The EU AI Act and the Agentic Economy: What It Means When Software Hires AI Agents
A few months ago, I proposed a framework I call the Agentic Economy: software applications hire AI agents, packaged as software capabilities, to do a task, and pay for the result. This article looks at how the EU AI Act affects that framework when AI agents are hired as capabilities by other software.
If your AI, or its output, is used in the EU, the EU AI Act may apply to it, even if your company is based elsewhere. The duties are being phased in, the heaviest last, and fines can be large, with caps linked to worldwide turnover. This is a technical architect’s reading, not legal advice, of what that means for my Agentic Economy framework, where software hires AI agents as capabilities to do work.
- Someone is still the employer. Every hire should name who supplied the AI, who is using it, and what for. The registry, the hire contract and the receipt record all three.
- Risk is decided at the moment of hire. The same AI can be ordinary in one job and high-risk in another, because the Act sorts by use. Check the risk level against the purpose every time a hire happens.
- The receipt is now evidence. For high-risk AI, the Act asks for logs, in many cases explanations for people affected, and incident reports. A signed receipt per decision supports all three, in a way scattered logs do not.
- Evidence now has a price. I expect agents to prefer the supplier whose hire carries less legal risk. The track record becomes the moat.
Bottom line. Read simply, the Act and the Agentic Economy ask for similar things: say what the system does, keep records of what it did, name who is responsible.
Below: the full article, for technical teams ↓"Read simply, the Act asks three things of high-risk AI: say what it does, keep records of what it did, and name who is responsible. The Agentic Economy already runs on those three things."
Start here
If you have not read about the Agentic Economy before, read the five-minute explainer first. It covers the idea and the six terms I use below. If you prefer video, watch Hire the AI Agent: SaaC, the End of SaaS on YouTube. The written version is From SaaS to SaaC.
The short version: software used to be sold by the seat. In the Agentic Economy, a software application, or an agent working inside it, hires a piece of work for one task. That piece of work is a capability, and more and more often it is an AI agent itself. The application uses it once or many times, gets a signed record of each result (a receipt), and moves on. Think of a staffing agency, not a software licence.
Why I wrote this
Most writing on the EU AI Act looks at one picture: a company deploys an AI system, and people use it. My framework looks at a different picture. A software application needs a piece of work done, so it hires an AI agent that has been packaged as a capability, uses it for one task, gets a signed record of the result, and lets it go. Often no person is involved at the moment of hire.
That picture raises questions the usual commentary does not answer. When one piece of software hires another, who is responsible under the Act? When does an ordinary capability become high-risk? What must every hire carry with it? Where do the cost and the risk land?
So this article takes each part of the framework, the capability, the manifest, the hire, the receipt and the registry, and asks how the Act affects it, and what software needs to do about it. Six rules come out of it. If you lead a business, they are the six questions to put to your technology team. If you are the technology team, they are the six things to build.
Two things before we start.
I am not a lawyer. This is a technical architect's reading of the Act: what its rules seem to ask for, and what software should be doing about it. Your lawyers decide how the law applies to you. My job is to make sure the software is ready when they do.
The Act will keep changing. Deadlines have already been moved once, and more guidance and standards are still coming. For that reason this article does not lean on dates. It focuses on what the Act asks software to do, which has stayed stable while the timetable has not. Your lawyers will tell you when each duty applies to you.
Important definitions
Six words carry the whole model. Think of a staffing agency.
- Capability. A piece of work that software can hire for one job, the way a company hires a contractor. Often it is an AI agent itself, packaged so other software can hire it. You pay for the result, not for a seat. (From SaaS to SaaC)
- Manifest. The capability's CV. What it does, what it will never do, what it costs, how well it performs, who stands behind it. (The Anatomy of a Capability)
- Hire-token. The contract for one job. Who hired the capability, for what, under which rules, in which country.
- Receipt. The signed timesheet for every decision. What went in, what came out, which rules applied, which versions ran. Signed by both sides. Stored where nobody can change it.
- Registry. The agency's book. Where capabilities are listed, compared, hired and rated on their track record. (The Four Moats)
- OLA. The service level a capability promises in its manifest and is measured against on every receipt. Hire it, measure it, release it if it falls short. (The Agentic Economy, Explained)
Case study: a claims agent hires a fraud scorer
An insurer's claims agent needs a fraud score for one claim. It looks in the registry, finds a fraud-scoring capability, hires it, sends the claim, gets the score and a signed receipt, and releases the hire. The whole thing can take less than a second. No person is involved.
The claims agent is part of the insurer's own software. The fraud scorer is an AI model built by another company and packaged as a capability. That is the framework in one transaction: software hiring an AI agent for one task.
