Protecting the Ingredients Is Not Enough

For years, companies have focused on protecting data: customer information, source code, contracts, pricing, product plans, and financial information. That still matters.

But competitive advantage often comes from something harder to see.

It comes from what your best people know how to do with that information.

Two companies can have the same data and still make very different decisions, because one knows which signals matter, which assumptions to challenge, and what question comes next.

That is the difference between ingredients and a recipe.

You can lock the ingredients in a vault and still lose your advantage if someone else learns exactly how your best chef prepares the dish.

AI creates that risk in a new way.

Imagine an experienced employee working through a difficult problem with an AI assistant. The first answer is incomplete, so they correct it, adding context, changing an assumption, and explaining why one option will not work.

Eventually, the answer improves.

But during that process, the employee may have revealed something more valuable than the original data: how they solve the problem.

AI Sherlocking

This is where the idea of AI Sherlocking becomes relevant.

The term Sherlocking originally described what happens when a large technology platform absorbs functionality created by another company. In the AI era, the concern is less visible. Instead of copying a feature, the broader AI ecosystem can potentially learn from the problem solving patterns, corrections, feedback, and interactions generated by users.

This is not hypothetical. In February 2026, Anthropic disclosed that three AI labs in China ran more than 16 million interactions with its Claude models through about 24,000 fraudulent accounts, training their own systems on Claude’s reasoning.

The fraudulent campaign was an extreme case, but it exposes a broader issue: AI systems become more capable as they learn from interactions, feedback, outputs, and repeated patterns of use.

No database needs to be stolen. No employee needs to intentionally share intellectual property. A company can have strong controls around uploads and still miss this risk entirely, because it does not look like a breach.

But protecting the data does not automatically protect the reasoning applied to it.

Could Smaller Models Help?

The answer is not to stop using AI.

It is to stop sending every problem to the same model.

Large Language Models are the Swiss army knife of AI. They handle almost any task passably well, from drafting an email to debugging code to answering a legal question. That range is why so many companies default to one.

Small Language Models are the scalpel. Built for a narrower set of tasks, they are less flexible, but far more precise where it counts.

SLMs can be optimized for specific domains and tasks, and typically need far less computing power than a large general purpose model. Bayer built one to answer farmer questions about crop protection, and it tested 40% more accurate than a large model on the same task.

That points to a practical way to divide the work: send broad, low stakes questions to an LLM, and keep your most sensitive functions on an SLM inside infrastructure you control, where the reasoning never leaves the building.

That does not make SLMs automatically secure. The architecture, hosting, training data, and governance still matter. For companies serious about protecting both data and logic, an SLM running inside infrastructure they control can become one important part of the answer.

Keep more of the recipe inside the kitchen.

Companies have spent years learning how to protect data.

AI introduces another asset worth protecting: the judgment, processes, and reasoning that turn information into better decisions.

Protect the ingredients. But do not forget the recipe.