Technology
Almost every disappointment with these tools traces back to asking for something in the second category: arithmetic, current fact, or authority.
The current generation of these tools is excellent at language and unreliable at arithmetic, authority, and knowing what it does not know. Almost every disappointment traces back to asking one for something in the second category.
Drafting from material you supply. Summarizing long documents. Rewriting the same explanation for a different reader. Finding the passage you half-remember. Turning notes into prose and prose into notes.
These share a shape: you already have the substance and the work is transformation.
Numbers that must be right. A model produces text that looks like a calculation. Sometimes it is one. You cannot tell which from the output, and the wrong ones are formatted as confidently as the right ones.
Current facts. Contribution limits, bracket thresholds, and rules change. A model will state last year's number in this year's sentence without hesitation.
Anything requiring a source. Asked for a citation it does not have, a model will often produce one that does not exist. This is not lying; it is the same generation process producing a plausible string.
Judgment about a particular client. It has not met them, does not know what they did not say, and will happily reason from the half of the picture you typed.
Never let a number or a rule reach a client without a human having confirmed it against a source. Not because the tool is usually wrong — because when it is wrong, nothing in the output looks different.
Treat it as an extremely fast first draft from a capable assistant who has never met your client, has no access to your files, and is too confident. That framing produces the right instincts without requiring you to understand the technology.
Questions this did not answer? Ask them directly — that is what the twenty minutes is for.
Schedule a call to see for yourself