Why regulated industries still do not trust AI

What I kept seeing

I work on private cellular core networks at Druid Software, the kind of infrastructure that carries enterprise, IoT and public safety traffic. I also built internal AI tooling for engineering workflows there. Two things were true at once.

The technology was clearly useful. And nobody wanted it anywhere near a real system.

The objection was never that the model was not smart enough. It was simpler than that. If something goes wrong on a live network, someone has to explain it. You cannot explain a probability. You can explain a rule.

Falling behind is also a risk

The result is a strange kind of stalemate. Organisations in finance, healthcare and telecoms are watching AI change how work gets done, and are stuck in pilots because nobody will sign off on autonomy they cannot bound. Meanwhile competitors with less to lose move faster.

Caution here is rational, not backward. But the cost is real and it compounds. Every quarter spent in a pilot is a quarter of process improvement that did not happen.

Three objections I hear constantly

• The data is proprietary. Patient records, trade positions, subscriber data and network topology cannot leave the environment. A hosted guardrail service is a non-starter before the technical conversation begins.
• The behaviour is not verifiable. A model that behaves well in testing has demonstrated a pattern, not a guarantee. Risk teams are asked to accept the pattern as evidence, and they are right not to.
• Nobody can sign it off. Approval requires an individual to accept accountability. Without an artefact showing why an action was permitted, that signature is unsupported.

Notice that none of these are solved by a better model. They are solved by a different kind of evidence.

Local and verifiable, or not at all

The two words that change the conversation are local and verifiable.

Local means the enforcement runs inside your environment. Prompts, actions and records stay where they already are. There is no vendor holding your data, and no new transfer to justify to a data protection officer.

Verifiable means the boundary is proven rather than trained. Every proposed action is checked against a formalised policy before it executes. The same action against the same policy returns the same verdict, every time. A compromised or manipulated agent still cannot take an action the policy forbids, because enforcement sits outside the model's reach. Every decision is logged, so the auditor gets a record instead of an assurance.

This is also the direction regulation is moving. The EU AI Act expects record keeping, human oversight and robustness for higher risk uses. Those obligations are far easier to meet when your controls produce evidence as a by-product of running.

What I would tell a risk officer

Do not ask whether the model is safe. Ask what the model can do, and who guarantees it cannot do anything else. If the answer involves confidence, scoring or careful prompting, you have not been given a boundary. You have been given a hope.

Trust in this field will not come from better behaviour. It will come from proof.