On September 29, OpenAI held its developer event, DevDay, in San Francisco, and the dominant theme was agents. According to the trade press, Anthropic and Microsoft are also working on similar features. Translation: in the coming weeks someone will offer you an AI agent for your practice, your shop, your calendar. It pays to arrive with clear ideas. This guide isn't selling you an agent: it helps you understand whether you need one.
What an agent is, without jargon
A chatbot answers questions. A traditional automation runs a fixed sequence: an email arrives, the data goes into a spreadsheet, a confirmation message goes out. An agent sits in between: you give it a goal and some tools, and it chooses the steps itself.
- Chatbot: it answers you, and that's it.
- Automation: it always does the same thing, in the same order, and is as predictable as a clock.
- Agent: it reads, decides, uses other programs and writes. For example: "sort the quote requests, reply to the simple ones and alert me about the others".
Tools are the things the agent can touch: the inbox, the calendar, the management system. The more tools you give it, the more it can do, and the more it can get wrong.
The interesting part is also the risky one: if it chooses the steps itself, it can choose badly.
Where an agent works
It works when the task repeats, the rules are clear and the input changes every time. Customer emails, for example, arrive written in a thousand different ways: a classic automation gets lost, an agent understands them.
A hypothetical example. A practice receives about twenty requests a day, by email and chat, about hours, prices and availability. An agent can answer the simplest ones, put the others in a prioritized list and prepare a draft for those that need a person. The person decides, the agent prepares.
Other suitable tasks:
- sorting incoming requests and answering frequently asked questions;
- preparing the draft of a report from data scattered across several tools;
- updating the management system by reading an email or a PDF.
The common criterion is that, if it gets it wrong, the damage is small and quickly repaired. A reply to correct is a five-minute problem. A wrong refund is not.
Where it's better to leave it alone
- Money and legal consequences. Quotes, refunds, contracts: the last word stays with a person.
- Confused processes. If even you don't know how you do it, the agent automates your confusion, only faster.
- Tasks solved with an automation. If the steps are always the same, a classic automation usually costs less and surprises less.
- Sensitive data. Before putting it in the loop, ask where it ends up. There are also models that run on your own server, where the data doesn't leave.
The three-task test
Before buying anything, do this exercise, which costs a pen and ten minutes.
- Write down the three tasks you repeat every week.
- For each one, ask yourself whether the input is always the same (an automation is enough) or changes every time (an agent may help).
- Ask yourself what happens if it gets it wrong. If the answer is "nothing serious", it's a good candidate.
An example: "answering customers who ask about hours" has a changing input and a low risk, so it's a good candidate. "Issuing credit notes" has a high risk, so it isn't.
Then start small: two weeks, a single task, and one person checking what the agent did before trusting it.
Human oversight, without becoming bureaucrats
A good way to begin is draft mode: the agent prepares, the person approves. After a few weeks you look at how many drafts you approved without touching them. If it's almost all of them, you can let it send the simplest cases on its own and keep checking the others. Autonomy widens one task at a time, and only where the facts prove you right.
Five questions for whoever is selling it to you
- What does it do when it doesn't know what to do: does it stop and alert, or improvise?
- Can I see what it did, step by step?
- Where does my data go and who can read it?
- How much does it cost at steady state, not just the first month?
- Who fixes it when one of the connected tools changes?
What we know, and what we don't
For now we know what has been announced. For the products presented this week we don't yet know how much they will cost or who they will really suit, and it's better to wait for the numbers before getting excited. Agents are one more tool, to be used where the task deserves it.
Sources: CNBC, DevDay 2026 live coverage and Crypto Briefing, preview of OpenAI's Managed Agents.
Have a similar case or a doubt? Write to us at info@sarabi.cloud.