What You Actually Own When You Pay for AI

Somebody who bought a film on their PlayStation a few years ago went looking for it recently and found it gone. So did a lot of other people, because Sony's licensing deal with the studio lapsed and more than 500 titles came out of customer libraries at once. These were purchases, not rentals, and there were no refunds.

The same month, Microsoft retired five features from Copilot. Group chats, AI podcasts, Copilot Labs, Deep Research and the animated assistant character all went on August 18. Microsoft's support page is careful about the wording: group chat threads, messages and images "will not carry forward after your account updates," and anything not saved beforehand "may no longer be available afterward." OpenAI had already closed the Sora app back in April, with its API due to shut on September 24 and users told to download their videos before the deadline.

Three companies, three unrelated reasons, and one thing in common. Something people had made or paid for stopped being available to them, and in each case the only warning was a page on a support site.

Nobody here did anything wrong, which is the point

Licensing agreements expire. Products that almost nobody uses get retired so the teams behind them can work on something else. These are normal decisions, made for normal reasons, and it would be strange to expect otherwise.

What has changed is what sits inside the thing being switched off. An AI product disappearing used to mean losing a novelty you had played with for a fortnight. Now there is a decent chance your team has been building on it. The prompts that took a month of fiddling before they produced anything useful. A weekly process that only works because of one particular feature. Research, drafts and summaries that were written in a tool nobody thinks of as a place that needs backing up, which is why none of it has a copy anywhere else. When the feature goes, there is nothing to restore from, because there was never anything to restore.

Where Copilot actually earns its money

We are not about to tell anyone to cancel Copilot over this, and we would say the same to a client who rang us today.

If your business already runs on Microsoft 365, Copilot is worth putting in front of a few people. It is unusually good at going back through a year or two of your own email and working out where a long-running conversation has actually got to. Most projects at most small businesses are run through email rather than through any system designed for the job, and the result is that nobody is quite sure who owes whom an answer. Copilot will tell you, on both sides, and that is a genuinely useful thing to hand somebody on a Monday morning.

Across our own clients we see fewer businesses paying for Copilot than for ChatGPT or Claude, spread across a mix of personal accounts and the team accounts that have only recently started to appear. That is what our own book looks like rather than a claim about small business in general.

The two things people mean by "build your own"

When somebody says they would rather build their own, they usually mean one of two quite different things, and only one of them is realistic.

The first is owning the hardware outright, running a model in your building so that nothing leaves the premises. For a top tier open model this is out of reach for a small business at the moment, and the obstacle is not the price. Today's large open models want several hundred gigabytes of memory to run properly, and the machines capable of holding that much were pulled from sale during this year's memory shortage. What decides which models you can run in 2026 is memory capacity rather than budget. Smaller models are a different matter, and there are good reasons to run those locally. We do it ourselves for image generation. But anyone telling a twenty person company they can host a frontier model in the back office is selling something.

The second meaning is the one that works, and it is considerably less exotic than it sounds. Your information stays in a database you control. The sequence of steps, the checks, the small repetitive loops that encode how your business actually operates, all of that is written by you and belongs to you. The model is asked for reasoning through an API, and reasoning is the only thing it supplies. Arranged that way, a vendor retiring a feature costs you nothing, because the feature was yours to begin with. A vendor putting prices up becomes a comparison to make rather than a crisis to manage. And the expensive part, which is the months of work encoding how the business really runs, stays where you can get at it.

Why this is getting easier rather than harder

The gap between open models and the commercial frontier has narrowed considerably. NIST's AI standards centre looked at one of them over the summer and concluded it was probably the most capable open weight model available when it launched, with overall capability in the region of a leading commercial model from late last year. Assessors in the UK put open weights somewhere between four and seven months behind the frontier.

None of which is a suggestion that you go and run one. That same assessment found the safeguards on it weaker than on the American commercial equivalents, and recommending it would be a peculiar thing for a security company to do.

The reason it matters is pricing. When something capable is free to download and only a few months behind, no vendor has permanent pricing power, and the rent you are paying this year is not necessarily the rent you will pay next year.

Two things worth doing this week

The first takes about five minutes. Open whichever AI tools your business pays for and look for anything that exists only in there, then copy it out somewhere ordinary. With Copilot the update is arriving account by account and Microsoft has not published a schedule, so there is a reasonable chance yours has not changed yet.

The second takes a conversation. Find out what your team is already paying for, including whatever has quietly gone onto somebody's personal card, and move that work onto a proper team account. Anything that matters should not be sitting in a private login that walks out of the door when its owner does. We have been helping clients through that move, mostly onto Claude and then OpenAI.

The useful question was never which AI to buy. It is what you would still be holding if you stopped paying for it.

If you would like a hand working out what your team is on, and what would survive the bill stopping, book a free 30-minute strategy session.