Every vendor in this category says their summary is grounded. Almost none of them will tell you what that means operationally, and the ones who will usually mean "we put citations in the output".
Citations are the easy part. Grounding is a property of the whole pipeline: what was retrieved, what was excluded, what the model was allowed to conclude from it, and whether a reviewer can reconstruct the path six months later when the case is challenged. If any of those four is missing, the summary is not grounded. It is plausible.
Here is how I would test it.
Retrieval coverage. Take a closed case with a known outcome. Ask the system to produce a summary and then ask it, explicitly, which sources it did not consult. A grounded system can answer that, because retrieval is a separate step with a defined source list. A system that cannot answer is generating from a prompt, not from evidence. The design question is spelled out in RAG in AML investigation, and the test is simple: if the source list is not a first-class object, you cannot audit the output.
Exclusion logic. Most real investigations are shaped by what was left out. The transaction that looked odd but had a documented explanation. The prior case that was closed with a clear rationale. A summary that cites everything and weighs nothing is not analysis. Ask the system how it handled a specific exclusion. If the answer is a sentence rather than a reference, you have a writing assistant.
Claim-to-source mapping. Every factual sentence in the summary should map to a source record, and every inference should be labelled as an inference. This sounds pedantic until the first time a supervisor asks why the summary says "unusual counterparty jurisdiction" when the customer had declared that corridor. That single error, repeated across a portfolio, is how a model risk finding starts.
Temporal integrity. AML cases are time-bound. A summary written today about activity from fourteen months ago must not import today's screening result as if it were known then. Systems that rebuild the customer profile at query time rather than at case time will silently rewrite history. This is the least discussed failure mode and the most damaging in a SAR context.
Regeneration behaviour. Ask the same question twice. If the summary changes materially with no new evidence, you do not have a controlled process. Some variation is fine in prose. Variation in the findings is not.
Now the harder point.
Grounding is not only a retrieval problem. It is a memory problem. An investigator reaching a conclusion needs to know what the institution decided in similar cases, not just what the transaction data says. That is investigation organisational memory, and it is usually the last thing built because it is the hardest to model. Cases are messy, rationales are inconsistent, and the older records are often PDFs.
If you skip it, you get a system that is excellent at describing a single case and useless at explaining why this case is different from the last twenty. Analysts notice within a week. They start opening the old cases in a second window, and the productivity gain quietly disappears.
There is a regulatory dimension too. Model governance expectations do not stop at "the model works". They extend to whether you can explain the output, evidence the oversight, and reproduce the decision. FCA expectations for AML AI governance is the practical frame here: documentation, validation, monitoring and a named owner. A citation in a PDF is not a control. A retrievable evidence chain with a human sign-off is.
Two implementation notes, from having watched this go wrong.
First, log the retrieval, not just the generation. Store the query, the source list, the version of each source and the timestamp. Storage is cheap. Reconstructing a case from memory is not.
Second, make the summary rejectable. If the reviewer can only accept or edit, you cannot measure whether the summary was any good. Give them a one-click "not grounded" flag with a reason, and review those weekly. That single field will tell you more about model performance than any benchmark, because it captures the cases where the model was confidently wrong.
The conclusion is unglamorous. Grounded means reproducible. Reproducibility is also the throughput argument: reproducible output is a throughput argument is the commercial version of the same point. If a different reviewer, six months later, cannot follow the same evidence to the same conclusion, the system is generating text about a case rather than investigating it. Everything else, including the demo, is decoration.
Top comments (0)