Who is Accountable When AI Agents Start Talking to Each Other? 

Everyone is building AI agents. Who is asking what comes next?

At Web Summit Lisbon, nearly every stage talked about what AI agents can do: research products, book trips, negotiate, buy. Companies are now building agents to act for their brands, but almost no one is talking about what happens next, when those agents start dealing with each other and with customers. 

This is something marketing professionals need to be asking. 

• If we are creating AI agents to do work on behalf of our brand, how do we ensure consumers know they can trust them? 
• When two agents transact and something goes wrong, who is liable? 
• How do we trace it back? 
• How do we find the agents that are quietly working against our business or our brand? 

Nick Thompson, CEO of The Atlantic, asked Mazin Gilbert, executive director of the Agentic AI Foundation, all the right questions. 

Who is responsible when an AI agent makes a mistake? 

As of November 2026, it is unclear who is responsible when an AI agent gets it wrong.

Tell your assistant to book a flight home. It hands the job to another agent, which negotiates with the airline’s agent, and you end up routed to a different destination on your corporate card. Who pays? You, the agent, the airline?

Today, who knows? Not because the law is unclear, but because it’ll be complicated to reconstruct what happened. There is no standard way to log a transaction from the first request to the last, so you can’t clearly say which party made the mistake. At least, not without a long dispute.

Then say the agent was authorized to spend $20 and spent $1,000. Or a merchant’s smarter agent talked your agent into a purchase you never wanted. Can’t be immediately solved today. 

What is the Agentic AI Foundation trying to build? 

The Agentic AI Foundation is defining open rules for how AI agents identify themselves, what they may do, and how their actions are recorded. 

It sits under the Linux Foundation and has more than 300 members, including Anthropic, OpenAI, Google and Microsoft among others Its goal is to define the rules of the road for agents, in the open, the way TCP/IP and HTTPS did for the internet.

These are the building blocks they shared: 

• Identity: which person or business an agent belongs to. 
• Authorization: what it may do, including how much it may spend and with which payment method. 
• Attestation: proof the agent is who it claims to be, via registries that agents must appear in. Companies would run their own registries, and registries would connect to each other. 
• Logging: a record of every step, so a transaction can be traced from A to Z. 
• Sandbox and gateway: agents stay inside a defined space, and every outside action passes through a checkpoint that enforces policy. 

They also draw a useful line between two protocols: 

Can the Agentic AI Foundation actually do it? 

Hopefully. They are building a baseline.

Their approach is for rival companies to debate use cases, vote, and publish open reference implementations that anyone can use. But can they put their interests above themselves? 

Standards cover protocols, interfaces and logging. They do not settle liability, bribery of agents or the business model. Those will take laws and contracts, which move far more slowly than software. Mazin Gilbert stated that whoever deploys a product is responsible for it, which is sensible, but he also said the accountability question is still being debated.

There is also a timing problem. The agent-to-agent standard matters most when businesses need their agents to transact with each other at scale, and today most don’t. Standards that arrive after the habits form are hard to enforce. 

Finally, a standard only protects you from agents that choose to follow it. A scraper pretending to be a human, or a visitor spoofing a trusted source, is not going to register itself. 

What do marketing professionals need to be aware of? 

Tell real agents from imposters, take responsibility for the agents you deploy, and keep a record of what they do. 

The agents you deploy are your responsibility 

Treat every agent as a spokesperson for the brand, with a defined role, a named owner and clear limits on what it can say and spend.   

Agents will be your customers, and your imposters 

We need to understand the difference. 

How can we tell real agents from imposters? 

Today, the best answer is proof rather than appearance. An agent’s name, IP address and origin can all be faked, so the industry is moving toward agents that cryptographically sign their requests, letting a site verify who sent them.

Visa and Mastercard are building their agent payment systems on the same idea. But this is still being standardized, not every agent signs yet, and a verified identity doesn’t prove good intent. 

Three steps to start: 

• Ask our IT or security team whether our websites verify signed agents. 
• Ask for analytics that separate verified agents, unverified agents and humans. 
• Require agents you deploy to sign their requests, carry a spending limit and keep a log of what they did. 

Make sure you can see how agents conduct themselves 

Before you let an agent spend money or speak for your brand, ask whether you can see every step it took, whether it has a defined spending authority, and what the contract says if it goes wrong. If the vendor can’t answer, treat that as the answer. 

Be clear about who is accountable inside of your company for orchestrating your AI tools 

Add agent transactions to that, and any gap in accountability becomes a liability. 

Put your own safeguards in place now 

In my view, even the best case means a gap before broad adoption. The AI agentic foundation’s advice is to build only on open standards and to put a sandbox, a gateway and an identity registry in your system now. 

What should we ask before the next agent knocks? 

The people who will feel an agent’s mistakes first are the brands and customers on the receiving end. I would start with these questions: 

• Which agents are visiting my channels, and can I tell the real ones from the fake? 
• Who in my organization owns agent risk? 
• What is every agent I deploy allowed to spend and say, and who approved it? 
• Can I trace what it did, step by step, if something goes wrong? 
• What do my contracts with agent and platform vendors say about liability? 

The next wave of marketing won’t only be about getting chosen by our customers’ agents. It will also be about proving what our own agents did on our behalf. 

And yes, Hotwire can help 

Hotwire advises brands on AI agent transformation and AI visibility. Our AI Labs team develops the consultancy work and our proprietary tools and works closely with clients and practitioners on AI visibility and transformation projects including Cloudflare, Honeywell and Flexential.  Our AI solutions are built with and for marketing and communications clients, so we know how these systems are built from the inside out.

Right now, most of us have more questions than answers, and that’s fine. Knowing which questions to ask, and starting to build the answers, is the real work. If you’d like a partner, we’re happy to help.

FAQ