Sovereign AI, or a rented assistant?

As more of the working day moves into an AI assistant built into the firm’s office software, control over that work moves with it. The vendor decides the price, which features are available, when the service is available, and where the AI reads the firm’s files. This article describes that shift and explains the alternative: private AI that the firm runs on its own premises for confidential client work. It is not a prediction that cloud services will fail, and it is not legal advice. If you would prefer the short version, read Why private AI?.

1What Sovereign AI means here

When IntelekIQ uses the term Sovereign AI, we mean one simple thing: the firm owns and operates the AI assistant that reads its confidential client work, instead of renting that capability as part of a cloud software subscription.

It helps to say what the term does not mean, because the same words are used elsewhere. It is not a government programme for national AI. It is not the “sovereign cloud” offering sold by Microsoft or Google; that is still their service, running on their systems, under their contract terms. And it is not a new product name. Secure Law and Secure Accounting are private AI products. Sovereign AI is the reason a firm would choose one of them rather than stopping at the assistant that came with its office software.

The word “AI” is currently being used to answer two quite different questions, and it is worth separating them.

1. Who sold us the licence? Copilot, Gemini and similar assistants answer this question well. They are built into Word, Outlook and Google Docs, they are easy for a firm to approve, and they are often described as included in a package the firm already pays for.
2. Who operates the system at the moment the AI reads a client file? This is a separate purchasing decision. A vendor can encrypt the firm’s data while it is stored, keep each customer’s data separate from every other customer’s, and promise not to use the firm’s prompts to train its models. All of those things can be true, and the client file is still being processed on the vendor’s computers, outside the firm.

2Whoever runs the assistant controls it

None of this means the firm should stop using cloud software. Storing email and documents in Microsoft 365 or Google Workspace is a decision about where files are kept, and it is a normal business decision that most firms have already made. The new question is about the AI assistant that now sits inside those same applications.

Once drafting, summarising and searching are done by an assistant built into the tools people use every day, the firm no longer controls how that assistant works. It rents the capability. The vendor can change which AI model is used, raise the price, limit how much staff can use it, or change the terms of service. Staff will still click the same button and may not notice that anything has changed. To be clear, this is not a claim that cloud services are insecure. It is an observation about who has the power to make those decisions, and that is the vendor, not the firm.

For a small or mid-size firm this matters more, not less. A large organisation usually has a specialist team that can renegotiate a licence, work out what a new usage limit means for the business, or rebuild a process when a vendor withdraws a model. A smaller firm rarely has that. What it does have is a set of clients who engaged the firm, not the software vendor, and who expect the firm to be able to account for what happens to their information.

The question we find most useful is the same one we ask about any processing of confidential work:

At the moment the assistant reads this file, where is that data, and who operates the system that could be legally compelled to hand it over?

Saying “the assistant is included in our package” answers a purchasing question. It does not answer the question above. For a fuller explanation of the difference between a locked-down desktop, an enterprise Copilot subscription, and processing that happens inside the firm, see Sealed Desktop, or Sealed Inference?.

3How a bundled assistant becomes a dependency

The response we hear most often is calm and reasonable: we already use Copilot, or Gemini, or the AI feature built into our practice management system. Those tools are genuinely useful for general drafting, email and other work that does not involve client information. They are a sensible way to run an office.

They are also how a working habit forms. An assistant bundled into an existing subscription feels free. After a few months, moving away from it no longer means giving up “another chatbot”; it means changing how the working day is organised. Staff naturally put work into whatever tool is already open in front of them. If the only assistant the firm has approved is the one inside Word, then confidential work will either be done without AI, or it will be put into that assistant because it is convenient and already paid for.

Banning public AI without offering an alternative is not a policy that staff can realistically follow. Sovereign AI exists so that there is a firm-approved way to use AI on the work that genuinely needs it: this client’s file, this tax return, this matter, this engagement. The firm keeps Copilot or Gemini for non-confidential work, and puts the confidential work through a system the firm itself operates.

4Comparing the costs fairly

A private AI system installed at the firm costs more at the outset than a monthly add-on to an existing subscription. That is true and should be stated plainly. The conclusion that usually follows, that renting must therefore be cheaper, is the wrong comparison if the thing being rented is the firm’s confidential workflow.

The more useful question is not whether one option is cheaper in the first month. It is this: what is it worth to the firm to keep the processing of confidential work, and the working habits its staff have built around AI, under its own control if rented access later becomes more expensive, slower or more restricted?

Cloud AI is easy to adopt while access is cheap and plentiful. Costs, access and confidence can all tighten. The AI industry is investing at a scale that makes cheap, unlimited access a risky thing to treat as permanent. If the easiest commercial path for a vendor is to raise prices once customers find it hard to switch, the firm that only rents will be the one paying. That is a precaution against a possible future, not a forecast that any particular vendor will fail.

A private system is not free to run either. It uses electricity, needs maintenance, and needs its AI models updated from time to time. What the firm is buying is a capability it operates itself, not a promise that the invoices stop.

5Are models that run inside the firm good enough?

The remaining objection concerns capability: a model that runs on the firm’s own hardware cannot match the largest models offered by the big AI companies, so the firm should either wait or rent the larger model. For some tasks that is a fair point. It is not an accurate description of most of the work a small or mid-size professional firm needs AI to do.

The most valuable work is research, document review, and first drafts that cite their sources, all carried out over the firm’s own files. That requires two things: a system that finds the relevant material in the firm’s files, and a capable model to work with it, with a person still responsible for anything that leaves the firm. It does not require the largest model in the world. Models that run on the firm’s own premises are already good enough for that job, and they continue to improve. The firm can replace the model with a better one without changing anything else about how the system works or how staff use it. With a subscription, the firm gets whichever model the vendor chooses to provide, at whatever price the vendor sets.

Working from the firm’s own material does not remove errors altogether. It does make the output checkable: the assistant starts from material the firm has authorised, shows which sources it used, and a lawyer or accountant reviews the result before it becomes advice or is sent to a client. For professional work, that ability to check matters more than which model scores highest in a general test.

When a larger external model is genuinely useful, there is still a way to use it. In IntelekIQ’s products that route is called Frontier Mode. Before anything leaves the firm, names and other identifying details are replaced with placeholders. A person reviews and approves exactly what will be sent. After the external model has responded, the original details are restored inside the firm. Frontier Mode is optional and is not the default. Sovereign AI does not mean the firm never uses an external model. It means the default route for confidential work is one the firm operates.

6Keep the office software. Own the route for confidential work.

Sovereign AI does not mean replacing the firm’s existing IT. Email, document storage and general-purpose AI stay where they are. Private AI is for the confidential part of the work: typically one or two workflows, a named group of people who review the output, and a controlled environment. That is a much smaller decision than replacing the firm’s systems, which is what many owners assume is being proposed when they first hear an argument that sounds like criticism of Copilot.

Clients look to the firm, not to the firm’s software vendor. Being able to tell a client that their confidential information is processed under the firm’s own control is a commercial advantage, not a technical nicety. For a small practice, a single anxious client can represent a significant share of its income.

For how private AI fits alongside practice management software, research databases and public AI tools, see Control in the office. For the difference between storing files with a cloud provider and sending them to an AI model, see Secure Cloud, or Sealed? and The Assistant You Already Trust.