Grounded AI for legal, clinical, and standards work

In regulated work the source matters as much as the answer. A clinician, a lawyer, and a standards authority are all doing a version of the same job: reaching a conclusion they can defend afterwards, to someone whose job is to check it.

That is a harder problem for AI than it looks, and it fails in a way that is easy to miss.

Why is a fluent answer not good enough here?

A general model produces text that reads like a citation whether or not the citation exists. It also has no reliable sense of which revision was in force on a given date, so it will answer from a superseded version with exactly as much confidence as from the current one.

Neither failure announces itself. This is silent omission: the answer looks identical whether the governing revision was read and set aside or never opened at all. The checking work falls back on the person who asked, and the time the tool was supposed to save goes into verification instead.

How does the grounding service handle provenance and time?

Every claim is traced back to the document and passage it came from, enforced on the retrieval side, so a specific claim either carries its source or it does not go out. You can open the source and read it in context rather than taking the answer on trust.

The service carries the date a document was in force, so “as of when” is a question it can actually answer. It keeps the native language of a document as the source of truth rather than working from a translation, which matters when the wording is the thing under dispute. And it returns a coverage report saying what was and was not included, so a gap is visible instead of silent.

What kinds of question does this make possible?

The queries that matter in regulated work are usually not lookups. They are coverage questions (“does anything in this corpus address X”), gap questions (“what is missing relative to this requirement”), and comparison questions across jurisdictions or revisions. Similarity retrieval is poorly suited to all three, because none of them are answered by finding the passage that most resembles the question.

Crafting Tomorrow, working across curriculum requirements in twelve jurisdictions, brought us exactly this set of queries. That corpus ran to roughly 8,200 documents, and measuring how little of it the best available model could actually work with is what started the company. It is a large part of why the service is built the way it is.

Where does a first project start?

With a teardown on a narrow, real slice of your corpus and the questions your current tools handle badly, followed by a scoped pilot if the teardown shows something worth pursuing. Data access, what cannot leave your environment, and what needs masking are settled at scoping rather than discovered later. The pilot page covers the process step by step.

If you are working with document sets like these, tell us which questions your current tools get wrong and we will tell you whether this is a problem grounding solves.