When an AI agent pays off, and when it's wasted budget

19 Jun 2026 · David Sak · 5 min read

Everyone sells AI agents these days, and demand often starts with "our competitors already have one". That's the worst possible reason. An agent is an infrastructure investment, and like any investment, you can calculate up front whether it makes sense. Just answer three questions honestly.

1. Does the task repeat often enough?

An agent pays off on volume. If your team handles twenty identical queries a day, automating eighty percent of them saves sixteen interactions, every day, including peaks and holidays. If it handles five a week, you save a few minutes and the deployment never pays for itself.

A practical threshold from my experience: below roughly 15 repeated interactions a day, don't consider an agent. It's cheaper to simplify the process or write a better FAQ.

2. Do the answers exist somewhere in documentation?

An agent can't invent knowledge the company hasn't written down. It can find it, compose it and back it with a citation. If the answers live only in the heads of your two most senior colleagues, the first step isn't an agent, it's knowledge capture. That, by the way, tends to be the most valuable side effect of the whole project: documentation that would otherwise never exist.

3. Can you bear the cost of an error?

Every system, human or machine, makes mistakes. What matters is how much a mistake costs and whether you can catch it. For an internal helpdesk, a wrong answer is an inconvenience; for legal advice to clients it can be existential. That's why a good deployment is recognized by its escalation protocol: the agent answers only where it has support in the documents, and hands uncertainty to a human. It never guesses.

The calculation, on an example

A company with a four-person support team handles 120 tickets a day, averaging 12 minutes per ticket at CZK 450 per hour of work. Daily cost: 120 × 0.2 h × CZK 450 = CZK 10,800. An agent conservatively handles 70% of first-level tickets, saving roughly CZK 7,500 a day, over CZK 150,000 a month. A deployment in the lower hundreds of thousands is paid back within two months.

The same calculation with twenty tickets a day comes out to savings around CZK 25,000 a month, payback over a year. Still defensible, but worth weighing. And with five tickets a day, I'll talk you out of the agent myself.

Summary

If you answer yes to all three, the agent will very likely pay for itself within half a year. If not, you've just saved yourself hundreds of thousands, and that's a good outcome too.

→ Write to me and I'll verify the payback for your operation in an initial audit