Ask about dry powder, stale deals or who has not opened the Q2 report and get an answer from your live data. Ask it to change something and it shows you the diff first. Nothing mutates until you apply it, and it never touches your LPs on its own.
Reads answer instantly. Every write is a card with a diff. Nothing changes until you click.
Publishing and notifications open the real dialog, never run from chat.
Source “Agent · date” on fields, “asked by you” in the feed.
A docked chat is one tap away on every screen, with suggested prompts that change with where you are. Agent Mode turns the whole app into a full-screen conversation for longer work. Both share a single thread, so context never resets.
“How much is left in Fund II?” returns the number and a live table of your vehicles. “What’s stale in Diligence?” returns names, days idle and a pipeline strip. “Pull up Acme Robotics” returns the company card with momentum and fundraise timing. Each artifact opens the real screen with one click.
Every change renders an action card with the field-level diff, old value struck through, new value highlighted, computed from the live record rather than from the model’s memory. Apply or dismiss. Batches of changes, like adding three companies from your notes, become a checklist where you pick which to apply.
Publishing a report set or sending notifications is never executed from chat. The card opens the app’s real publish review, and the agent says so in plain words. The boundary is architectural, not a prompt.
Fields the agent writes carry the source “Agent · date” with a confidence chip, using the same provenance system as enrichment, and the manual-edit rules apply unchanged. Every applied action lands in the activity feed as “Agent … · asked by you”, and raises a notification.
The agent is a client of the same versioned API the interface uses. Same validation, same guards, same audit hook on every write. There is no side door.
An autonomous agent that edits your records while you sleep sounds efficient until it moves a deal, mis-records a round or emails an LP on the strength of a guess. In fund operations the cost of one wrong write is higher than the cost of a hundred clicks.
So the agent in Investryx does all the work up to the write, and stops. It reads your live data, computes the exact change and shows you the diff. Applying is one click; dismissing is one click. Batches are checklists. LP-facing actions open the same dialog you would use yourself.
The result is the speed of delegation with the accountability of doing it yourself, and an audit trail that says who asked for every change.
Google Vertex AI Gemini, behind one provider module so the model can change without changing the product. Your data is sent to the model for the request only and is not used for training.
No. There is no autonomous or background agent in v1. It acts only inside a conversation, and only after you apply a proposed change.
It follows the AI language set in Settings, English, Japanese or Chinese, for prose. Data values stay as stored.
Dismiss it. A dismissed card is recorded but changes nothing. If a value is wrong after applying, edit it like any other field; the provenance chip shows who wrote it.
Fourteen days free. Card required, charged only when the trial ends. Cancel anytime.