The free AI tier is quietly turning into a trial with a closing date. A 15-minute audit that tells you which of yours are load-bearing before the window shuts. 

What's the difference, and why does it matter now?

A free tier is a permanent floor: limited, but yours indefinitely. A free trial is a window that closes on a date somebody else chose. Through 2025 most AI vendors sold the first and behaved like it. Through 2026 a lot of them have quietly been running the second.

The clearest recent example is dated and checkable. Meta's Business Agent free window opened on 1 July 2026 and closed on 1 August 2026 — thirty-one days — after which AI-handled conversations bill at $2 per million tokens. A second change is already scheduled: from 1 October 2026, Meta resumes charging for service messages, which have been free since November 2024. That's a two-year "free" that turned out to have an end date.

It isn't only Meta. OpenAI moved workspace-agent and spreadsheet tasks onto token-based credits in early July 2026, and Microsoft raised 365 prices on the 1st of the same month. One quarter, several pricing changes, all in the same direction.

Nothing here is scandalous — introductory pricing has always worked this way. What's new is how much unpaid infrastructure small businesses have quietly built on top of free tiers, on the assumption that free was a floor rather than a promotion.

Which of your free tools are load-bearing?

That's the actual question, and it's not the same as "which do I use most". A tool is load-bearing if something a customer sees, or something you'd have to do by hand tonight, stops working when it stops being free.

Run through it honestly:

• The free automation that files your enquiries — if it stopped on Monday, who notices?
• The free tier answering out-of-hours messages — is that customer-facing?
• The free transcription you use for every client call — how many hours is that, really?
• The free scheduler holding your posting calendar — what's the manual fallback?

Most people find one or two genuinely load-bearing free tools and a long tail of things they'd never miss. Both findings are useful. The tail is cancellable clutter; the load-bearing ones are unbudgeted dependencies, and an unbudgeted dependency is just a surprise with a delay on it.

There's a particular trap in the free tools that came bundled with something you already pay for, because they don't feel free — they feel included. Included is a pricing decision the vendor can revisit at renewal, and typically does, with the increase presented as a plan change rather than a new charge. Those belong on the list too, even though nothing about them looks like a free tier.

How do I run the 15-minute audit?

Four columns, one pass, no spreadsheet required.

1. List every AI tool you use for free. Include the ones bundled inside things you already pay for — those are the easiest to forget and the most likely to be re-priced, because the vendor knows you're already a customer.

2. Mark each load-bearing or not. Using the test above: customer-facing, or work you'd have to do by hand.

3. Find the terms. Specifically: is there a stated end date, a usage limit you're near, or language like "during the preview period" / "introductory"? That phrasing is a schedule, not a mood.

4. Write the fallback. One line per load-bearing tool: what you'd do on the Monday it starts charging. "Pay it" is a perfectly good fallback — as long as you know roughly what "it" is, which means finding the paid price now rather than at the moment of interruption.

That's the whole audit. It's deliberately the same shape as the mid-year stack review, pointed at the tools that never appear on your card statement — which is precisely why they escape every review you already run.

What do I do about the ones that fail?

Three responses, in ascending order of effort.

Budget it. If a free tool is load-bearing and worth paying for, find the paid price and add it to your expected costs now. Then set the ceiling before you need it, not after the first metered invoice.

Reduce the dependency. Keep using it, but make sure the work isn't trapped inside it. Can you export? Is there a second tool that does the job adequately? A dependency you could leave in an afternoon is not really a dependency.

Retire it. If it's not load-bearing and you're only using it because it's free, that's not a saving — it's a small ongoing tax on your attention, and it will eventually ask for money at an inconvenient moment.

Do this on a quiet Saturday rather than on the morning something stops working. Fifteen minutes now buys you the right to be unsurprised in October, and being unsurprised is most of what good cost control actually is. Anything that survives this audit and now bills per job has earned a one-page contract rather than a renewal by default.

Not financial advice — vendor terms, free-tier limits and prices change frequently; verify current terms with each provider before relying on them.

For the vetted list of what we actually pay for and what we don't — free, 5-min read: 10 AI Tools That Save 10 Hours a Week.

Related Decodes

• AI Just Became a Utility Bill — the week's hub: the shift from flat seats to per-use meters.
• The £0 AI Stack — the four roles you can still fill free, and the real price tags to check.
• Your Chatbot Now Has a Meter — what the first closed window actually cost.

AI the Brief (@aithebrief) — AI advice you can act on and trust, for time-starved founders.

TL;DR: Free AI tiers are increasingly free trials with a closing date — Meta's Business Agent free window ran just 31 days from 1 July to 1 August 2026 before token billing at $2 per million began, and service messages free since November 2024 become chargeable again on 1 October 2026. Run a 15-minute audit: list every free AI tool, mark which are load-bearing (customer-facing, or work you'd otherwise do by hand), check the terms for end dates and preview language, and write a one-line fallback for each. Then budget it, reduce the dependency, or retire it.