Direct answer: Why should AI finance answers cite sources?
AI financial answers need citations, freshness signals, source boundaries, and clear uncertainty so users can inspect the evidence instead of accepting an unsupported answer.
How Celestice helps
Celestice makes citations and source state part of the answer workflow. In practice, advisors can see which document, account state, research artifact, or policy source supported an answer before it becomes a recommendation or client-facing explanation.
<!-- celestice-query-answer:end -->The problem with confident, unsourced answers
Large language models are fluent, and fluency is dangerous when it is wrong. An AI that answers a question about your portfolio policy, a tax rule, or a regulatory requirement in a confident paragraph — with no indication of where that answer came from — is asking you to trust it blind. In casual use that is a nuisance; in wealth management it is a liability. The antidote is grounding: answers built from governed source material, with citations specific enough to verify.
Knowledge is not memory, and not account data
A useful distinction underpins this. A wealth platform deals with three different kinds of information that are easy to conflate:
- Structured account data — your holdings, balances, and transactions: facts from systems of record.
- Memory — what the assistant has learned about your situation and preferences over time.
- Knowledge — governed source material: documents, policies, research, regulatory references, internal docs, third-party captures, and reviewed wiki pages.
Knowledge is the body of reviewed reference content an answer can be drawn from. Keeping it separate from memory and from account data is what makes it possible to cite — because a knowledge answer points to a document, not to a vague recollection or a raw data row.
What makes a citation actually useful
"According to our research" is not a citation. A genuinely useful citation carries enough locator detail to find the original — the specific document, section, chunk, web capture, wiki node, or app surface the claim came from. Beyond the locator, a good cited answer exposes the metadata that lets you judge whether to trust it:
- Source name — what document or system it came from.
- Source type — a policy, a research note, a regulatory reference, an internal doc, a third-party capture.
- Freshness — how recent the source is, because a tax figure from two years ago may be obsolete.
- Trust level — how authoritative the source is, since an official policy outranks a scraped web page.
Those four signals turn a citation from decoration into a tool for judgment.
Freshness and trust: the signals that prevent quiet errors
The two most underrated citation signals are freshness and trust. A confident answer drawn from a stale source is one of the most common ways AI quietly misleads — the tax bracket changed, the policy was updated, the regulation was superseded, but the model cited the old version with full confidence. Surfacing freshness lets a reader catch that. Trust level does the parallel job for authority: when sources conflict, the official, governed document should win over an unvetted capture, and the citation should make clear which kind you are looking at.
