How we judge

Three tiers would be tidier and would hide the two most common real answers. There is a large difference between “the agent does the work and you approve it” and “the agent does the filing and you do the thinking”, and collapsing them is how people end up disappointed.

YES19

An agent does the whole job. You look at the output, not the work.

MOSTLY25

An agent does the work and hands you a decision. Minutes a day, not hours.

HALF14

An agent clears the mechanical majority. The part people are paid for stays yours.

NOT YET7

An agent can assist around the edges. Handing it the job would be a mistake.

1. 01

   Break the job into steps

   Nobody has a job that is one task. Accounting is nine things. The step list is the actual work of the page, and the verdict at the top is only a summary of it.

2. 02

   Rate each step on capability, not on hype

   The test is whether a competent person could set this up with today’s tools and stop checking it daily within a month. Demos do not count. Roadmaps do not count.

3. 03

   Then rate it again on consequence

   Payroll is not technically harder than bookkeeping, and it is rated lower, because a mistake hits a colleague on a fixed date and is painful to reverse. Where the downside is severe and irreversible, the verdict drops.

4. 04

   Check what the law says

   Some steps are not a capability question at all. Signing accounts, making hiring decisions, giving legal advice and clinical triage are constrained by regulation regardless of how good the model gets.

5. 05

   Summarise honestly

   The job verdict is roughly the weakest link that matters, not the average. A job with eight excellent steps and one you cannot afford to get wrong is not a YES.

There is a second question behind the first one: if an agent can do the job, what are you still paying for? The software pages use the same four tiers, because the reader should only have to learn one scale — but the consequence is a billing decision rather than a delegation decision.

YESCancel it

The product was the labour. An agent and somewhere to put the data covers it.

MOSTLYDowngrade it

You still want the storage and the interface. You no longer need the tier you are on.

HALFKeep it as the record

The agent does the work inside it. The tool stays because that is where the data lives.

NOT YETKeep it

The moat is real — a network, a dataset, or a liability someone else carries.

One pattern does most of the work here: an agent replaces the labour, not the data. Products sold on the hours they save you — the scheduling link, the receipt scanner, the AI writing wrapper, the meeting notetaker — have very little left. Products that own a ledger, a network, a payment licence or a compliance obligation are untouched, because none of those are things you can prompt your way into.

Which is why so many software verdicts are half: keep the tool, cut the seats. The bill usually was not the licence. It was the people the licence was counting.

Prices are indicative list rates for a small team at the tier most people end up on, in euros. Software pricing changes constantly and almost nobody pays list. The cancel-list total counts only products rated cancel it — including the downgrades would inflate it with money you would still be spending.

Hours per week are an estimate of what an agent takes off one person doing that job full time, counting only the steps rated agent-ready. Costs are indicative ranges for a small European business — a bookkeeper in Amsterdam and one in Lisbon are not the same line item, and neither is yours. Treat both as orders of magnitude, not quotes.

The total on the front page adds up every job on the site. No business does all sixty-one, so read it as the size of the category rather than a promise about your calendar.

This site is made by Ogento, who build agents for a living. That is a conflict of interest, so it is worth saying plainly what we do about it: the verdicts are written to be right rather than to be useful to us. Of 65 jobs, 21 (32%) are rated half or not-yet, and 7 say plainly that you should not hand the job over at all. A page that oversold an agent would cost us more than it earned, because the person who believed it would be back within a month, angry.

These verdicts have a shelf life. Several of them were a tier lower eighteen months ago and will move again. If your experience contradicts a page — either because you have an agent doing something we said it could not, or because one of our YES verdicts blew up in your face — that is the most useful thing you could tell us, and the second case more than the first.

Every page carries the date it was last reviewed. This one: 2026-08-15.