The Operational Era, Priced and Enforced

In the two weeks to early August 2026, two structural changes arrived from opposite directions and pointed at the same conclusion. The market sharply compressed the cost of access to capable AI. The regulator activated the power to enforce accountability for how AI is used. Neither change is about which model a business chooses. Both are about the system a business builds around whichever model it chooses.

Consider the economics first. On 30 July, OpenAI reduced the price of its GPT-5.6 Luna model by 80 percent and its Terra model by 20 percent, three weeks after the family reached broad availability, a repricing confirmed by the company and reported by CNBC. The move followed a period of unusually intense performance-and-price competition among frontier laboratories through July, with cheaper open-weight alternatives applying steady downward pressure. The direction of travel is clear: inference costs are falling, capable models are converging in quality for many enterprise tasks, and buyers have more flexibility across providers.

Model selection still matters. A frontier model retains a real advantage on the hardest reasoning, coding and long-context work, and choosing well remains a genuine engineering decision. What is changing is durability. When several providers offer comparable capability at falling and volatile prices, the specific model a firm selected last quarter is a weaker foundation for lasting advantage than it appeared to be. The decisive variables move to what surrounds the model: routing between providers, evaluation, orchestration, and the data foundation underneath.

This is why the substitution test belongs in the boardroom rather than the engineering backlog. The question is simple and uncomfortable: if your primary AI provider withdrew on Monday, how much of your business would still be running by Friday. A firm that owns its evaluations, its logs, its orchestration and its data treats a price cut or a deprecation notice as a routine operating event. A firm that does not is exposed to every pricing change, access decision and outage its provider issues. In a market repricing this quickly, that exposure is an operating risk, not a technical footnote.

If your primary AI provider withdrew on Monday, how much of your business would still be running by Friday

The regulatory change runs in parallel and reinforces the same point. The EU AI Act, Regulation (EU) 2024/1689, entered into force in 2024 and phases in over several years. Two elements took effect on 2 August 2026. First, the European Commission, acting through its AI Office, became able to exercise its supervision and enforcement powers over providers of general-purpose AI models; the substantive obligations on those providers had already applied since August 2025, and models placed on the market before that date have until August 2027 to comply. Second, the Article 50 transparency obligations became applicable, covering disclosure that a user is interacting with an AI system, the marking of AI-generated or synthetic content, and the labelling of deepfakes.

The common belief that the AI Act was postponed is only partly correct, and the wrong part matters most to operators. The Digital Omnibus, proposed in November 2025 and agreed politically in 2026, deferred the heaviest regime, the high-risk obligations, to December 2027 for standalone Annex III systems and to August 2028 for AI embedded in regulated products. It did not defer the Article 50 transparency duties or the enforcement powers. Those landed on schedule.

The obligations do not fall uniformly on every business, and this distinction deserves care. Provider obligations attach to those who develop or place systems on the market; deployer obligations, including several transparency duties, attach to those who put systems into use, and the specific duty depends on the system type and the use case. Territorial reach extends beyond the EU: an organisation established elsewhere can fall within scope where it places a system on the EU market, or where the output of its AI is used within the EU. Penalties are tiered, reaching up to 35 million euros or 7 percent of worldwide annual turnover for prohibited practices, and up to 15 million euros or 3 percent for general-purpose AI providers and for most other obligations, including transparency. Whether any particular obligation applies to a particular firm is a legal determination that turns on legal role, system type, use case and territorial connection, and firms with material EU exposure should obtain specialist legal review rather than rely on a general reading.

Taken together, the market made access a less durable advantage, and the regulator removed the option of deploying AI without accountable governance. Both forces relocate advantage to the operating layer that a firm owns and controls.

The evidence on where firms actually struggle confirms that this layer, not access, is the binding constraint. In Domino Data Lab's 2026 research, return on AI still failed to outpace spend for 57 percent of enterprises, unchanged from the prior year, even as 93 percent reported improved production; the firm's own reading is that getting a model into production is no longer the milestone that matters, and that the gap has moved to the last mile between a model in production and a business user unlocking value from it. This converges with several independent studies through 2025 and 2026 finding that a large majority of AI pilots produced no measurable profit-and-loss impact. The consistent explanation is not model quality. It is that AI applied to a process designed for human hands produces little, because value is captured only when the surrounding process, governance and metrics are rebuilt around the capability. Stated plainly, most firms do not have an AI problem. They have a systems problem, and a better model cannot compensate for missing infrastructure.

Most firms do not have an AI problem. They have a systems problem, and a better model cannot compensate for missing infrastructure

The governance side of the same gap is equally documented. In a survey by Smarsh and FTI Consulting, 55 percent of enterprises were actively deploying AI while only 26 percent judged their governance frameworks fully aligned with the pace of that deployment. Deloitte's 2026 research found that a majority of organisations expect to run agentic AI within two years, while only a minority have mature governance for autonomous systems. Gartner, having forecast worldwide AI spending of 2.59 trillion dollars in 2026, separately projected that up to 234 billion dollars of enterprise application spending could be at risk from agentic AI through 2030. Firms are adopting far faster than they are governing, and that gap is now the dominant operational risk.

This is not confined to large technology enterprises. In property, often assumed to be a technology laggard, Buildium's 2026 research reported that AI adoption among property management firms rose from 20 percent to 58 percent in a single year, while only 8 percent had fully automated any process. Professional-services firms show the same signature: rapid tool adoption, negligible operational integration. The constraint is never access to the tool. It is the operating system that converts the tool into an outcome.

So what should a chief executive do differently. Not choose a better model. The higher-leverage actions concern the operating layer. Run the substitution test as a governance exercise and treat the answer as a risk metric. Own the portability layer, the evaluations, logs, orchestration and data, so that switching provider is a routine decision rather than a rebuild. Establish, if any material part of the business touches the EU, where transparency obligations now apply, and take specialist advice on legal role and scope. Redirect budget from access, which every competitor can now buy cheaply, to the architecture and governance that convert model output into measurable return. And redesign the process before automating it, because automation layered onto a workflow built for humans is the most reliable predictor of wasted spend in the current data.

The operational era does not mean the technology has settled. Capability is still improving, and often cheapening in the same week. It means something more precise and more useful for planning: the model layer has become a less durable source of advantage, so the advantage that compounds now sits in the operating system a firm builds around it. Access has become a procurement decision. Accountability has become a legal obligation. The operating system that governs both is the part a competitor cannot simply buy.

About the Author

Dale Anderson

Founder, Dale Anderson Consulting

Dale Anderson is the founder of Dale Anderson Consulting, an advisory practice operating from London and Cape Town. He has spent more than twenty years in UK and international real estate, with over £1.5 billion in transactions, and advises founders, property companies and professional-services firms on AI Operating Systems, business infrastructure and commercially disciplined growth.