Research agents arrive: what they are genuinely good for, and where they mislead
OpenAI launched deep research on 2 February, days after o3-mini. A model that works for twenty minutes produces a new kind of output, and a new kind of error.
On 2 February 2025 OpenAI launched deep research, a mode in which the model browses the web over many steps and returns a cited report. Two days earlier it had released o3-mini, a cheaper reasoning model, made available to free users as well.
The pattern is the same in both: spend more time and computation on one answer. With deep research the time is minutes rather than seconds, and the output is a document rather than a paragraph.
Where this is genuinely useful for an SME
Market and competitor groundwork. « Who else in German-speaking Switzerland offers this service, at what positioning » is a real half-day of work that comes back in twenty minutes with sources you can check.
Regulatory orientation in an unfamiliar area. Not legal advice — orientation. What the rules broadly are, what the terms mean, which questions to bring to a lawyer so the meeting is short.
Supplier and technology surveys. Who makes this component, in which countries, at what rough price level.
In each case the value is the same: it replaces the first, tedious layer of a research task, and it leaves you a citation trail.
Where it misleads
It reads what is public and confident. A well-optimised vendor blog outranks a quiet expert. The report inherits the biases of what is easy to find, and a polished twenty-page output makes that invisible.
It cannot know what is not written down. The reason your industry does something a particular way is often in nobody’s blog post.
The format signals more authority than the content earns. This is the real risk. The same claim in a chat reply invites scepticism; in a structured report with footnotes it reads as established.
The rule that makes it safe
Open three sources from the report at random and read them. Not all of them — three.
If those three say what the report says they say, treat the rest as a solid starting point. If one does not, treat the whole document as a list of leads rather than a set of findings. That check takes five minutes and it is the entire difference between a useful tool and a confident one.