A plausible answer is not a checked file

If AI reads a mortgage deed and returns a date, has it completed the task?

Only if the date is right, it belongs to the right event and it has been checked against the source. A plausible answer is not enough.

This is the validation gap I keep coming back to. In a conveyancing file, a small difference can matter. A name may include an extra word. A date may look reasonable but be wrong. The system needs to distinguish between a match, a mismatch and something it cannot establish.

The same problem appears in post-completion work. A requisition may have been answered in part, with points still outstanding. If the system only tells you 'responded', the useful detail has been lost. What remains open? When was it last chased? How close is the deadline?

My starting point is cautious: if the required check does not pass, do not quietly mark the file as complete. Keep the source visible, explain the exception and route it to someone who can decide.

That does not mean every difference is a legal problem. It means a difference should not disappear simply because the output reads well. The review process should separate harmless variations from issues that need action.

The next step is to learn from those exceptions. Which checks fail repeatedly? Is the problem the source data, the extraction or the rule itself? Improve each part without removing the need for judgment.

To me, intelligent case management is a system that knows what it has checked and what it has not.