What it costs, honestly

What you'll be able to do afterwards: Estimate the monthly cost of your own AI use, including the costs nobody puts on the pricing page.

AI is unusual in that its headline price is both true and misleading. The subscription is real and easy to budget. The cost that decides whether it was worth it sits somewhere else.

The two pricing models

Subscription. A flat monthly fee per person, usually with usage limits. Predictable, easy to approve, and how most people start. The trap is seats: one person's plan becomes eleven people's plans because it was easy to add and nobody removed anyone.

Usage-based. You pay for what you process, measured in tokens — roughly, pieces of words. Cheap per item, unbounded in total. This is how anything automated is billed, and it is where costs surprise people, because a workflow running quietly every night generates a bill without anyone approving it.

A practical rule: use subscriptions for people, and usage-based pricing for programs. Then keep an eye on the second one, because nobody is watching it by default.

A realistic monthly bill

Ignore list prices, which change. Think in bands for a small team of five:

A few people using assistants daily: the cost of one modest software subscription per person. This is the easy, low-risk end.
Assistants plus one or two production tools for images or video: noticeably more, still a line item you would recognise on sight.
Automated workflows processing volume: the least predictable category. Could be small, could exceed everything else combined. This is the one that needs a ceiling set before you start, not after.

Whatever you plan, set a hard monthly ceiling on the automated category and treat it like a budget rather than a forecast.

The three costs that are not on the pricing page

Your review time. This is the biggest and the most consistently ignored. A draft that takes ten seconds to produce still takes you three minutes to read, correct and approve. If that review is not accounted for, the saving is imaginary.

Do the arithmetic honestly once. If a task took thirty minutes and now takes twelve minutes of reviewing plus two minutes of prompting, you saved sixteen minutes — not thirty. Still good. Still worth doing. But it is the honest number, and it decides which tasks are worth automating.

The corrections you will not notice. A subtle error in a routine document is worse than an obvious one, because it gets published. Build in one check that is cheap and always applies. For anything with a number in it, that check is: where did this number come from?

The subscription you forgot. Four tools that each felt essential in February and one is used in June. Review the list quarterly. This is the single most common source of waste in small teams, and it has nothing to do with AI — it is just a subscription list nobody prunes.

When it is genuinely cheap

Three conditions make AI use clearly worth the money:

The task is repetitive. You do it weekly or more. One-off tasks rarely repay the setup.
The output is verifiable. You can tell at a glance whether it is right — a table that must add up, a format that must match.
A mistake is cheap. An internal draft costs nothing to be wrong. A published quote does.

When all three hold, the arithmetic is easy and the saving is real. When any of them fails, be sceptical.

Where it is expensive and people pretend otherwise

Tasks needing deep context. You spend longer supplying the background than doing the work.
Anything with legal, medical or financial consequence. You will review it line by line anyway, and correctly so.
High-volume output nobody reads. Producing more documents is not progress. If the volume was never the bottleneck, automating it creates a queue.

Do this now

Take your last month's tool subscriptions and write the total down. Then estimate — honestly — how many hours you got back. Divide. If the number of hours is comfortable, you are doing fine. If you are not sure, that uncertainty is the finding: you do not yet know what this is saving you, which is the state most teams are in.

Next: lesson 5, the failure mode you must understand before trusting any of it.