AI concerns shouldn't extend to the enterprise agent

AI can do remarkable things. It’s increasingly capable of handling ambiguous or complex tasks end to end, but given a complex task with a clear objective, it will do whatever it can to achieve that objective.

As a result of some well-publicised, unintended methods of the models achieving an objective, concerns — even from the heads of the leading frontier model companies themselves — about continued development have been expressed.

As calls for regulation grow, what I’d prefer to see first is cooperation between AI companies.

Dario Amodei, for one, has proposed independent evaluators with access to Anthropic’s models, and Sam Altman pledged OpenAI would adopt the same approach. With these in mind, the heads of the frontier labs could agree on a basic set of guidelines every model must follow, shared red lines and disclosures when something goes awry.

These are not unsolvable problems, and this is not a binary choice between pumping the brakes and sprinting blindly. It’s a bit more nuanced than that, but that nuance should extend to understanding the difference between consumer-grade AI and the controlled environments of enterprise AI. Deployed correctly, with the right guardrails in place and within an organisation’s own governed trust boundary, enterprise AI and, increasingly, agentic AI, can continue to deliver on its promise.

Enterprise applications and agents run on their own data, which is kept away from the public, governed and, importantly, manageable. In effect, an agent built on an enterprise’s own foundation is limited to the virtual walls of the business environment and the guardrails set by the organisation. These parameters are constantly checked and maintained.

Tellingly, several of this year’s incidents happened when test environments were misconfigured and gave models internet access they were never meant to have. Enterprises have every reason to keep those walls up: they don’t want to share their IP, and unlike consumer products, their internal agents can’t simply be downloaded and tested by anyone.

These are practical reasons for enterprises to build governed, secure and proprietary applications for their own purposes.

Given the restricted scope, AI guardrails are easier to implement in an agentic enterprise environment. For example, a guardrail can stop an agent from acting on malicious instructions a staff member has unknowingly copied from a website.

Agent policies can also set, uniformly across the enterprise, what each agent is allowed to access and do, and limit each session to that scope. Agentic observability, meanwhile, helps companies see how AI agents are being used, and detect and prevent malicious use.

Underpinning all of this is the semantic layer (what many now call the intelligence layer) that translates an organisation’s raw data into shared business definitions, metrics and relationships that agents can understand. Rather than guessing what ‘revenue’ or ‘active customer’ means, an agent draws on a single, governed source of truth, so its answers are consistent, explainable and grounded in how the business actually operates. It’s what turns a capable model into a trustworthy one, and it means the guardrails govern not just what an agent can reach, but how it interprets what it finds.

The safety concerns around AI aren’t without merit, and the frontier models, to their credit, are cognisant of the risks. Whether boundaries come through voluntary industry cooperation on shared red lines, or through reactive government regulation, is an open question.

Enterprise AI is a distinctly different proposition however, and we need to make sure the public’s concerns don’t have a chilling effect on legitimate business use and innovation.

Image credit: iStock.com/Orhan Turan