AI in business: vendor lock-in and risks to avoid
November 7, 2026
Once again, the biggest problem is us, the consultants. Companies adopting AI face real risks: costs that can change, terms the vendor can rewrite, and knowledge they own on paper but no longer control. A good consultant should plan for those risks from the first day.
I already see this in the ISO 9001 world. Companies rely on professionals who should have a broader view. Those professionals are supposed to guide the company through improvement, anticipate risks and point it toward the right choices. Too often we solve the problem in front of us and ignore what the decision will mean in the long run. With artificial intelligence, we risk making the same mistake.
What does AI really cost, and for how long?
AI is an extraordinary technology, but it is enormously expensive. It needs infrastructure, energy, hardware, maintenance, development and cloud services. Growing demand for computing capacity is also pushing up the price of the resources AI depends on.
Today a subscription of a few hundred dollars a month can cover usage that would cost thousands of dollars at API list prices, and more than that in some heavy use cases. Meanwhile, the major AI developers are making enormous investments. We don’t know whether today’s prices, usage limits and business models will still hold up in five or ten years, or whether they will still be a good deal.
Picture a company that, over the next few years, hands entire processes to AI agents: customer management, administration, design, technical documentation, production and quality control. Then one day the vendor makes drastic changes to its prices, contract terms or usage limits. If the company can no longer afford those services but can no longer work without them, the problem is no longer technical. We could end up with companies that miss their payroll costs and even the headaches of managing staff.
Is it safe to start with cloud AI?
This is where a competent consultant should step in. If a company doesn’t have the resources to buy suitable AI infrastructure, starting with commercial services is a perfectly reasonable choice. From the first day, though, the system should be designed so the company can switch vendors. When it makes technical and economic sense, it should also be able to move to AI models running on the company’s own hardware.
That doesn’t mean buying expensive servers right away. It also doesn’t mean local AI is always the cheaper option. It means building an architecture that doesn’t lock the company into a single vendor.
Does owning your data mean you control it?
There is a second problem, and in my view it matters even more than cost: the company’s knowledge.
Technical drawings, projects, procedures, standards, past mistakes, solutions, decisions and production data are often a company’s real assets. They took years, sometimes decades, of work, investment and sacrifice to build. Today we are moving more and more of that knowledge onto cloud platforms and external AI systems.
Contracts can state that the customer keeps ownership of the data, as they should. But legally owning information and keeping real control over it are two different things. You need to know who can access that data, how it is processed and where it is stored. You need to know which parties and systems can use it, and what assurances you have that it won’t be used for purposes other than the ones agreed.
I’m not claiming that every cloud or AI vendor uses customer data to train its models or build competing products. Contracts, controls, security systems and specific safeguards do exist. Still, no company should simply trust that those protections will always be enough. The risk goes beyond someone stealing a design. Someone could take the value of the knowledge that made the design possible. A technical drawing can hold years of experience. A production process can contain hundreds of problems already solved, and a procedure can be what sets a company apart from every competitor.
Now picture the knowledge of thousands of companies concentrated in the infrastructure of a few tech multinationals, with vast financial resources, computing power and analytical tools. Nobody has to claim that this knowledge will be misused. It is enough to ask what would happen if it were, and who would be strong enough to fight back.
A recent case: what happens when the vendor changes the rules?
I won’t bore you with my own examples or my own solutions. I’d rather use a recent public case involving a platform that many content creators and agencies use.
On July 23, 2026, Higgsfield, a platform for creating AI video and images, published new terms of use. Users kept ownership of their content, but the platform reserved a very broad license over it: perpetual, irrevocable, transferable and sublicensable. The license also covered generated content and included promotional use and training of its models. According to a line-by-line comparison of the archived versions, the perpetual license had been there since the August 2025 version. The new version expanded it, and that was when many users finally noticed.
Within a few days, the internet pushed back. On July 25 the platform announced a revision, and on July 26 it published a new text, effective August 27. The license is no longer perpetual or irrevocable and, with some exceptions, ends when the user deletes the content or the account. Promotional use now applies only to content made public or authorized for that use. According to the same comparison, though, the sentence allowing content to be used to train the models stayed exactly the same. The commitment to give at least 15 days’ notice before material changes became “reasonable notice.”