Now read the EU AI Act against that one transaction. The Act applies because what is being hired is an AI system; if the fraud scorer were a simple rules engine with no AI in it, most of what follows would not apply. As I understand it, the transaction now has a provider, a deployer and a risk level and, depending on that risk level, duties to keep records and to monitor, and a risk of fines. The Act mainly pictures a company that deploys an AI system and people who use it. But every one of its rules lands somewhere in this little transaction. That is what the rest of this article walks through.
What the Act asks for, in four lines
The Act sorts AI by what it is used for. A short list of uses is banned outright. A longer list, which includes hiring, credit, education and essential services, counts as high-risk and carries the heaviest duties: risk management, record-keeping, human oversight, monitoring and registration. AI that talks to people must generally make clear that it is AI, and AI-generated content must generally be marked as such, with some exceptions. Everything else carries light duties, such as supporting AI literacy among staff. The duties are being phased in, the heaviest ones last, and the dates have already been moved once; your lawyers will have the current timetable. Two things do not depend on the timetable: the fines can be large, with caps linked to worldwide turnover, and the Act can apply to companies outside the EU when their AI, or its output, is used in the EU.
Rule 1: Someone is still the employer
A company buys a document-summarising capability. Its agents hire it to summarise reports. In the Act's words, as I read them, the vendor is the provider and the company is the deployer. Each has duties.
Here is the catch. If the company's recruiting agent starts using that capability for something on the Act's high-risk list that the vendor never intended, for example ranking job applicants, the Act appears to treat the company as the provider of a high-risk system, with the provider's duties. Nobody at the company will have noticed. Whether that happens in your case is a legal question. That it can happen is reason enough to design for it.
The Agentic Economy already has the answer, and it is the staffing agency. The registry says who published the capability, which points to the provider. The hire-token says who hired it and what for, which points to the deployer, and the stated purpose shows whether the use was changed. The receipt is the timesheet, signed by both sides. Three signed records make "who is responsible for this decision?" much easier to answer.
What software should do: every hire records a provider, a deployer and a purpose, at the moment it happens.
Rule 2: Risk is decided at the moment of hire
The same scoring model is ordinary analytics when it flags a duplicate invoice. On my reading of the Act's high-risk list, the picture changes when a score is used to assess whether a person is creditworthy, or for risk assessment and pricing in life and health insurance. The list even draws lines inside a category: it carves fraud detection out of the credit item. The Act sorts by what the system is used for, not by what the component is.
With people, the deployer reads the instructions and decides. With agents, that decision has to happen inside the hire. So the manifest declares a risk class, the hire-token declares the purpose, and the registry checks one against the other. In the framework, a capability declared low-risk should not be hireable for a high-risk purpose, and a high-risk capability should only be hireable by a caller that takes on the deployer's duties, including human oversight.
The Act asks for one more thing: a provider who decides that a system on the high-risk list is not high-risk must document that assessment and register the system in the EU database. "We decided it was out of scope" leaves a record.
What software should do: risk class in the manifest, purpose in the hire-token, checked every time.
Rule 3: The receipt is now evidence
The Act never uses the word receipt. But as I read it, it describes one, over and over. High-risk systems must be able to record events automatically over their lifetime. Providers and deployers must keep those logs for a set period. In many high-risk cases, a person affected by a decision can ask for an explanation of the role the system played. Providers must report serious incidents within set time limits.
Put those together and what you get is the receipt from the definitions above. What went in, what came out, which rules applied, which versions ran, who hired it, for what, in which country, signed. The thing you settle on is now the thing you will be asked to produce.
So the first thing I would build this quarter is simple: send every receipt to write-once storage that cannot be edited or deleted. The major clouds offer this. Logs are not receipts. Logs are what you dig through when you do not have receipts.
What software should do: if your agents do not produce signed receipts today, start there.
Rule 4: Measure and release are now duties
In the Agentic Economy you hire a capability, measure it against the service level it promised, and release it if it falls short. I set this out in From SaaS to SaaC and in the episode Hire the AI Agent. As I read it, for high-risk systems the Act contains the same cycle. Providers must monitor their systems after launch. Deployers must monitor too. A provider who finds that a system does not meet the rules must correct it, withdraw it, disable it or recall it, and tell the deployers.
The difference is who moves first. In the Act, the provider and the regulator do. In the Agentic Economy, the receipts do: a missed service level shows up in the numbers, the hire is released, the registry's rating moves. The second way is faster, and it produces the evidence the first way asks for.
What software should do: measure every capability against its promise, and keep the evidence.
Rule 5: The label has to survive the chain
A customer asks an assistant for a letter. The assistant hires a drafting capability. The drafting capability hires a research capability. Only the first one talks to the customer. Only the last one wrote the words.
