Episode 13: The Shadow AI Problem

If you run a company and believe AI isn't being used inside it, Holly Joint and Ewan MacLeod have unwelcome news: it almost certainly is, just not in ways you can see. This episode turns from their personal experiments toward the harder question of what AI is actually doing inside organisations, and why the gap between what employees do and what leaders know has become a serious problem.

The conversation opens on shadow AI, the unsanctioned tools staff reach for when official channels fall short. Ewan shares a striking figure from a 500-person consumer goods firm: in 90 days, office staff uploaded around 50GB of company data into outside AI systems and pulled 30GB back out, prompting a swift clampdown. More memorable still is a room of Gulf compliance professionals who almost unanimously denied using AI at work, then described, in unison, an elaborate workaround involving photographing confidential documents on their phones to get around their own security. The honesty problem, the hosts argue, is compounded by a deeper shift: when everyone has instant access to knowledge, knowledge itself stops being a reliable currency.

From there the discussion widens to value. Most organisations, Holly and Ewan agree, are still stuck at the task level, summarising meetings, drafting policies, checking documents, useful but hardly transformative. A few have gone further. They describe a financial institution that restructured its contact centre with conversational AI and released hundreds of staff, the kind of change rarely discussed publicly for fear of how it looks. Holly pushes back usefully here, noting that some firms slap an "AI" label on ordinary downsizing to flatter their share price, and that chatbots alone aren't the systemic, end-to-end transformation people imagine.

The episode's real argument is about mindset. There are companies chasing AI to shrink, and companies using it to grow. The first squeeze out cost efficiencies and improve profit; the second keep their people, make them dramatically more productive, and rethink their business model entirely. Holly's conviction is that the shrinkers are leaving the larger prize untouched, because they never look at their top line or the disruption they could cause.

What does starting well look like? Ewan defends "test and learn," while admitting it can sound like a cop-out in a boardroom impatient for answers. Holly's refinement is to go narrow rather than wide: hand one capable tool to a legal, procurement or engineering team and see how much better, not smaller, they become. The pair close on what excites them most, the moment a custom OKR system or a replacement for expensive enterprise software appears on screen in an afternoon, and a room of leaders realises the constraints they have always worked within may no longer apply.

Key Topics

• Shadow AI and the data-leakage risk hiding inside most companies
• Why employees underreport how, and how much, they use AI at work
• Knowledge as a devaluing currency when AI is universally accessible
• The gap between task-level use and genuine systemic transformation
• Cutting costs versus growing capability: two opposing AI strategies
• AI as cover for ordinary layoffs, and the share-price games involved
• "Test and learn" versus targeting one narrow, high-value use case
• Building custom software in hours and what it means for SaaS vendors

Links & References

• Microsoft Copilot — https://www.microsoft.com/microsoft-copilot
• Anthropic (Claude, Claude Code) — https://www.anthropic.com
• ChatGPT (OpenAI) — https://openai.com