Adding AI to a platform is easy, adding it without breaking your security isn't

There is a question every business handling sensitive information should be asking before it ships another AI feature: when we give this model access to our data, what stops that access from quietly becoming access to everything? 

In IBM's 2025 breach research, found that 63% of organisations had no AI governance policy in place, and that among those experiencing an AI-related security incident, 97% lacked proper AI access controls. That is the real story behind the current rush to adopt AI. Integrating a model into a product is no longer the difficult part. The real challenge is making AI effective for the work it is meant to support, while ensuring security, governance, and access controls are robust from the outset. 

You might ask what qualifies Ansarada to speak on AI integration and security. For more than two decades, we have operated as a secure system of record for high-stakes transactions. We have been ISO 27001 certified since 2009, and our platform is trusted to hold highly sensitive data across live M&A, capital raises, and government procurement processes, where a single leak can jeopardise a transaction or even a business. 

That experience has shaped how our systems are built. Security, governance, and control cannot be treated as optional or added later. 

With that in mind, I have set out some thoughts on the topic, with the aim of encouraging organisations to ask the same questions of any vendor they engage with. 

Should an AI tool ever see more than a person would? 

No. This is the principle most teams get wrong, and it is the one that matters most. The instinct, when you add a capable model, is to give it broad reach so it can be useful. That instinct is the vulnerability. 

Treat an AI tool the way you would treat a brilliant new hire. You do not hand a talented new employee the master key to every room in the building on day one because they are capable. You scope their access to the job in front of them. An AI assistant is the most capable new hire you will ever onboard, and the temptation to skip the onboarding, to grant it everything because it is so useful, is exactly how access becomes exposure. 

In practice, this means the model gets no privilege that the platform does not already control. Our in-room assistant operates only inside the single transaction it serves. It introduces no new attack surface over the platform it already lives on, and it has no reach into any other deal. The same least-privilege discipline that governs every person in a data room governs the model. When people talk about agentic AI in dealmaking, that confinement is the whole point: a capable agent that can only ever act within the four walls of one transaction. 

Is your customers' data training the model? 

It should not be, and you should be able to prove it. The most common quiet compromise in AI integration is letting customer data improve the model, because more data makes a better product and the commercial pull is strong. For anyone holding confidential information, it is also a betrayal of the entire promise. 

We do not train our models on the contents of customer data rooms. We use third-party large language models inside a controlled environment, and our contracts with those providers prevent them from training on our customers' data. The data stays where it belongs: stored in region, encrypted at rest, and moved across a border only for defined, contracted reasons the customer has agreed to. Data residency rules do not get waived because what is reading the document is an AI rather than a person. If anything, AI integration should raise the bar for adhering closely to your security guidelines. 

Does AI get its own, looser set of rules? 

This is the trap. The most dangerous idea in any organisation integrating AI is that AI is special enough to deserve its own governance, which in practice often can mean a newer, faster, lighter-touch regime running alongside the real one. 

Every AI capability we ship goes through the same risk, compliance and audit process as every other feature, measured against the same certifications we have passed for two decades. We have formally aligned our AI governance framework with ISO/IEC 42001, the international standard for AI management systems, precisely so the AI is held to a defined, audited standard. And every action the AI takes is captured in the same activity log as every human action, because when a regulator or a counterparty asks what happened during a transaction, "we believe" is not an answer. The log is the ultimate source of truth here. 

Are you being too cautious? 

In this business, you can never be too cautious. And I would say the same sentiment rings true for other, highly sensitive and highly regulated industries. The breach data referenced in my opening paragraph points to the risks of a rushed AI integration better than I can. I.e., the organisations that moved fastest by skipping access controls are the ones showing up in the incident statistics.