Don't marry your business to one AI model

There’s a worst case for AI regulation, and it isn’t necessarily the strict version. It’s the unpredictable version — no published rulebook to plan around, just the standing chance that the model your business depends on disappears on a Tuesday, with the explanation arriving afterwards.

That’s roughly where the United States has just landed. And it’s a bigger problem for a ten-person firm than for the AI labs the rules are aimed at.

Three Financial Times articles over the past week paint a remarkably consistent picture.

First, OpenAI’s chief executive, Sam Altman, used the FT’s opinion pages to argue for a US-led international body to oversee frontier AI. His proposal resembles an International Atomic Energy Agency for artificial intelligence: countries would agree to common safety standards, companies would be certified against them, and access to the most advanced models would depend on compliance. Whether you agree with him or not, the underlying message is clear — if AI is becoming critical infrastructure, it needs a predictable rulebook.

The FT’s Lex column accepted the diagnosis but doubted the prescription. The closest historical comparison, it argued, is the Basel Committee on Banking Supervision — which made banking safer, but also demonstrates two uncomfortable truths: international standards take years to emerge, and complex regulation almost always favours the largest incumbents, because big organisations can afford compliance departments and smaller competitors often can’t. Any global AI framework would also face a reality banking largely avoided: China. A genuinely global agreement would need Beijing’s participation; without it, the world risks splitting into competing AI blocs — different standards, different models, restricted access between them.

Then came the most revealing article of all. The Trump administration’s departing AI adviser insisted there would be “no FDA for AI” — no formal licensing regime, no central regulator, no bureaucracy deciding which models reach the market. Yet in the same interview he confirmed that the White House had recently used emergency powers to force Anthropic to withdraw its most capable model and hold up OpenAI’s next release on national security grounds, while introducing a framework that gives the government time to review frontier models before deployment.

Read together, those positions expose the real issue. America hasn’t chosen deregulation, and it hasn’t built a comprehensive regulatory system. It’s regulating frontier AI through discretionary intervention — and that distinction matters.

Markets cope surprisingly well with rules. Businesses complain about regulation, but they’re remarkably good at adapting to it: give a company a clear constraint — even an expensive one — and it will budget for it, redesign its processes and move on. What businesses can’t cope with is uncertainty. Uncertainty can’t be priced, can’t be scheduled, and can’t be built into next year’s operating plan. It becomes its own tax.

That isn’t a figure of speech; it’s one of the better-measured effects in economics. Faced with an unclear future, firms don’t panic — they wait. The Stanford economist Nicholas Bloom showed in 2009 that when uncertainty spikes, companies temporarily freeze investment and hiring: with the picture unresolved, the rational move is to sit on your hands until it clears. He and colleagues later built an Economic Policy Uncertainty index from newspaper coverage and found that, in the United States, a jump in policy uncertainty of the size seen between 2006 and 2012 foreshadowed roughly a 6% fall in business investment. The uncertainty didn’t have to be bad news. It just had to be unresolved.

Britain has a fresher example, and a cautionary one. In the three years after the 2016 referendum — before a single trading rule had actually changed — Brexit uncertainty alone reduced UK business investment by around 11%, according to work by Bloom and colleagues using the Bank of England’s Decision Maker Panel survey of firms. Not a recession, not a new regulation — just years of not knowing the rules, and companies quietly deciding to wait. Whatever your politics, the economics are the point: the cost wasn’t the eventual outcome, it was the fog on the way there. That same fog is now forming around AI, only faster.

Path	Business investment (indexed, expected = 100)
Expected — without Brexit uncertainty	100
Actual — three years after the 2016 vote	≈ 89 (about 11% lower)

For the frontier labs, that fog is a lobbying problem. For small and medium-sized businesses, it’s an operational one.

Most SMEs don’t think they’ve built their business on AI. They think they’ve bought some software. Increasingly, that’s no longer true. If your CRM drafts emails using GPT, your support desk summarises tickets using Claude, your accountants review documents with Gemini, or your developers rely on Claude Code, then part of your operating model already depends on decisions made by organisations — and increasingly governments — thousands of miles away. Most days, that’s perfectly acceptable. One day, it may not be.

And this isn’t only about politics. Models are retired, APIs change, pricing changes, performance rankings shift every few months, new open-weight models appear, and providers merge or discontinue products. Government intervention simply adds another source of volatility to a stack that’s already moving at extraordinary speed.

So the mistake many businesses make is assuming the important question is which AI model they should choose. It isn’t. The better question is: how easily can we stop using it?

That might sound counterintuitive, but it’s how resilient businesses have been built for decades. Manufacturers diversify supply chains. Finance teams manage counterparty risk. IT departments avoid unnecessary vendor lock-in — because suppliers change prices, discontinue products, or occasionally fail altogether. AI deserves exactly the same discipline.

The emerging best practice isn’t to build directly against one provider’s model. It’s to put an orchestration layer between your business processes and whichever models sit behind them. That layer decides which model performs which task; it lets providers be swapped with minimal disruption; it keeps you running if one service goes dark; it can move sensitive work onto privately hosted open-weight models where appropriate; and it cuts cost by routing simple jobs to inexpensive models while reserving premium ones for work that genuinely needs them. In short, it turns changing AI providers from a redevelopment project into a configuration change — which is worth doing even if governments never intervene again.

The events of the past week simply make the point harder to ignore. You don’t need to predict whether Washington, Brussels or Beijing will regulate AI next year, or whether OpenAI, Anthropic, Google or an open-weight model leads the market in 2028. You only need to accept that the answer will change.

The businesses that benefit most from AI over the next decade won’t be the ones that picked the “winning” model in 2026. They’ll be the ones that designed their systems so they never had to care. Technology changes, politics changes, markets change — your business architecture should assume all three.

If you’re not sure where your business is quietly depending on a single AI provider, that’s exactly the sort of risk a free process audit should uncover — before the next Tuesday headline becomes Wednesday morning’s operational problem.

---

Sources

• Investment-under-uncertainty evidence (primary): N. Bloom, “The Impact of Uncertainty Shocks,” Econometrica (2009); S. Baker, N. Bloom & S. Davis, “Measuring Economic Policy Uncertainty,” Quarterly Journal of Economics (2016) — US data, index built from newspaper coverage; N. Bloom et al., “The Impact of Brexit on UK Firms,” NBER (2019), using the Bank of England’s Decision Maker Panel survey of UK firms (~11% lower investment over the three years to 2019).
• Regulatory developments as reported and argued in the Financial Times: Sam Altman’s op-ed (1 July 2026); the Lex column “Altman’s AI safety proposal: let us win, or everybody loses” (2 July 2026); and the interview with the departing White House AI adviser (3 July 2026).