The tool trapped on one laptop

A specialist firm had a secret weapon, and it lived on one laptop.

One person had built an AI agent that did genuinely valuable work, the kind of analysis clients paid the firm real money for. It ran on a pile of scattered prompts, a consumer subscription, and that one person’s memory of how the pieces fit. It had never been written down. It had never been tested. If that laptop died, or that person left, or the subscription plan changed its limits on a Tuesday, the secret weapon was gone.

Sound familiar? It should. 49% of employees say they use AI tools their employer never sanctioned (BlackFog, 2026), and 47% of the people using generative AI at work run it through personal accounts, outside anything the company can see or keep (Netskope / Infosecurity Magazine, 2026). Almost every organization now has a tool like this. Clever, quietly load-bearing, and one login away from breaking.

Everyone at the firm could feel the bigger prize. Clients would pay for this directly. But you cannot sell a pile of prompts on somebody’s machine. So the question became the one that separates tinkering from building. What would it take to make this real?

The answer was four moves, and none of them was magic.

• Recover and version every asset. Every prompt, every configuration, every half-remembered trick got pulled out of the laptop and into a versioned system of record. The tool stopped being a memory and became a codebase.
• Build test cases that prove it works. Real inputs, expected outputs, run again after every change. This is the unglamorous step most people skip, and it is the one that turns “it usually works” into “it works.”
• Move it to an API with a web app in front. Now a client could touch it without borrowing anyone’s laptop, and the firm could see usage instead of guessing.
• Make it configurable. Standing up a new client instance went from a rebuild to a few hours of setup. That is the move that turned a tool into a product line.

A fragile personal script became a tested, repeatable product the firm now sells to other clients under its own brand. Same underlying idea. Completely different asset.

The distance between a clever personal tool and a sellable asset is method, not magic.

Here is the lesson executives should take from this, because it is not about this firm. Version it. Test it. Make it repeatable. Those are management ideas, not engineering ones, and leaders already know how to demand them. Most just have not pointed that demand at the AI experiments scattered across their org, which is exactly where the shadow-AI numbers above say the value is hiding.

Somewhere in your company, and maybe on your own laptop, a tool like this already exists. It deserves better than dying with a subscription plan.

Book a discovery call. Bring the fragile thing that already works. We will build it into something you can stand behind.

Sources

• BlackFog, 2026. A Sapio Research survey of 2,000 UK and US workers at organizations with 500+ employees found 49% using AI tools not sanctioned by their employer, a measure of how much real AI work happens outside official systems. https://www.blackfog.com/blackfog-research-shadow-ai-threat-grows/
• Netskope / Infosecurity Magazine, 2026. Netskope’s Cloud and Threat Report found 47% of people using generative AI tools at work do so through personal accounts, outside company oversight and continuity. https://www.infosecurity-magazine.com/news/personal-llm-accounts-drive-shadow/