The market pays more for agents right now than for anything else in applied AI. The skill "agentic AI" grew 280% in US job postings in a year, to roughly 90,000 listings, according to Stanford AI Index 2026. An engineer who builds agents earns 15 to 20% more than a comparable ML engineer, per the KORE1 hiring survey for 2026. And the same market, according to Gartner, will cancel more than 40% of agentic AI projects by the end of 2027.
I don't see a contradiction there. The premium is paid for a rare skill. The cancellations hit projects where an agent was put somewhere it wasn't needed, or where nobody could prove it worked. When someone asks me for an agent, my first question is usually not "which agent" but "do you need an agent at all".
When plain automation wins
There is a whole class of tasks where a language model is simply not needed. The signs:
- the input is structured: a form, a spreadsheet, an API, an email that follows a template;
- the rules fit on one page and don't change every week;
- the answer has to be identical every time, and a mistake is expensive;
- the volume is high, but there are few decisions per unit of work.
Moving orders from the CRM into accounting. Payment reminders. Parsing a supplier price list that always comes in the same format. Assembling a weekly report from three sources. All of that is a parser, a rule, or a scenario in any automation platform. It is cheaper, it behaves the same way every time, and it has no token bill.
I would not put an agent on a task you can draw as a flowchart. The agent will be more expensive, slower and less reliable than a plain if-then.
When an agent earns its cost
An agent belongs where the input is messy and there are many small decisions.
Inbound emails to a distributor, where one message contains an order, a complaint and a question about delivery dates. Leads who write in a messenger in free text. First-line support where the questions repeat but the wording is new every time. Invoices, contracts and requests that arrive as PDFs, photos and scans.
What these have in common: the rules are too many and too fuzzy to write down. A person handles it through judgment. An agent reproduces that judgment well enough to take most of the routine off the person and leave them only the ambiguous cases.
The second sign is a chain of actions. The agent doesn't just classify an email. It finds the order in the database, checks stock, drafts a reply and creates a task for the manager. Each step on its own is trivial. The value is in linking them.
How to measure ROI
I use four numbers, and I insist they get measured before launch, not after.
- Hours. How much time staff spend on the process today. Not a guess of "about half a day", but a one-week measurement. Multiply by the fully loaded hourly cost.
- Response time. How long a customer or a colleague waits. In sales this is conversion, directly: a lead who got a reply in a minute and a lead who waited until morning behave differently.
- Error rate. Orders typed in with a mistake, invoices sent to the wrong place, requests that got lost. Errors have a price, and usually nobody has counted it.
- Revenue being lost right now. Leads nobody answered. Requests that arrived after 6 pm. Repeat sales nobody got around to.
If the sum of those four lines is smaller than implementation plus a year of maintenance, the project should not happen. That is a normal outcome of an audit, and I say so to clients.
Evaluation first, agent second
In the KORE1 survey, employers named the top skill of an agentic engineer as neither prompts nor frameworks but the ability to build an evaluation harness: a set of test cases the agent is checked against automatically after every change.
So before the agent is written, you need real examples from the process with the correct answers, ideally at least a hundred. Emails and the way your best manager replied to them. Documents and what should have been extracted from each. The agent runs against this set every time the instructions, the model or an integration changes.
Without it you don't know whether the agent got better or worse after an edit. You find out from customers. The projects Gartner counts in the cancelled 40% are very often exactly this: the demo worked, production drifted, and nobody could say why.
Data matters more than the model here. If there are no examples because nobody ever recorded the process, the first days go into collecting them. Boring work, and it decides the result.
Guardrails and handoff to a person
An agent in production has to do three things: never do what it isn't allowed to do, stop when it is unsure, and hand the case to a person in a way that lets the person understand the context in ten seconds.
In practice this is a list of allowed actions (reading an order is fine, changing a price is not), confidence thresholds below which the agent opens a task for an operator instead of answering, and a log of every action. Plus a spending cap, because an agent stuck in a loop can burn a month of tokens overnight. That one is not a scare story.
The operator doesn't disappear. They stop doing the routine and start handling exceptions. Usually it is the same person who did it by hand before, and in the first weeks they are the one who spots where the agent is wrong. That is the most useful feedback in the whole project.
Three steps to decide
- Audit the process. One or two weeks: where the input comes from, who does what with it, where time and money leak, whether there is data for evaluation. The output is a list of processes with an estimated effect and a verdict for each: rule, agent, or leave it alone.
- Pilot one process. Two weeks, a first working version on real data, with an evaluation set and guardrails. Not five processes at once. One.
- Measure. The same four numbers as before the start. If they hold, extend to neighboring processes. If not, stop, and the loss is limited to the pilot.
This order is how my services are built: an AI readiness audit from 450 EUR, an agent for one specific process from 1,500 EUR, first working version in two weeks. Why the market pays a premium for agentic engineers, and what that means if you are the one buying, is in my review of the AI job market in 2026.
If you have a process that eats hours and you are not sure an agent belongs there, send me a short description: what comes in, who handles it, how long it takes. I will reply with what I would do in your place.
Originally published at ilinmaks.com.
Top comments (0)