Beyond the Obvious

Recent disclosures from OpenAI should concern enterprise leaders, but not because they prove frontier AI is inherently unsafe. The more important lesson is architectural.

As frontier AI moves from generating answers to operating tools, navigating systems and pursuing objectives, organisations are no longer simply consuming intelligence. They are potentially importing part of a frontier lab’s control environment into their own. That changes the risk.

A chatbot mainly creates an information problem. An agent connected to email, databases, cloud platforms, operational systems or payment rails creates an authority problem, and that speaks directly to the emerging role of the CISO as the Architect of Authority as I wrote about earlier.

The question is no longer simply ‘Can I trust the model?, it becomes ‘Who decides what the model is allowed to do, with whose identity, against which data, through which systems, and under what independently enforced controls?l

For most enterprises, the answer should not be the frontier lab. Nor should the reflex be to abandon frontier models and run everything locally.

Frontier services offer enormous capability, rapid innovation and infrastructure few enterprises could economically reproduce. Local models offer sovereignty, resilience and greater control. Both have value. The balanced architecture is therefore likely to be hybrid.

Use frontier models where capability matters and use local models where sovereignty matters, but most importantly of all place an enterprise controlled orchestration and policy layer (harness) around both wherever authority matters.

That harness becomes the trust boundary which controls identity, tool permissions, data classification, network egress, transaction policy, human approval and auditability. The model can propose an action. The enterprise control plane decides whether that action is permitted. This also addresses a deeper issue, notably the trustworthiness of frontier labs themselves.

Organisations should not need to assume that OpenAI, Anthropic, Google, Meta or any future provider will always behave perfectly, configure perfectly, update perfectly or even expose every relevant failure mode perfectly. Trustworthiness cannot simply be inherited from the vendor.it has to be engineered.

That is why the future should move from Explainable AI to Provable AI; not merely asking why a model made a decision but being able to prove that every consequential action passed through independent controls, used the right identity, stayed within policy and left an auditable record.

Trustworthiness is not merely a matter of confidence in how a system behaves; it demands demonstrable accountability which is what I have argued about extensively in my blogs – ‘The AI Risk Nobody Is Measuring’ An organisation must be able to reconstruct what happened, establish whether policy and delegated authority were respected, attribute responsibility and independently verify the integrity of the resulting evidence. These are not additional requirements beyond trustworthiness; they are fundamental to establishing and sustaining it.

Done properly, the model almost becomes irrelevant to transactional trust, it becomes powerful, replaceable and deliberately untrusted and perhaps that is the real enterprise lesson from frontier AI.

There is one final benefit to be noted. By not plumbing the enterprise directly into expensive frontier lab subscriptions, organisations have a chance to apply FinOps (Financial Operations)  discipline from day one, something many failed to do during the first wave of cloud transformation, only discovering later that convenience had quietly become dependency and consumption had become cost.

A properly designed AI control plane can govern not only identity, authority and data, but also model selection, token usage, workload routing and price-performance trade-offs. The enterprise can use frontier capability when it genuinely adds value, cheaper models when it does not and local inference where economics or sovereignty make more sense. Industry has already learned the hard way that unlimited consumption wrapped in an easy API eventually appears on somebody’s balance sheet. AI does not need to repeat cloud’s FinOps lesson at vastly greater speed and scale.

So my advice is the safest (and most prudent) AI architecture is not the one with the most trustworthy model. It is the one that remains trustworthy even when the model is not.