In secure boundaries

If someone cloned your voice this afternoon, could your finance team tell the difference? If a perfectly written email landed in your inbox requesting an urgent payment, would your processes catch it or comply with it?

If one of your employees quietly copied sensitive client information into their personal AI account, would you even know?

This is not science fiction. This is not five years away. The tools required to do all the above cost less than a monthly streaming subscription. And yet, the real security issue in most organisations isn't artificial intelligence, it's direction.

We've been here before. Cast your mind back to the early cloud days. Before Microsoft 365 matured into what it is now, staff were quietly signing up to Dropbox accounts and files drifted into their personal email accounts.

The IT department didn't approve it, but IT couldn't realistically prevent it either.

Innovation sprinted ahead. Governance jogged behind. Eventually, the products stabilised. Microsoft 365 became administratively robust and compliant, Dropbox and Google matured, governance caught up with product maturity and centralised control became normal.

We are now standing in that exact same moment again, but this time with AI. The pace of innovation is outstripping policy. New models, new integrations, new capabilities, companies embedding AI into everyday consumer devices.

If you already operate within Microsoft 365, Copilot becomes the natural gravitational centre. It respects existing permissions, it inherits compliance policies and sits inside your controlled environment.

If you don't set the direction, one of two outcomes follows. Best case: staff gravitate towards Copilot Chat anyway because it's convenient. More likely: shadow AI use from all the copious alternatives out there, with paid-for consumer AI outperforming the free versions of business tools.

For under £20 per month, anyone can subscribe to ChatGPT Plus personally. They will use it to draft emails, solve problems, summarise documents and augment themselves. Augmentation is not the problem. Data location and ownership is.

When confidential emails, client details or sensitive commercial information are copied into personal AI accounts, organisational control weakens. When that employee leaves, so might fragments of your operational knowledge.

Voice lowers the barrier further. Transcription is now accurate enough that sensitive discussions can simply be spoken aloud into an AI application. Intellectual property becomes text in seconds.

Context matters. Asking a general process question to your favourite AI overlord is different from talking to it about personal data. But without explicit boundaries, those lines blur quickly.

Security does not stop at data handling. Artificial intelligence is also a deception multiplier. It can generate persuasive text, fabricate convincing images, synthesise video and clone voices.

The technology required to imitate a senior executive's tone is no longer reserved for nation states. It is commercially accessible. Now.

That's why 'guard', the fourth pillar of the A2G2 framework I introduced previously, is essential. Guard means guarding against others' use of AI against your business processes, and that starts with making sure that every login you have has a secure password and is protected by multi-factor authentication.

This is, and has long been, an existing IT threat, but with the dawn of AI, social media accounts should not be protected by passwords sitting in someone's notes app.

It would not take much for a poorly secured social account to be compromised and used to publish misinformation under your own brand, with your own logo and brand voice, sounding just like you, using AI.

Recruitment processes may need refinement. If text and interview responses can be AI-assisted, verification regains importance – face-to-face interaction, structured questioning, critical thinking exercises. Assume tools are being used and design accordingly.

Guard also means strengthening verification culture internally. Urgent payment requests should trigger secondary confirmation. Sensitive instructions should be validated through alternative channels. Immediately is a word that needs removing from our vocabulary.

This is not paranoia, it is proportionate adaptation for those who don't want to get caught out.

Innovation will not slow down, product maturity will improve, governance will catch up, as it did with cloud software. But until then, leadership matters.

Artificial intelligence does not automatically undermine security, but lack of direction does.

Dealt with deliberately, AI can operate inside a mature, controlled and well-understood environment. The organisations that succeed will not be those that retreat from AI out of fear, but those who adopt it with a 'this is what we are doing and how we are doing it' mindset.

The tools are here to stay. What are your rules?