AI agents in business: what actually pays off and what is a well sold promise
The short version. The average small business now runs around five AI tools, most often for research, marketing, customer support and admin. Market figures point to 40-60 % time saved on repetitive work and payback within 60-90 days. They also mostly come from the vendors selling those tools, so treat them as orientation, not a promise.
Plainly: AI pays off where you have many identical tasks with a clear correct outcome. Everywhere else it is an expensive toy.
Where it works
Most activity sits in support, sales, finance and operations, where volume is high and results are measurable. For customer support specifically the reported numbers are:
| Metric | Reported figure |
|---|---|
| Time saved on repetitive tasks | 40-60 % |
| Cost per machine-resolved ticket | roughly 1-2 EUR |
| Cost per human-handled ticket | roughly 6-12 EUR |
| Enquiries handled without escalation | around 89 % |
| Break-even at 500 tickets per month | just under seven months |
Even if these were optimistic by half, the gap between one euro and ten per resolved ticket is still worth attention.
Where it does not work
- Without volume. Twenty emails a month is not an automation case.
- Where the correct answer is unclear. If people cannot judge it, a machine will not.
- On messy data. An agent on top of chaos produces faster chaos.
- On consequential decisions. Credit, HR, health. That is also where regulatory duties appear, see AI Act transparency.
How to verify before you buy
This is the only part of the article you really need.
- Take one process that repeats at least a hundred times a month.
- Measure the current state. Minutes per case and cases per month. Without this number any vendor can prove anything to you.
- Calculate the ceiling. Minutes times hourly cost. That is the maximum, and reality is usually half.
- Run it as a one month pilot on that single process, not across the company.
- Compare using the same measure as step two. Not a feeling.
Payback under six months means it is worth doing. If not, you just saved yourself a project.
Off the shelf versus custom
A ready made tool is fast and cheap to start, but you hit a wall when you need integration with your own systems or when the vendor changes pricing. Custom costs more upfront and gives control.
The sensible approach is usually a mix: off the shelf for what is the same everywhere, and a custom business application for what makes you different from your competitors. That difference is your advantage, and no generic box will deliver it.
FAQ
Will it replace people? In practice it removes routine. Reported cases describe handling routine enquiries without escalation, not disbanding teams.
How fast is payback? 60-90 days is quoted for simple deployments. Treat it as the best case.
Are those numbers trustworthy? Mostly vendor sourced. Hence the measurement method above. Your own number is the only one that carries weight.
Where should I start? With the highest volume, most boring process. Not the most interesting one.
Figures come from public 2026 market overviews, largely vendor reported. Treat them as orientation, not a guarantee.
