AI: adopt fast, or first protect what sets you apart?
"Stealing the weights and alpha of my business": what big companies' catch-22 means for small and mid-sized businesses
In July 2026, Palantir CEO Alex Karp told CNBC's Squawk Box what, by his account, heads of large companies say in private about the providers of language models: "These people are stealing the weights and alpha of my business." And further: "Why would they get access to my data if they're gonna build my alpha?"
In finance, alpha is the return above the market average. In a business, it is what sets you apart from competitors: recipes, process settings, costings, prices, customer knowledge, the way you solve problems others cannot. Karp is not a neutral observer, as Palantir sells its own platform as the alternative. But the fear he describes is real, and it affects companies that have never heard of Palantir.
The catch-22
On one side is pressure of a kind technology has not seen before. Exceptional amounts of capital are going into models, data centres and chips, and that capital has to earn a return. It can only do so if AI is used everywhere as fast as possible: in every department, for every task, by every employee. Hence the message boards hear from all sides: whoever does not adopt falls behind.
On the other side are companies that know their value lies in knowledge others do not have. For AI to really help, they have to give it that knowledge: documentation, production data, costings, customer correspondence. The more useful the use, the more knowledge flows out of the business.
The result is familiar: many pilots, few company-wide rollouts, bans on public tools and long legal reviews of contracts. Companies pay for usage that creates no value and fear the usage that would.
What the reality is
What actually happens to data depends on which tool you use and how.
Free and personal tools. With personal accounts, whether your input is used to improve models often depends on a setting the user never looks at. If an employee copies a costing into a personal account, the company has no control over it.
Business plans and API access. The major providers state in their business terms that they do not use customer data to train models. That is a contractual commitment to read, not to assume.
The biggest risk is often people, not models. In 2023 Samsung restricted the use of public AI tools after engineers entered parts of source code into them. The model stole nothing. The data left because there were no rules.
Karp means something subtler. Even if a model does not train on your data, the knowledge of how to use AI in your industry accumulates with the provider, not with you. You build processes on a platform you do not control, pay by consumption and find it hard to switch. "Why wouldn't I control the weights?" is his question. For a large company it is a strategic decision. For a smaller one it is mainly a question of which knowledge may leave the building at all.
What this means for small and mid-sized businesses
According to Eurostat, 20% of EU enterprises with at least ten employees used AI in 2025: 55% of large enterprises, 30% of medium-sized and 17% of small ones. Enterprises that had considered AI but not adopted it most often gave these reasons:
- lack of relevant expertise (71%),
- unclear legal consequences (53%),
- concerns about data protection and privacy (49%).
So the fear of losing data matters, but it is not the biggest barrier. The biggest barrier is that a company does not know where and how AI would pay off.
For small and mid-sized businesses this catch-22 has two sides.
The downside: a smaller company cannot negotiate terms with a global provider. It accepts the standard terms or does not use the tool. It has no legal department to review the terms and no IT department to run its own model.
The upside: while large companies hold back, a smaller one can move faster, provided it knows what may go into a tool and what may not. Most useful applications in a manufacturing business need no alpha at all: quotes, meeting summaries, searching documentation, writing instructions, processing invoices and payments. There is nothing to steal there, and the time saved is immediate.
The paradox lies elsewhere. In smaller companies, alpha is often not written down at all. It sits in the process engineer's head, in the production manager's spreadsheet, in the owner's experience. To use AI, it first has to be written down. That is valuable in itself, because the company becomes less dependent on individuals. At the same time, written knowledge is easier to take away, with AI or without it.
And micro businesses?
Eurostat's survey does not even cover enterprises with fewer than ten employees. In a micro business, the alpha is often the owner: their knowledge, their customers, their methods. The fear of losing intellectual property is therefore personal and understandable.
In practice, though, the risk for a micro business often lies somewhere other than the owner expects. A model will not learn your recipe from one conversation. It is more likely that you breach a commitment to a customer: if you signed a non-disclosure agreement, putting the customer's data into a tool the agreement does not allow is not a technical question but a contractual one. That is why three rules are enough for a micro business: a business account instead of a personal one, training on your data switched off, and never customer data without the customer's consent.
How to adopt AI without giving away your alpha
1. Sort your knowledge into three classes.
| Class | Examples | Rule |
|---|---|---|
| Public | quotes without prices, general instructions, published catalogues | freely in approved tools |
| Internal | minutes, internal procedures, invoices, payments | only in business tools with training switched off |
| Alpha | recipes, process settings, costings, prices, customer data | only in tools under your control, or not at all |
2. Start where there is no alpha. Make the first use cases ones built on public and internal data, where the benefit is quickly measurable. People learn to use the tools before they get near sensitive data.
3. For alpha, choose control. Where the benefit is large and the knowledge sensitive, consider solutions where data stays in your environment: open-weight models on your own infrastructure, or in a private cloud with a clear contract. It is more expensive and slower, so use it only where it pays.
4. Give every use case a business case. Karp's complaint about "tokens that create no value" applies to smaller companies too. Subscriptions for every employee are not a strategy. For each use case, write down how much time or money it brings, what it costs and who is responsible for the result.
5. Write the rules before you need them. One page is enough: which tools are allowed, which data goes into them and which does not, who approves new tools.
Five questions for your business
- Do you know which AI tools your employees use today, and whether through personal or business accounts?
- Have you written down what your alpha is and where it is stored?
- For the tool you use, have you read the terms on using data for training?
- Do your non-disclosure agreements with customers allow their data to be processed with AI?
- For at least one AI use case, can you show how much time or money it brings each month?
If you have no answer to two or more of these, AI is probably already running in your company, but without rules and without measuring the benefit.
Where to start
For one week, collect which tools people use and for what. Then list the ten tasks that take the most time and contain no alpha. On one of them, measure the time before and after. That is your first case to show the board, without a single sensitive piece of data leaving the company.
Sources: Alex Karp’s statements are from his interview on CNBC’s Squawk Box, 1 July 2026. AI usage figures are from Eurostat (Use of artificial intelligence in enterprises), 2025 data for enterprises with at least ten employees.
Case study: Payments automation (PDF) ↗
Inventory management: from master data to excess and obsolete materials
Why the warehouse fills up again after a one-off inventory write-off, and six steps that prevent it: master data, posting movements, cycle counting, BOM standards, S&OP and excess and obsolete stock.
Less inventory, more cash, but not more profit
Reducing inventory releases cash, but EBITDA usually falls in the year of the reduction. Three traps, a case from practice, a board table and six steps to keep the warehouse from filling up again.

