When a Frontier Model Can Be Switched Off Overnight, Value Is the Only Hedge
It may seem counterintuitive, but an uncertain regulatory climate is a reason to build AI, not a reason to wait, so long as you build the kind that can prove its value rather than just claim it, for the business and the customer alike. That discipline is worth having whichever way the rules fall.
In June, one of the most powerful AI models in the world stopped working. It had not failed, and it had not run out of compute. A government had told it to switch off.
Anthropic’s top tier, the Mythos and Fable models, went dark for every user everywhere after the United States issued an export-control order. There was no wind-down and no notice for the companies that had already wired the models into live workflows. One letter from Washington, and a capability businesses were building on was gone by the evening.
That is the world UK firms are now investing in. The most advanced tools can be granted or withdrawn by a government, and the rules around them are still being written.
The UK has chosen uncertainty on purpose
Britain has decided not to pass an AI law.
Where the EU wrote a long, risk-based statute, the UK kept its existing regulators and gave them five principles to apply: safety, transparency, fairness, accountability, and the right to challenge a decision. The Information Commissioner handles anything that touches personal data. The Financial Conduct Authority leans on the rules it already has, chiefly the Consumer Duty. The competition and communications regulators cover the rest. The cross-sector AI Bill that many expected has been set aside, and ministers now prefer to act case by case. The newest measure, an AI Growth Lab that lets firms test AI in live markets under relaxed rules, was launched with an open admission that the rules today lack clarity.
There is a real case for this. Writing detailed law around a technology that changes every quarter risks locking in rules that are useless within a year. But the uncertainty does not vanish. It moves onto the companies that have to commit money now, without knowing where the rules will land. And the sums are not small: the four largest cloud providers alone are sitting on more than two trillion dollars of contracted, not-yet-delivered AI and cloud business, spending that customers have already promised.
Nor is the uncertainty only British. The EU’s AI Act reaches across the Channel: a UK firm whose AI is sold into the EU, or whose outputs are used by people there, falls under it wherever the firm sits, with fines reaching seven percent of worldwide turnover. The UK has meanwhile loosened its own rules on automated decisions, so a company trading on both sides now answers to two sets of requirements that no longer match.
The wider picture: fragmentation, and a new kind of control
Look wider and the pattern is divergence. Europe regulates hard, Britain regulates lightly, and the United States has swung back toward almost no federal rules at all, while individual states, California and Colorado among them, write their own. The treaties meant to pull these together stay voluntary.
The Mythos shutdown points to the bigger shift. Governments are starting to treat frontier AI the way they treat advanced chips and weapons parts: as a strategic technology whose export they control. The capability itself has become a matter of national security, switched on or off at a government’s discretion. A company that runs its core operations on a single such model has handed a foreign authority a veto over its work.
Uncertainty makes firms behave badly
Faced with rules they cannot read, companies tend to make one of two mistakes.
Some freeze. They wait for the rules to settle before committing, and while they wait, the competitors who kept building pull further ahead.
Others overreach. They buy the most powerful model available and push it into everything, mistaking activity for progress. That is the Mythos trap in miniature: a critical process resting on one capability that someone else can take away overnight.
Neither is a plan. Both confuse doing things with AI for getting value from it.
What regulators want is what value needs
It is tempting to read all this as the reason AI disappoints: the rules are unsettled, the ground keeps shifting, the best model can be switched off overnight. Clear those obstacles, the thinking goes, and the value will follow. It will not….
Most AI already fails in conditions far more open than these, and rarely because the technology fell short. The thing standing between a firm and the value of its AI was never the environment. It is whether the firm can show its AI did the work. And the discipline that takes is, of all things, what a regulator now insists on.
The demands a regulator makes look, at first, like a burden. Assess the risks. Document the data. Keep a person in the loop. Be able to explain why the system did what it did. Name who is accountable. Those are the same habits that separate the few AI projects that pay off from the many that do not.
The failure rate is the part people underplay. Gartner found that fewer than three in ten enterprise AI projects fully meet the return their sponsors expected, and one in five fail outright. MIT put the failure rate for corporate generative-AI efforts at around ninety-five percent. McKinsey found that almost every company now uses AI, yet only about four in ten can point to any effect on profit.
The cause is rarely the model. It is missing data, an outcome no one defined, and a pilot that was never tied to a number anyone could check. Strip away the detail and the failed projects share one flaw: nobody can show that the AI caused the result it claims. They can show that the AI was in the room when the result happened. They cannot show that it did the work.
That gap, between something being present and something being its cause, is the whole problem of proving value. Closing it takes evidence: a record of what the system did, and what followed from it.
A regulator asks for that same record, from the other side. To justify an automated decision you have to show what the model did and why. The compliance officer and the finance director are not asking the same question, but underneath it they want the same thing: proof that you can account for your AI, not a claim that it works. Build that proof for one, and you have most of it for the other.
Trust is where the discipline pays
There is a second return, and customers are the ones who deliver it. AI that is governed well is AI people can trust, and trust is what makes them use it. PwC found that around six in ten executives say responsible AI improves both their returns and their customers’ experience. In a bank, using AI carefully in fraud checks or customer service is not only the regulator’s demand under the Consumer Duty. It is what turns a wary customer into a loyal one. The sequence is short: careful work earns trust, trust drives use, and use is the only place value was ever going to come from.
What this means for deciding what to build
So what should a company do while the rules keep moving?
Start every AI project with the question a causal analyst asks first: what would have happened anyway, without this? If you cannot answer it, you cannot prove the project worked, and you will struggle to defend it to a regulator either.
Never let one model or one supplier become a single point of failure. The Mythos shutdown is the warning. Keep an alternative ready, and build so you can switch.
And treat the record-keeping of governance, the risk register, the named owner, the tracked result, not as a tax but as the evidence you will want the day a regulator, a finance director, or a customer asks you to prove the value is real.
AI is a means, not an end. The rules will keep shifting, in Britain and everywhere else, and no company should wait for them to settle. The question worth building around does not shift with them: where will AI create real value, for the business and for the customer, and can you show it caused that value rather than just happening to be there when it arrived?
Answer that, and an uncertain regulatory climate stops being a reason to hold back. It becomes the discipline that was worth having all along.
Sara Tiron is Causal AI Advisor at FernAI, where she brings econometric rigour to closing the gap between what AI does and the value a firm can prove it creates.
