Skip to content
valiss AI
EN DA
← All posts
· 7 min read · strategy

Agent to agent: what the A2A protocol means for SMEs

A2A lets one company's AI agent deal directly with another's. Here is what that changes for small and medium businesses in Europe, and what to do this year.

A container ship being loaded at Terminal Burchardkai in Hamburg at dusk, with cranes reaching over the deck.

The advice small and medium businesses get about AI agents sounds the same everywhere right now: build your own. That is the wrong place to start. The decisive change in 2026 is not that you get an agent. It is that your customers, your suppliers and your competitors get one, and that their agents have to find you, read you and verify you without a human ever opening your website.

A2A is the protocol that makes that possible. It stopped being a research project some time ago, and it is worth understanding before someone sells you something on the strength of it.

What is A2A?

A2A, short for Agent2Agent, is an open protocol that lets two AI agents belonging to different companies find each other, agree on a task and deliver a result with no human intermediary.

Google published the protocol in April 2025 and handed it to the Linux Foundation that summer. In April 2026 the specification reached version 1.0 with more than 150 organizations behind it, among them AWS, Cisco, IBM, Microsoft, Salesforce, SAP and ServiceNow. It ships inside Azure AI Foundry, Copilot Studio and AWS Bedrock AgentCore. There are finished development kits for five programming languages, and the first real production work sits in supply chains, insurance and financial services.

The protocol gets confused with MCP constantly. The difference is ownership. MCP connects an agent to tools and data inside your own world, while A2A connects your agent to another company's agent. One crosses a system boundary. The other crosses a company boundary, and it is the second one that changes anything for you.

The difference in scale between the two is worth carrying with you. MCP is already infrastructure in practice, with around 97 million development kit downloads a month and more than 18,000 publicly known servers. A2A is younger and thinner in real use, but it solves the problem MCP cannot touch, which is what happens when the two systems belong to different companies and neither one trusts the other.

Sitting on top of it is AP2, a protocol for payments between agents, backed by around 60 organizations. That is where money starts moving without a manual approval at every step.

Why this is not about building your own agent

Most European small and medium businesses will never build an agent. They do not need to. But they will be assessed by one, and that arrives faster than most people expect.

In 2025, 42 percent of Danish companies with at least 10 employees used AI, against an EU average of 25 percent. Among the smallest, those with 10 to 49 employees, the figure was 37 percent, up from 24 percent the year before. Among large companies it was 75 percent. Denmark sits first in Europe on that list, and the number is usually presented as a success story.

Read the same figures as a statistic about buyers instead, and they look different. Three in four large Danish companies already have AI in the building. They are your customers. And large companies are exactly the ones who put a system on the preparatory work of a supplier assessment first, because they have the most suppliers to keep track of.

The new buyer does not read your front page. It pulls whatever you made machine readable, compares you against three others, and puts a recommendation on a table you never get into.

It is the same mechanic we describe in what AI visibility is, moved somewhere with more money on the table. From a private person asking ChatGPT for a recommendation, to a procurement department that has set a system to draw up the short list.

Here is what that looks like in practice. A property manager needs the windows replaced across fourteen stairwells in Valby. The assistant takes the job, finds twelve possible suppliers, and then starts eliminating. Do they carry valid insurance? Is their service area stated anywhere? Is there a price, or a price range, that can be compared, or do you have to call to get one? Eight of the twelve fall out on questions like that. Not because they are poor tradesmen, but because the answer did not exist anywhere the system could read it. The four that remain get an email from a human being. The eight get nothing and never find out they were in the running.

Denmark has been through this before

There is a Danish story that explains it better than any protocol specification can.

In 2005 it became mandatory to send invoices to public authorities electronically. NemHandel and OIOUBL became reality overnight for every tradesman who had a municipality as a customer. The lesson was brutal and it was useful. If the invoice was not in the right format, it did not arrive. Nobody called. No case worker set it aside and asked for a new one. It simply vanished, and the money did not come.

The shape of it is identical now. The difference is that it no longer applies only to the invoice. It applies to the quote, the delivery time, the warranty, the certification and the proof that you carry insurance that actually covers the job.

Peppol has since replaced most of the Danish arrangement, and the companies that learned the lesson in 2007 were the ones who found the switch easy. The format changes every five years. The habit of keeping your own facts machine readable does not.

The registries already exist, and they do not talk to each other

This is where the picture gets messy, and it is worth knowing before anyone sells you a listing.

A count from the first quarter of 2026 found more than 104,000 registered agents spread across upwards of 17 separate registries, eleven competing ways of finding each other, and at least ten active IETF drafts on how the problem ought to be solved. The registries do not interoperate. An agent registered in one place is invisible to a client looking in another.

The comparison the industry keeps reaching for is the right one: the agent ecosystem in 2026 is roughly where the web was before DNS existed.

For you that means one thing. Do not pay for a listing in a registry this year. Most of them will not exist in two years, and none of them has the coverage the seller claims.

What lasts is what you own. A2A defines an agent card, a small JSON file published at a fixed path on your own domain, stating what you can do, how to reach you and how to verify that it is really you. Version 1.0 added cryptographic signing of that card. Your domain is the only address in this entire picture that does not get acquired, renamed or shut down.

What you can actually do this year

Five things. None of them require you to build an agent.

  • Put prices, delivery times, service area and terms on your own domain as text. Not inside an image, not inside a PDF behind a form.
  • Collect the facts a professional buyer always asks for in one place: company registration number, certifications, insurance coverage, response time, and references with the year attached.
  • Make sure what you say about yourself sits in the places the models already pull from. Your own site, your Google Business listing, your trade registries.
  • Publish an agent card on your domain the next time you have the site open anyway. It is an hour of work, and you do not need an agent behind it yet.
  • Meet the disclosure duty if you run a chatbot or a phone assistant.

That last point gets misread often enough to be worth stating precisely. Article 50 of the AI Act applies from 2 August 2026 and requires you to tell people when they are speaking to a system rather than a person. The obligations for high risk systems were pushed to December 2027 by the digital omnibus package. The disclosure duty was not pushed. It applies now. The ceiling is €15 million or 3 percent of worldwide turnover, and for a small company the lower of the two figures is the one that gets used. It is still a number that closes a business with twelve employees.

What to watch

If you want to know where you stand before deciding anything at all, that is what our audit is for. Signal runs live queries against ChatGPT and Gemini in your real category, scores you across five pillars and benchmarks you against three competitors. The output is a 10 to 15 page PDF and a 30 minute walkthrough call within one business day. €690.

The interesting question over the next year is not whether A2A beats a competing protocol. It is when the first of your customers asks for something machine readable in a tender and phrases it as though it were obvious. At that point the question stops being technical. You either have the answer ready, or you do not.

Ready

What’s your brand’s score?

Find out where your brand stands. Book a discovery call and we’ll run Signal on your brand together.