There’s a particular problem with AI-generated summaries of professional documents: they can be right and unverifiable at the same time.

A lease abstraction that says “base rent: $28.50 per square foot” is useful. A lease abstraction that says “base rent: $28.50 per square foot (Section 4.1, page 3)” is trustworthy. The number is the same. The difference is that the second one can be spot-checked in thirty seconds. The first requires trusting the output, which professionals in high-stakes document workflows don’t tend to do with AI systems they haven’t verified yet.

Source citations solve the trust problem without solving the accuracy problem. They don’t make the AI more accurate. They make accuracy checkable. That’s a different thing and in many ways more valuable, because it means the analyst can build trust incrementally through verification rather than having to decide to trust the whole output at once.

This matters more in professional workflows than in consumer applications. A consumer asking an AI to summarize a movie doesn’t need to know which scene a plot point came from. An analyst reviewing a lease before a closing needs to know exactly where in the document the renewal option clause appears, because they may need to pull that clause for the purchase agreement, or discuss it with the counterparty, or verify that the abstract accurately captures an unusual provision.

Source citations also change what the tool is for. Without citations, the tool produces outputs. With citations, the tool produces auditable outputs. Auditable outputs can be shared, reviewed, archived, and challenged. They become part of the deal record rather than a throwaway summary.

The tools that have figured this out have built source citations into their core output, not as an add-on. The tools that haven’t are one analyst complaint away from realizing why it matters.

Cite your sources. It’s not optional. +++