AI Security in Professional Services

Professional firms hold information whose confidentiality is not merely commercial but often a professional and legal obligation. AI introduces new ways for that information to be exposed, layered on top of the security concerns firms already manage. Understanding these new exposures — without either dismissing them or being paralysed by them — is now part of running a firm responsibly.

The risks are not exotic, but they are different in character from familiar cyber threats. They arise from the way AI tools consume information, from the difficulty of controlling what users paste into them, and from the novel ways in which these systems can be manipulated.

Data leakage is the everyday risk

The most common exposure is mundane: confidential material entered into a tool that retains it, processes it outside the firm, or uses it to improve a provider’s models. This rarely results from malice; it results from convenience, as people reach for the easiest tool to hand. The mitigation is a combination of approved tools with appropriate contractual terms, clear guidance, and technical controls that make the safe path the easy one.

New attack surfaces

AI systems introduce genuinely new vulnerabilities. Prompt injection — where malicious instructions are hidden in content the system processes — can cause an assistant to behave in ways its owners never intended. Systems connected to a firm’s data can be induced to retrieve or reveal more than they should. These require security thinking that traditional controls were not designed to address.

• Which AI tools are approved, and are their data-handling terms genuinely fit for confidential work?
• How do we prevent confidential material from reaching unapproved tools?
• Where AI is connected to firm data, how is retrieval constrained and monitored?
• How would we detect and respond to shadow AI usage across the firm?

The risks rarely result from malice. They result from convenience, as people reach for the easiest tool to hand.

Shadow usage is the blind spot

The exposure that most often goes unmanaged is the AI a firm does not know it is using. When approved tools are cumbersome or unavailable, people turn to consumer alternatives, and confidential information follows them. Addressing this is as much about providing good, approved options as about prohibition; people route around controls that get in the way of their work.

Proportionate, not paralysed

The answer is not to ban AI, which simply drives usage underground, but to secure it: approved tools, sensible controls, clear guidance and the ability to see what is happening. Handled proportionately, AI security is an extension of the confidentiality discipline firms already understand — and demonstrating that discipline is increasingly something clients and regulators expect to see.