Start with your invoices or your inbox: high volume, fixed rules, mistakes reversible. Never start with customer contact. The best first agent is boring, measurable, and live within a few weeks. That sounds unambitious, and that’s exactly why it succeeds where the flashy pilots run aground.
Because most first agents don’t fail on the technology, they fail on the choice of process. Choose wrong and even a good agent collapses. Choose right and even a mediocre agent works. Here are the three properties you select on, and the trap most companies walk into.
Why not start with customer contact?
Because customer contact combines everything a first agent should avoid: emotion, exceptions, and liability. An angry customer isn’t a pattern a language model tidily handles, it’s a relationship at risk. And what your bot says, you said: Air Canada had to honor a discount its chatbot invented, Forbes reported, and the German Higher Regional Court of Hamm ruled that a company is directly liable for its chatbot’s statements.
Klarna showed how that plays out at scale. The company said its AI assistant performed work equal to roughly 700 FTE and cut handling time from eleven minutes to under two. The CEO later acknowledged that the emphasis on cost had affected quality and reinvested in human support for complex cases. Customer contact is the final exam, not the first lesson.
Which three properties make a process suitable?
A good first candidate scores on all three. Miss one and you keep looking.
High volume. An agent pays off on repetition. A task that comes up three times a month isn’t worth the setup, even if the agent handles it fine. Look for the work that shows up daily and that nobody enjoys.
Fixed rules. The more explicit the rules, the better the agent performs. “If the invoice matches the purchase order, book it; if it doesn’t, set it aside” is a rule. “Judge whether this complaint is reasonable” isn’t a rule but a judgment, and judgments belong with humans for now.
Mistakes reversible. This is the property that gets skipped most often. If the agent makes a mistake, how bad is it, and can you undo it? A mislabeled email you fix in ten seconds. A wrongly paid invoice or an invented promise to a customer costs money or trust. Start where a mistake is a correction, not a damage.
So which processes actually make good first candidates?
Four examples that almost always score on all three properties.
- Invoice processing. Match incoming invoices to purchase orders, book them, and set the exceptions aside for a human. High volume, hard rules, and an exception is exactly the kind of mistake you want: visible and reversible.
- Email triage. Sort, label, and route incoming mail to the right department. Knowledge workers spend an average of about 2.5 hours a day on email, a widely documented figure, so the time saved is real and the risk is low.
- Document extraction. Pull data from contracts, receipts, or forms and drop it into a system. Boring, error-prone when humans do it, and easy to spot-check.
- Internal knowledge questions. An agent that answers colleagues’ questions about internal procedures. Mistakes are reversible because it stays internal, and no customer is on the line.
Notice what these four have in common: none touches a customer directly, and in all four a mistake is a rule you adjust rather than a damage you compensate.
How do you know within a few weeks whether it works?
By deciding up front what you measure success against, in hours or euros, not in enthusiasm. That sounds obvious and is not: a field report from Project NANDA at MIT said 95 percent of the organisations studied saw no rapid GenAI P&L impact. The report underlines the importance of workflow integration and a concrete yardstick.
External partnerships outperformed internal builds in the implementations studied. That does not prove guidance automatically doubles your odds, but it is a reason to bring in relevant experience if you have not done this before.
Practically: pick one process, count how many hours it costs now, agree which saving counts as success, and set a date. Miss the yardstick and you stop without it hurting, because you deliberately chose something small and reversible.
The choice in one sentence
The best first agent is the one nobody wants to give a presentation about: boring, measurable, and live fast. In today’s labor market that’s often enough, because per CBS, the Dutch national statistics office, automation or AI is the primary measure in 2026 for nearly half of companies facing staffing shortages. A boring agent that catches daily work beats an empty chair.
Renforza places agents as the third candidate on your shortlist, and in an agents intake we start exactly here: which process scores on high volume, fixed rules, and reversible mistakes. Not sure which of your processes is the best first candidate? That’s the question the intake exists for. More on how we look at these choices lives on our page for employers.
