Back to blog
agents-ia
November 15, 2026
8 min read

AI Regulation: What the Government Can Switch Off in Your Small Business

AI regulation doesn't hit small businesses through the law, it hits them through their vendors. The Fable 5 case, US lobbying and a 5-step plan.

Vincent

Vincent

AI expert, AI First

AI regulation and government: the real risk for an SMB is hidden dependence on a single model. Fable 5, the AI Act, lobbying and a 5-step action plan.
Contents
  1. What does AI regulation change for a small business?
  2. Why can a government decision cut off your AI tool in 3 days?
  3. Who actually writes the rules on AI?
  4. What does compliance really cost a small organisation?
  5. What action plan can a small business start this week?
  6. My verdict: should small business owners fear AI regulation?
  7. FAQ

Government regulation of AI affects a small business in two ways: through the laws it has to comply with, and through decisions a government can make about a model the business uses without even knowing it. The second is the more dangerous, because no contract covers it. I went through four recent sources to separate what concerns you from what doesn't.

  • ⚠️ Hidden risk: a government decision can cut off a model within 72 hours.
  • 🏗️ Invisible dependence: your software vendor picks the AI model for you.
  • 📊 Targeted laws: the toughest rules aim at AI labs, not the small businesses using their tools.
  • ✅ Simple safeguard: a register of your AI uses and a second model ready to take over.

Here is my argument: the point isn't to read every piece of legislation, it's to avoid building your automations on a single opaque vendor. I'll start with what the law actually requires of you, then the case that shows where the real risk lies, then who writes these rules, and finish with an action plan.

What does AI regulation change for a small business?

AI regulation is the body of rules (European, national, regional) governing how AI systems are designed and used. For a French SMB that uses off-the-shelf tools without training its own model, the concrete obligations boil down to data protection, transparency and staff training. The heaviest rules are aimed elsewhere.

Who is covered by which law?

In the European Union, the AI Regulation (AI Act) is being phased in. According to the European Commission, the obligation to train staff in AI and the ban on certain practices have applied since 2 February 2025. A small business rolling out an assistant or an agent acts as a "deployer", not as a model provider.

In the United States, it's the other way round. Alex Bores, a New York State Assembly member and congressional candidate, pointed out on the show Breaking Points that the states are the only level of government actually regulating AI. His New York law targets developers of frontier models (he names Anthropic, OpenAI, Google and Meta): they must publish a safety plan and report critical incidents. Small businesses don't feature in it at all.

For the details of your obligations in France, my article on GDPR and the AI Act for SMBs covers what a team can and can't do.

Why isn't reading the legislation enough?

Because the law tells you what you must do, not what you can be prevented from doing. A transparency rule won't protect you from an export directive that cuts off a tool overnight. That is exactly what happened, and it's the heart of this article.

Why can a government decision cut off your AI tool in 3 days?

A government can disable an AI model by simple directive, without the end user ever being consulted. According to the Business of Tech podcast (MSP Radio), the US government ordered Anthropic, through a Commerce Department export-control directive, to cut off access to its Fable 5 and Mythos 5 models for all foreign nationals, 72 hours after launch.

What exactly happened?

Anthropic launched Fable 5, billed as its most capable model. According to the company's published benchmarks, it outperformed OpenAI's GPT 5.5 across the board. Three days later, it was gone for non-Americans. Anthropic complied with the order while publicly challenging it, arguing that the risk cited was limited and already present in other tools.

I'm not taking sides on whether the directive was justified. What matters for a small business owner is the mechanism: a frontier model became unavailable to a whole category of users, by political decision, over a single weekend. If you want to know whether the model was worth the investment before this episode, I wrote a review of Claude Fable 5 for SMBs.

How can a small business be hit without ever choosing this model?

The podcast lays out the mechanics precisely: vendors no longer sell you a model, they sell you a feature. Pax8 built a tool into its platform for managing Microsoft 365 and Copilot for its SMB customers. CloudRadial launched an agent that triages tickets before a human even looks at the queue. Proofpoint plugged its compliance product into Claude's compliance API.

The model disappears inside the product. You aren't buying Claude or GPT, you're buying an "AI-powered support service". And when the model goes down, the outage lands on you with no warning and no clause in your contract.

The same episode contains a detail that should worry you if you work with a team abroad: the measure targeted foreign nationals. A support team in Latin America or analysts in India whose workflow depends on that model, even hidden inside a third-party product, lose their tool overnight.

You don't depend on a model, you depend on a vendor whose own vendor you don't know.

How many companies are ready for this?

Very few. According to the study cited by MSP Radio, Netriox surveyed 401 IT leaders at mid-sized companies: 82% already use AI in production, but only 26% have deployed and managed it across the whole organisation. More than eight in ten use it, fewer than three in ten actually manage it.

That's the gap I see in my training sessions: teams that use AI every day, and nobody who can say which tool depends on which model.

Who actually writes the rules on AI?

AI rules are being written under heavy lobbying pressure, especially in the United States, where money from the labs and their investors weighs on elections. For a small business, that means the regulatory framework will stay unstable, so you need to build automations that can survive a change in the rules.

Why does lobbying matter to you?

On the same show, Alex Bores describes narrowly losing his congressional race after AI-linked super PACs spent around $8 million against him, having initially pledged at least $10 million. Among the major donors he names Marc Andreessen, Ben Horowitz and Greg Brockman, president of OpenAI.

He also describes a split between lab employees, who in his view mostly want regulation, and a small handful of executives who want none at all. I'd treat this with caution: it's the account of a politician who lost to these groups, so it's a partisan source. But the amounts and the names are verifiable facts, and they explain why the rules change so quickly.

