5 min read AI Process Improvement

Your AI Walled Garden Has a Thousand Human-Shaped Backdoors

As leaders, we’ve rightly focused on building secure ‘walled gardens’ for AI. We are investing in private instances, drafting acceptable use policies, and running training sessions.

And we need to accept a difficult truth: for a significant portion of our data, it is already too late. The walls have been breached, not by hackers, but by our own well-intentioned staff using their personal ChatGPT, Gemini, or Claude accounts.

The Reality of the Everyday Leak

This isn’t happening in dramatic, high-level breaches. It is happening in the mundane, everyday tasks of modern work, across all departments. Think of:

  • The Editor, pasting a sensitive manuscript from a high-profile author and asking, “Suggest structural edits and check for plot holes.”

  • The Sales Executive, pasting a long email chain with a client, including commercial terms, asking, “Help me draft a persuasive reply to close this deal.”

  • The Rights Manager, uploading a complex international rights agreement and prompting, “Summarise the key restrictions and royalty clauses in plain English.”

  • The Junior Developer, pasting proprietary source code and asking, “Can you find the bug in this code?”

  • The Disgruntled Employee, pasting their entire project history, including confidential metrics, and asking, “Help me write a CV that shows my impact.”

The precedent is no longer theoretical

This risk was made stunningly clear in the last few days. News outlets like TechRadar reported that thousands of private ChatGPT conversations were appearing in Google search results. The issue stemmed from a feature allowing users to make their “Shared” chats discoverable. The backlash was so swift that OpenAI has already removed the feature, stating it “introduced too many opportunities for folks to accidentally share things they didn’t intend to.” The incident is a perfect, real-world example of how platform features and user behaviour can instantly nullify corporate data policies.

The Leadership Challenge: From BYOD to BYOAI

Remember the arguments and battles for and against ‘Bring Your Own Device’ (BYOD)?

Some organisations are still in that place (I feel you). We are now at the start of a much bigger, faster shift: ‘Bring Your Own AI’ (BYOAI). With Meta focusing on building deeply personal AI assistants, the line between an employee’s personal AI and their work life is set to evaporate.

The old approach of prohibition is obsolete. It leads to the attribution trap: trying to discipline employees is technically difficult and, in the UK, legally perilous. 

Leaders need a more sophisticated, four-pronged strategy:

1. Accept Reality: Target Harm Reduction, Not Prevention. The first step is to accept that you cannot stop this behaviour. The strategic goal must shift from a zero-tolerance policy of prevention to a pragmatic policy of harm reduction. This is where building governance from the ground up matters — the people already using shadow AI are telling you exactly where your sanctioned tools fall short.

2. Educate & Empower: Anonymisation. Albeit a temporary fix, we must teach staff how to use public tools safely by anonymising prompts. And while this is a crucial short-term measure, we must be realistic that human nature and convenience mean this won’t always happen. This makes the next steps even more urgent.

3. Provide & Promote: Your Internal AI Must Win on Merit. If you don’t want your teams using public AI, you must provide an internal, secure alternative that is just as good, if not better. 

  • The Trust Mandate: Guarantee Internal Privacy. This is about psychological safety, not just data security. An employee is often more worried about their manager seeing a “silly” question than they are about a foreign state seeing their data. You must provide explicit, public, and technically transparent assurances that employee inputs into the approved internal AI are anonymous and not visible to IT, HR, or management. This commitment must be a cornerstone of the internal communications plan, demonstrating that the tool is for empowerment, not surveillance.

  • The Leadership Mandate: Adopt a Product Mindset. As I’ve argued before, AI adoption is a major business transformation programme. Crucially, this requires moving beyond a ‘project mindset’ with a defined end-date, to a ‘product mindset’ focused on continuous, in-life management. This means dedicated resources, post-launch, to perpetually product manage and iterate on the tool, provide ongoing education, champion its use, and govern its evolution. Treating the internal AI as a living product, rather than a completed project, is the only way to ensure it remains the path of least resistance for your employees.

4. Prepare for Impact: Update Your Crisis Plan. Accepting that leaks are inevitable means preparing your response. This isn’t about creating a new AI-specific crisis plan; it’s about ensuring your existing corporate incident management and communications strategy is updated to include AI-related scenarios. Your Comms and Legal teams need to be ready for the unique challenges a data leak via a public AI can present.

The era of perfect control is over. The most effective leaders will be those who accept the new reality of porous data boundaries and shift their strategy from building walls to building skills, better internal tools, and robust crisis response plans.

How is your organisation preparing for the inevitable arrival of ‘Bring Your Own AI’?