AI pricing confuses people because it arrives in unfamiliar units, and vendors are happy to leave it foggy. Here is the whole cost picture in plain English, followed by the only ROI calculation most businesses need.
The cost layers
Subscriptions. Chat tools like Claude are sold per person per month, comparable to any software seat. For an individual, this layer is trivially cheap against the hours it saves, and it is where everyone should start.
Tokens. The moment you build systems, agents, integrations, automations, you pay for usage, metered in tokens. A token is roughly three quarters of a word, and you pay per million of them, with bigger models costing more per token than smaller ones. The intuition that matters: a single conversation costs cents, so individuals can ignore tokens entirely, but a system processing thousands of documents or calls a month has a real usage bill, and model choice at volume is a genuine financial decision. This is why the models article banged on about routing: running high volume simple tasks on a frontier model is the AI equivalent of doing deliveries in a road train.
Build and integration. Getting AI wired into your actual systems, the practice management software, the phones, the inbox. This ranges from a few hours of setup to serious engineering, and coding harnesses have collapsed it dramatically, but rarely to zero.
The hidden lines. Your time learning and supervising, the change management of getting a team to adopt, and ongoing maintenance, because software that touches other software needs an owner. These dwarf the token bill in most small deployments and almost nobody budgets them. When an AI project disappoints, the missing money was usually spent here without being counted.
The only ROI formula you need
Value per year equals hours saved per occurrence, times occurrences per year, times the loaded hourly cost of whoever was doing it. Compare that to total cost per year, all four layers above. Divide to get payback.
Worked example. A trades business does ten quotes a week, and AI assisted quoting saves forty five minutes each. That is 7.5 hours a week, call it 375 hours a year. At a loaded cost of 85 dollars an hour, the saving is roughly 31,900 dollars a year. If the all in cost of the solution, subscriptions, usage, setup amortised, and a realistic allowance for supervision, comes to 8,000 dollars a year, the return is about four to one and payback is around three months. Those are the shapes of numbers I see in real deployments when the target is chosen well.
Two honesty rules keep the maths trustworthy. First, only count hours that convert into something, more jobs quoted, earlier invoices, owner time redirected to selling, saved hours that evaporate into slack are not savings. Second, run the calculation before you build, as a filter, and again ninety days after, as a scoreboard.
Where the money is actually lost
Not on subscriptions, they are noise. The real losses are building for a low value target, the fix is the bottleneck article, model maximalism at volume, the fix is the models article, and abandonment, paying for things nobody drove, which is the subject of the next article. Notice none of these are technology failures. They are management failures wearing a technology costume, which is good news, because management is fixable.
Cheap to try, compounding when aimed well, expensive only when unowned. That is the honest cost story. Now let us talk about why so many pilots die anyway.