these foundations are in place is it worth looking at protocols such as MCP, ACP, or AP2 – and even then only if your business model involves self-service or standardized purchases.
Sources
- Anthropic: “Introducing the Model Context Protocol” (Nov. 25, 2024)
- Google Cloud Blog: “Announcing Agent Payments Protocol (AP2)” (Sept. 16, 2025)
- Stripe Newsroom: “Stripe powers Instant Checkout in ChatGPT and releases Agentic Commerce Protocol codeveloped with OpenAI” (Sept. 29, 2025)
- OpenAI: “Overview of OpenAI Crawlers”
- Google Search Central: “Intro to Product Structured Data on Google”
- OWASP Gen AI Security Project: “LLM01:2025 Prompt Injection”