This is exactly the gap between owning content and controlling it. The case shows something else, too. Creators and agencies had alternatives and a shared voice, and they got part of the contract changed in a few days. A manufacturer would find that much harder. If twenty years of drawings sit in a single CAD program’s file format, switching vendors costs so much that accepting the new terms is often the only option.
Technical software has already been down this road. Autodesk has not sold new perpetual licenses since 2016 and stopped updating them in 2021; PTC, the maker of Creo, did the same starting in 2018. Anyone who had a license can still use it, but it stays frozen at that version: to get the new versions, which you really need after three or four years, the only way is a subscription. As a consultant I have worked inside many companies, and they all told me the same thing: if I could, I would change software, but it is too complicated. When the rules are written by whoever holds your data, your freedom to say no depends on how much it costs you to leave.
What past cases should companies keep in mind?
Some cases from recent years show that concerns about how information gets used have a basis.
- In 2020 the European Commission charged Amazon with using non-public business data from sellers on its marketplace to favor its own competing retail business. In 2022 Amazon made binding commitments, and the proceedings closed without a final finding of infringement.
- Anthropic reached a 1.5 billion dollar settlement with authors over books downloaded from pirate libraries, about 482,000 works. A judge had found that training on lawfully purchased books was legitimate. The case was about the pirated copies.
- OpenAI faces numerous copyright lawsuits, including one from The New York Times, over the use of copyrighted content to train its models.
- In July 2026 Apple sued OpenAI, accusing it of taking trade secrets through former employees. OpenAI denies the allegations, and no court has ruled yet.
These cases differ, with different responsibilities and different legal situations. They don’t prove that AI vendors take their customers’ industrial designs. They do show that conflicts of interest, intellectual property disputes and misuse of information are real risks.
Who pays when something goes wrong?
A large multinational may have the resources to handle lawsuits, damages and even very large fines. A small or midsize company that loses its competitive edge may have neither the time nor the money to defend itself. Compensation, possibly years later, could arrive after the company has already lost customers, market share and the ability to stay in business. By then the damage could be permanent.
That’s why I don’t think the answer is to give up on AI or to demonize the cloud. The answer is to design systems where the company keeps as much control as possible over its information. That means keeping strategic knowledge and documentation on the company’s own infrastructure, with independent copies and strictly controlled access. It also means sending data to outside vendors only when necessary, using architectures that let you replace services and AI models without rebuilding everything from scratch. Where possible and cost-effective, the most confidential information should be processed locally. The cloud should be a tool that serves the company, and the company’s survival should never depend on it.
What should a consultant do?
Back to where I started: the responsibility of consultants. A consultant shouldn’t just propose the solution that is easiest to install, cheapest today or most convenient to resell. A consultant should weigh the benefits and also the economic, technological, contractual and strategic risks. The questions to ask are concrete. What happens when the vendor raises prices? When a service is discontinued? When the company has to move to another technology, or when confidential information is compromised?
In ISO 9001 we talk constantly about risk-based thinking: identify risks, assess them and plan actions before they turn into real problems. There’s no reason to forget that principle just as we bring in a technology that will become central to how our companies run.
Five questions to ask before handing a process to an AI service
- Where does our data go, who can see it, and can it be used to train the vendor’s models? Look for the answer in the contract and the terms of use.
- If we want to leave tomorrow, can we take everything with us in a format other tools can read: documents, procedures, configurations, history?
- Can the system run on a different model or with a different vendor without being rebuilt from scratch?
- What happens if prices, usage limits or terms change? How much notice will we get, and what can we do in the meantime?
- Which information is so confidential that it must stay under our control, on premises or in independent copies, and who decides what can leave the company?
Why AI needs an industrial strategy
Artificial intelligence is an extraordinary opportunity. It can raise productivity, cut waste and give companies skills and tools that were unthinkable a few years ago. But it has to be adopted with an industrial strategy as well as a technology plan, and an industrial strategy considers the worst cases too. A responsible business owner doesn’t have to predict the future. The job is to build a company that can face it, without waiting for a risk to materialize before starting to manage it.
Using artificial intelligence shouldn’t mean becoming dependent on someone else’s artificial intelligence. Technology can be bought, rented or replaced. Knowledge a company has built over decades may be impossible to recover once it is lost.
Giovanni Commodaro, MechWorks Design Engineering (MW-DE). I design AI systems for businesses that run in the cloud, on premises or as a hybrid.