Most posts about AI tool costs list subscriptions and add them up. I don't find that number very useful.

Across the work I direct, at the agency where I lead SEO and on my own sites, the metered API bills are the smallest line. The expensive parts are somewhere else: in review, in output nobody checked, and in wrong numbers that reach a client.

The cheap part: compute

A few real numbers from my own work:

At a few dollars a run, the open question is whether you can trust the output.

Expensive part 1: review time

An illustrative example: say an agent drafts a 2,000-word article in a few minutes. If a person then needs an hour to verify every claim in it, the article costs an hour, not a few minutes. Most cost math counts the minutes.

The fix is to make review cheaper and harder to fake. In the content engine I designed at work, every factual claim is traced to a source in a claim ledger before a human sees the draft. On one measured batch, first-pass accuracy was 97.78% (14 errors in 633 claims), which I report as an upper bound. Review packets carry planted defects, so a reviewer who skims gets caught.

That machinery takes time to build, but it's still far cheaper than reviewing blind.

Expensive part 2: output nobody checked

One of the clearest examples I've dealt with wasn't even an AI tool. It was a $1,900/month link subscription that shipped without review. I cut it.

Same lesson with a link vendor: before approving a reorder, I audited the previous campaign and found 20% link retention. The reorder stopped.

The real price of any tool or service includes everything that goes out because nobody looked, and automation makes that number grow faster.

Expensive part 3: a wrong number that reaches a client

One month, a burst of bot impressions would have made an RV brand's +8.4% impressions look like −15.1% in a client deck. That's a 24-point error, pointing the wrong way. A check I built caught it before the deck shipped.

That kind of mistake never shows up on an invoice; it shows up as a client who stops believing your reports. An AI workflow that speeds up reporting without adding checks stays cheap only until that happens.

Build vs. buy changed

Agents also change which subscriptions you need at all. I built Gmail Mission Control, an in-house Gmail tool that replaced the agency's shared-inbox subscription, from idea to v0.8 in six days.

The caveat is that a tool you build is a tool you maintain. Building is cheaper than it used to be, but owning still costs something.

Comparison: hiring someone

The usual framing compares a tool bill to a salary. I don't think that comparison works. Tools don't replace a person's judgment; they change what that person spends time on, with less assembling and more checking and deciding.

The better question for any new tool: what will it cost to verify its output, and who is going to do that?

The dependency question

The bigger risk than the bill is dependence. If one provider changes pricing or quality, a single-vendor workflow breaks.

Two hedges I actually use:

I'm not planning to switch. I just want to know I could.

Summary

AI compute is cheap and getting cheaper, so it's rarely the big line. The money goes into the time to verify output, the cost of output nobody verified, and the credibility you lose when a wrong number reaches a client. Budget for those first and the subscription line takes care of itself.