Skip to content
Help Center Back to AgenticObjects

Chat

Ask an agent a question in everyday language and get an answer you can check — with its trust badge, its declarations and the evidence behind every number.

App 6 min read Updated

Chat is where you ask a question in your own words and get an answer you can verify. Every answer shows how far its numbers can be trusted, which assumptions it made, and a path down to the evidence behind each figure.

What this screen is for

Each conversation is answered by one agent. That keeps everything consistent: the metrics the agent may reach, the context it carries from one question to the next and the evidence it records all belong to that agent. When you want a different agent to take over, you move the conversation openly, and the thread marks the exact point where the change happened.

Answers are governed: the engine runs the queries, checks the agent's figures against the results, and says plainly when something could not be verified.

What you see

  1. The left rail — your company, the workspace (a list you can search when you work in several), its Projects (All projects first), New chat, Search conversations, then 📌 Pinned and your conversations grouped by agent under Today, This week and Older. A conversation that is still being answered shows Answer in progress.
  2. The conversation header — company, project, agent and conversation title as a breadcrumb, with the Conversation menu (▾) at its end. When your role allows it, a selector beside it lets you choose the Model answering this conversation.
  3. The work trace — while an answer is being written, its steps stream live below your question. When it finishes, it folds into one line such as ✓ 6 steps · 4.2s · detail ▸, with a Simple Query or Agentic Loop stamp and, if you can open run records, a Run detail ↗ link.
  4. The answer — the headline figure and its comparison, the written answer, and a row of badges: ✓ Trust A (or B, C), Data freshness, and Scope when row-level security narrowed what was read. Declarations badges list the assumptions the answer used.
  5. Evidence — a button under the answer; it opens a drawer from the right (from the bottom on a phone). Inside: how the answer was computed, the bound measure, the filters and where each came from, the period's dates and the freshness; How these numbers were verified lists each figure with its verdict, and Show the query opens the executed statement in the same drawer for those who hold that permission. You never leave the chat; Esc closes the drawer.
  6. Next analysis? — suggestion chips under an answer, when the agent offers them.
  7. The message toolbar — Copy, Helpful and Not helpful, and how long ago the answer arrived.
  8. The composer — Ask a question over governed data… with the send button. In a new conversation, starter questions built from the agent's own metrics appear above it.

How to …

Start a conversation

  1. Press New chat and pick an agent, or press + beside an agent's name in the rail.
  2. Type your question, or press one of the starter questions — it fills the composer so you can read and adjust it before sending.
  3. Send, and watch the work trace as the agent works.

Answer a clarification question

  1. When your wording leaves something open — usually the period — a card with the agent's question and a set of options appears instead of an answer.
  2. Press an option, or type your own period into Another period (e.g. 2024, Q1 2023)… and press Send.
  3. The card records your choice as Answered: Last month, and the answer follows with the conversation's context intact.

Check where a number came from

  1. Press the figure: a small one-sentence card says what it is — the measure, the period and its dates, the scope, how far the data is complete and whether the engine verified it. For more, press More → on the card or the Evidence button under the answer. A figure the engine computed from other figures says the calculation instead: each figure it was computed from with its value, the operation, the period, and the engine's own result — for example "63.0% = Patient revenue (248,660,570) ÷ Revenue excl. VAT (394,608,578) · year to date · recomputed by the engine: 63.01%". A figure the record does not hold is named without a value.
  2. Read the bound measure and the period resolution — the two assumptions that explain most surprises.
  3. In the drawer, How these numbers were verified shows each figure, how it was produced and whether it is verified; coming from a card, that figure's row opens highlighted.
  4. To see the statement itself, press Show the query in the drawer; personal-data values arrive masked and every opening is recorded in the audit log. If it is locked for you, press Request Technical Review to ask the people who hold that permission to review it with you.

Tell us an answer was wrong

  1. Press Not helpful under the answer.
  2. In What went wrong?, pick a reason — for example Wrong method / path or Wrong or missing content — and optionally Add a note.
  3. Press Send feedback. Your feedback reaches the people who tune the agent.

How to read it

  • Read the trust badge first. Trust A is an approved, governed metric; B is derived from approved metrics; C is exploratory. See Trust tiers.
  • A dagger marks a figure the agent wrote itself. The legend † model statement — not verified by the engine appears beneath the answer, and a badge such as 2 figures could not be verified sits next to the trust badge.
  • Dotted figures can be pressed. They are figures the engine wrote or verified; pressing one opens a card that says where it came from.
  • Carried context is shown, not hidden. When a filter or period is carried over from an earlier question, a Carried context strip names it, with a Change period chip beside it.
  • Scope means less data, not less permission. Row-level security narrowed what the agent read, so your total may legitimately differ from a colleague's.
  • A stop is stated, never disguised. If an answer reaches its time or cost limit, it says so in a sentence — and where a partial value exists it is labelled Partial result.

Good to know

  • One answer is written at a time in each conversation, so every answer is built on the one before it.
  • A conversation moved to another agent carries a permanent line — From this point on, questions go to Finance Analyst. — and a change of model is marked the same way, so the history always shows who answered what.
  • Starter questions and Next analysis? chips only fill the composer. Nothing is sent until you send it.
  • The work trace shows steps and timings only; the executed query is kept behind its own permission in the evidence drawer.