Short answer
Professionals are accountable for what they sign. Software for them must show its sources, be auditable, fail visibly rather than silently, and never require the user to spot a subtle error to stay safe.
The asymmetry
If a consumer app recommends a mediocre restaurant, the cost is one dull evening. If a drafting tool omits a required clause and a professional signs the act, the cost is a liability they carry personally, for years.
That asymmetry changes the design brief. The relevant question is not "how often is it right?" but "when it is wrong, how quickly does the user find out, and what does it cost them?"
Four things these users need that consumers do not
Visible sources
A professional will not act on an answer they cannot check. Every substantive output should show what it was based on, close enough to click. This is not a trust badge, it is how the work gets verified.
An audit trail
Regulated professions must be able to reconstruct what happened and when. If your tool touches the work product, it needs a history: what was generated, from what, when, and what the user changed.
Visible failure
Silent degradation is the worst property professional software can have. If retrieval fails, say so. If the model is uncertain, surface it. A blank space is safer than a confident guess.
Reversibility
Nothing the tool does should be hard to undo. Professionals will only adopt something they can back out of, because the alternative is trusting software with work they are answerable for.
What we learned watching people use it
Two things consistently surprise teams building for this audience.
First, the drafting is not always the valued part. Structure often matters more: knowing which documents are required, which are outstanding, and what has been checked. A checklist that is always correct can be more useful than generation that is usually good.
Second, speed is not the main draw. These users are not trying to work faster at the cost of certainty. They want to stop worrying about omissions. Software that saves time but adds a new thing to double-check has made the day worse, not better.
What this means for how you build
- Ship narrow and reliable before broad and impressive.
- Make refusal a designed state, not an error.
- Show your work by default, not behind a toggle.
- Treat every silent failure as a serious bug.
The engineering behind the first two is covered in stopping a model from inventing a citation.
Frequently asked questions
What makes software for legal professionals different?
The user is personally accountable for the output, so approximate correctness is not acceptable. The software must show its sources, keep an audit trail, fail visibly rather than silently, and make every action reversible.
Do professionals want AI to work faster or more accurately?
In our experience, certainty over speed. A tool that saves time but introduces something new to double-check has made the day worse. Reliable structure often delivers more value than impressive generation.
Should an AI tool for professionals ever refuse to answer?
Yes, and refusal should be an explicitly designed state. If the system cannot find a basis for an answer, saying so is far safer than producing a fluent guess the user may act on.