Can the rules change just as fast on the public-sector side?

Yes, and the video furthest from our topic offers a useful illustration. Consultant Eric Coffie explains that in September 2025 the US federal government changed a procurement rule that had stood for decades (FAR 19.108-11), opening doors to companies that had previously been shut out. He adds that nearly 90% of companies ignore open consultations ("sources sought" notices), even though responding lets you shape the requirements.

The lesson carries over: those who respond to public consultations on AI (the CNIL, the European Commission, trade associations) shape the rules, everyone else just lives with them. Your industry association often circulates these questionnaires.

What does compliance really cost a small organisation?

Compliance costs weigh more heavily on a small business than on a large group, because it has no dedicated legal department. The pitfalls to watch for are administrative errors penalised automatically and dependence on a software subscription to keep equipment compliant.

What does the Singapore case show?

Dennis Tan, a Workers' Party MP, backed an energy efficiency bill in Singapore's Parliament while warning about the burden on SMEs. A product listing automatically pulled over a documentation error can cost sales and search ranking, and SMEs don't have the compliance teams that multinationals do. It isn't an AI law, but the mechanism applies to any regulation enforced by algorithms.

More relevant for us, he highlights "software-defined" energy. AI-driven optimisation systems, such as ABB's, can revert to high baseline consumption if the business stops paying the subscription. He also cites a figure from the US Department of Energy: a chiller running in its default fail-safe mode uses 30 to 40% more energy without alerting the owner.

How can a small business's exposure be summed up?

The table below lists the five risks raised in these sources, with a safeguard for each. I've only put numbers where the sources do.

Risk Example cited SMB exposure Safeguard
European law AI Act, staff training from February 2025 Certain, moderate Register of AI uses
Directive targeting a model Fable 5 cut off within 72 h High if you have a team abroad Tested backup model
Hidden dependence Pax8, CloudRadial, Proofpoint High, invisible Clause in the contract
State laws on models New York law (Bores) Indirect Keep track of your vendors
Subscription that gates usage ABB energy optimisation Medium Demand reversibility

SOURCE: cited transcripts (Business of Tech, Breaking Points, The Online Citizen) · UPDATED 11/2026

What action plan can a small business start this week?

The plan comes down to five steps, all doable without a lawyer: list your AI uses, ask each vendor which model runs underneath, line up a backup model, train your teams, and write reversibility into your contracts. None of them requires an 18-month transformation project.

What are the five concrete steps?

  1. Map your AI uses: tool, task, data processed, team involved. One page is enough. It also forms the basis of your training obligation.
  2. Ask your vendors: "Which model runs this feature, and what happens if it becomes unavailable to part of my team?" The answer, or the lack of one, tells you a lot.
  3. Test a backup model: I cover this in my Mistral vs Claude comparison, since a European provider changes your risk profile.
  4. Train your teams: it's the only European obligation already in force that concerns you directly.
  5. Demand reversibility: exportable data, and features that don't vanish when the subscription ends.

On the underlying logic, digital sovereignty and AI for SMBs answers the question "should you bring everything back in-house?" (my answer: no, above all you should diversify).

Should you wait for regulation to settle down?

No, and this is where I part ways with the prevailing caution. Waiting means keeping slow processes and high costs while your competitors automate. The companies that come out ahead build AI into their actual operations, with a measurable use case, and stay in control of the architecture. To measure what you gain, see AI ROI for SMBs.

If you build software with technical teams, GoLive Software, our sister company, supports projects where this architecture question comes up right from the specification stage: golivesoftware.co.

My verdict: should small business owners fear AI regulation?

No. Fear dependence, not regulation. The rules that apply to you (AI Act, GDPR) are manageable with a register and some training, whereas a model being cut off by government decision, as in the Fable 5 episode, can bring a service to a halt within 72 hours with nothing in your contract to cover it.

My conviction, drawn from what I see in the small businesses I train: the value lies in integration with your processes, not in the model. A replaceable model is therefore an asset, and a tool that only works with one model is a weakness. The four sources agree on one point I'll take away: the rules change fast, and without warning.

FAQ

What is the difference between AI regulation and a government decision about a model?

Regulation is a lasting text that sets obligations (AI Act, GDPR). A government decision is a one-off act, such as an export-control directive, that can make a model unavailable to certain users. You prepare for the first with compliance, and for the second with a backup plan.

Does a French small business have to comply with the AI Act?

Yes, in its role as a deployer of AI tools. According to the European Commission, the obligation to train staff in AI has applied since 2 February 2025, with further obligations phased in according to the system's risk level. Check the latest timeline on the Commission's page before making any decision.

How can I find out which AI model runs inside my software?

Ask the vendor in writing: model name, hosting, and what happens if the model becomes unavailable. Many features sold as an "AI assistant" rely on a third-party model, which the Business of Tech podcast calls a repainted dependency. If the vendor doesn't answer, treat the risk as uncontrolled.

Do US AI laws affect my business?

Indirectly. The New York law described by Alex Bores targets developers of frontier models, not their customers. It can still affect the terms of use and availability of the models your vendors build in, which is why it's worth keeping an eye on what your suppliers say.

Is a backup model expensive to set up?

Not necessarily. It mostly means testing an alternative on your three to five main tasks, with reusable prompts, before you need it. The investment is testing time, not a second permanent subscription.

Take action with AI First

Let our developers build your website, application or automation with Claude, from brief to delivery.

Tell us about your project →

Read next