The real cost of AI in software — and why your tools should be honest about it

Every app has an “AI” badge now. Almost none of them tell you what that AI costs to run, which matters, because someone pays for it — and if it's not priced honestly, eventually it's you, by surprise.

The two costs that actually move

Most software is cheap to run per customer — it's database rows on servers the company already pays for. AI breaks that, because two things cost real money every time they're used:

• Tokens — every AI message in and out is metered. Cheap per message, but it adds up with heavy use.
• Voice and SMS — phone minutes, text segments, transcription. This is the expensive one: an AI that answers calls can cost real dollars per call.

Everything else — scheduling, invoicing, the client portal — costs the vendor almost nothing no matter how hard you use it. The AI is the part with a meter on it.

Why honest tools cap and meter, and shaky ones don't

If a tool bundles unlimited AI into a flat price, one of two things is true: the AI is barely capable, or the price will rise once enough people actually use it. The honest model is the boring one — a clear allowance, then metered usage past it, the same way your card processor charges a fee per transaction. No surprise, no “unlimited” that quietly isn't.

What we do about it

We split the cheap stuff from the metered stuff on purpose. The core tools are flat-rate because they cost us almost nothing. The AI agents and the phone line are add-ons with a setup, an included allowance, then usage past it — so the people who use them heavily pay for what they use, and the people who don't aren't subsidising them. It's less exciting than “unlimited AI!” It's also true, which is the point.