The Act's transparency rules are written for the moment a machine meets a person: say you are a machine, and label what you generate. On my reading, machines hiring machines do not need to announce themselves to each other, because the duty is owed to people. But the label on generated content must still be there when it reaches the customer, three steps later. The receipts carry it: if each hop's receipt points to the previous one, "which parts did a machine write?" is a lookup, not a guess.
What software should do: pass the label along in the receipts, hop by hop.
Rule 6: Borders are agreed at the moment of hire
As I understand the Act's scope, a capability published in Dallas and hired by an agent in Dublin is covered on the same terms as one published in Dublin. What matters is whether it is offered in the EU, or its output is used there. So the hire-token states the caller's country, the manifest states which regions the capability can run in, the hire is pinned to a region that fits, and every receipt records where it ran. "Did any of this leave the EU?" becomes a filter on the receipts.
What software should do: country on the manifest, country on the hire, country on every receipt.
What this means for the Agentic Economy
Three things stand out.
First, the law and the architecture ask for similar things. For high-risk AI, the Act wants a description of the system, records of how it ran, and a named owner. The Agentic Economy runs on a manifest, a receipt and a registry. The Act was not written around agents hiring agents, yet the two arrive at similar objects, because both are trying to make AI answerable for what it does.
Second, the cost of the Act falls on an architectural line, not an industry line. A company whose capabilities already produce signed receipts, declare a risk class and name an owner will, I expect, have a head start on much of the record-keeping the high-risk rules ask for, as a by-product of normal business. A company whose systems produce logs and whose contracts are PDFs has a reconstruction project, and the clock is already running.
Third, I expect Europe to become the first place where evidence has a price. An agent choosing between two similar capabilities will prefer the one whose hire carries less legal exposure for its owner. Registries will learn to rank on that. This is the audit-ready moat from The Four Moats, now with a law behind it.
Where my research goes next
These are the questions my next pieces of research will take on. Stay tuned.
- Who answers for a shared decision. When five capabilities from four companies in three countries share one decision, the Act sets out who counts as the provider. How responsibility is shared in practice, across so many parties, is still being worked out.
- How incidents reach the right people. The Act sets out who reports serious incidents. When receipts spot a problem in minutes and no person was involved, how it reaches the right provider quickly is a practical question.
- How duties combine. How the duties of a model provider combine with those of a capability publisher.
- How registration scales. Whether the EU database for high-risk systems can keep up with an economy of many small capabilities.
If this is your field, in law, policy, a regulator or a university, I would like to compare notes.
The small print
I am a technical architect, not a lawyer. Everything above is my reading of the Act as it stood on 18 October 2026, including the Digital Omnibus on AI (Regulation (EU) 2026/1744), which moved the dates for the high-risk rules later. It is written to work out what software should be doing. It is not legal advice, and other readers of the law may reach different conclusions. The Act has already been amended and may change again, and I have deliberately avoided quoting deadlines because they move. Check the duties and the current timetable against the official text before relying on them, and take advice on your own situation.
For readers who want to check my reading against the Act, here is the part of it each rule draws on:
| Rule, as I read it | The part of the Act it draws on |
|---|---|
| Someone is still the employer | The definitions of provider and deployer, and the rules on who becomes responsible along the supply chain |
| Risk is decided at the moment of hire | The rules for classifying high-risk AI, the list of high-risk uses, and the EU registration database |
| The receipt is evidence | The rules on record-keeping and keeping logs, the right to an explanation, and serious-incident reporting |
| Measure and release | The rules on monitoring after launch and on corrective action and withdrawal |
| The label survives the chain | The transparency rules for AI that talks to people or creates content |
| Borders at the moment of hire | The rules on who and what the Act covers |
| The fines | The penalties section |
| What the Act asks for | The prohibited practices, the high-risk list and its obligations, the transparency rules, and the amendment known as the Digital Omnibus on AI, which changed the timetable |
Important notice. This article is part of my ongoing research into the Agentic Economy and reflects my thinking at the time of writing. It is shared in good faith, for information and discussion. It is not legal, regulatory, technical, financial or other professional advice, and it is not a verdict on how the EU AI Act or any other law applies to any organisation, system or situation. Laws, guidance, technology and my own research all move on, and others may reach different conclusions. The examples, designs, diagrams and technology choices show one possible approach. Before relying on, adapting or applying any part of it in your own situation, you must test and verify it, design it for your own architecture, and take independent expert advice, including from qualified lawyers and from your own IT, security and data protection specialists. Any decision to act on it, and its consequences, rest with you. To the fullest extent permitted by law, the author accepts no responsibility or liability for any loss, damage or consequence arising from any use of, or reliance on, this article, which is provided without any warranty of accuracy, completeness or fitness for a particular purpose. Reading it does not create a professional or advisory relationship with the author. The views expressed are the author’s own.