In plain English
- What happened
- Chat assistants from OpenAI and Anthropic currently let a business upload files such as reviews, survey exports or support transcripts and ask for a summary of recurring themes, and Anthropic's Projects feature works the same way inside Claude. Neither tool goes out and asks customers anything new; both work only on text a business already has and feeds in.
- Why it matters to your business
- A busy owner who has a stack of reviews and no time to read them can genuinely save hours this way, but the summary is only as good and as representative as the sample fed in. Treating an AI summary as proof of what customers want, rather than a shortcut for reading what a handful of them already said, is where this usually goes wrong.
- Do this
- Export a batch of actual customer feedback — reviews, survey responses, or support emails — into one document, and ask a chat assistant to list the three most common phrases in it before deciding whether the pattern holds.
The honest answer is narrower than the question suggests. An AI assistant cannot tell a business what its customers want in any independent sense — it has no access to anyone's unspoken preferences. What it can do is read through text a business already has and pull out what keeps showing up.
What the tools actually do
OpenAI's current file upload feature lets a user paste or attach documents inside a ChatGPT conversation and ask for a summary or analysis of what is in them, according to OpenAI's own file uploads guide. Anthropic's Claude offers a similar capability through what it calls Projects, described on Anthropic's support page as a way to keep a set of files and reference them across a conversation.
Both are pattern-matching over the words provided, not research. If a business feeds in fifty reviews, the assistant can group them by complaint or compliment and surface the phrases that repeat. If a business feeds in five reviews, it can do the same thing, but the result is five opinions dressed up to look like a trend.
Where this goes wrong
Reality check
The risk is not that the summary is wrong about the text it was given — it usually is not. The risk is presenting a summary of a small, self-selected sample (say, only the five-star reviews someone happened to paste in) as if it were a finding about the whole customer base.
There is also a marketing-claims angle worth knowing about. The Federal Trade Commission has warned businesses to be able to back up any claim about what their product does or what customers think, whether that claim came from a person or an AI tool. A summary an assistant produced from a handful of reviews is not something to quote in an ad as 'what our customers are saying' unless it can actually be backed up with the underlying reviews.
What to try instead
- 1
Gather the source material in one place
Export reviews, survey responses, or support tickets into a single document rather than working from memory or a few recent examples.
- 2
Ask for themes, not conclusions
A prompt like 'list the three most common phrases in these reviews' produces something checkable. 'Tell me what my customers want' invites the assistant to guess.
- 3
Read a sample of what it points to
Pull up five or six of the actual quotes the summary is built on and read them directly, to confirm the pattern is really there.
- 4
Note how small the sample is
If the input was twelve reviews out of a few hundred customers, say so before acting on the summary or repeating it to anyone else.
- 5
Re-run it periodically
Feedback themes shift with the season or with a product change, so a summary from six months ago is not a standing fact about current customers.
Used this way, the payoff is time saved reading, not a new source of insight the business did not already have. A summary can point to the three complaints that show up most often in the last quarter's reviews faster than reading all of them by hand — it just cannot tell a business anything its customers have not already said somewhere in writing.
Sources
