Evidence & grounding
The assistant is built to be trustworthy on the record in front of it. It answers from what it can actually establish about the account or claim, and tells you when it can't establish something — rather than filling the gap with a plausible guess. This page explains the rules it follows, so you know how far to trust an answer.
Answers cite their evidence
Every factual claim traces back to something in the record: the event it came from, the acting user, the timestamp, and before → after values where they exist. When you ask who changed something, the answer leads with the change and then attributes it.
Names, amounts and dates
- Names. User IDs are resolved to real user names, so answers say who acted rather than showing an internal identifier.
- Amounts. Money is shown formatted, in the claim's own currency.
- Dates. Timestamps are rendered in your own timezone — the assistant is told your browser's timezone with each question and quotes dates as given rather than converting or estimating them.
If a tool gives no date for something, the assistant says the time is unknown instead of inferring one.
It reports what changed, not why
This is the rule most worth understanding. The assistant will tell you that a metafield changed, when, from what to what, and who changed it. It will not explain why it was changed, or what the field "typically" means, unless an event explicitly records that reason.
So for a metafield like withFailedCMS flipping to true, you'll get the date, the before → after, and whether it was a user or an automated change. You won't get "this is usually set after a delivery failure" — the event doesn't record that, so saying it would be an invention even hedged with "usually".
The one deliberate exception is why a claim is overdue, where no reason is stored anywhere and the assistant narrates the likely driver from the timeline. That answer is interpretation, and reads like it.
It doesn't do its own arithmetic
The assistant quotes the figures the system computes — each claim's own amounts, and the system's outstanding total per currency — rather than adding claims up itself. Totals in an answer are values the platform calculated, not the model's mental maths.
Nothing is carried over from earlier messages
A value mentioned earlier in the conversation — including one the assistant itself reported a minute ago — is treated as stale. Before it states, proposes or confirms anything about a record's current status or a field's value, it re-reads that value from the record.
This matters most in a long session: if you resolve a claim, or someone else changes a field while you're chatting, the assistant's next answer reflects the record, not the transcript. It only reuses earlier messages as a source of which claim or field you're talking about.
Claims are kept apart
Because the conversation is shared across the account and all its claims, the assistant internally marks which claim each earlier turn was about. A finding about one claim is never carried onto another — each answer is established from the record you're asking about now.
Plain language, not internals
Answers explain mechanics in business terms. The assistant deliberately avoids listing raw internal identifiers, blueprint IDs, metadata flag names or condition expressions. Where a claim sits in a strategy with many routing branches, it tells you the outcome for your claim — which branch it matched, or that none did so it took the default path — rather than enumerating the rules.
When there's no evidence
If the data doesn't support an answer, the assistant says so plainly. That's a feature, not a failure: "I can't find an event recording that" means the record genuinely doesn't show it, and you should look elsewhere rather than assume the assistant missed it.
Where the timeline doesn't explain a failure, it falls back to searching the relevant service's logs — and if those turn up nothing, it tells you the logs showed nothing instead of guessing a cause.
Off-topic questions
Ask something the assistant can't investigate and it steers you back to what it can answer — history, attribution, summaries and lookups about this record. Ask something it won't help with and you'll see "I can't help with that request. Try rephrasing, or ask me something about this account."
If an answer surprises you, ask the assistant which event it came from. It's designed to show its working.