> Ask your database anything in plain English. Kepler searches every candidate, CV and conversation, then makes the update — with your approval.

_Kepler product feature — AI CRM Assistant for Recruitment Agencies. Canonical: https://keplercrm.com/features/ai-assistant/_

# Ask your database a question. Get the shortlist — or the change — done.

It doesn't just find answers. It files the note, moves the stage and sets the reminder, as cards you approve.

## How it works.

### It finds what was never typed in

Keyword search dies at the name field. Kepler reads the unstructured stuff — notes, CVs, call transcripts, emails — with hybrid retrieval, so the words people actually said find the people they were about.

**Searches the words people actually used** — "Led a team through a migration" surfaces the CV that says exactly that — and the ones that say it differently in a note or a transcript.

**Counts before it lists** — Ask for a number, get a number — then the breakdown behind it. The list only happens when you want one.

**Every answer shows its sources** — Each result carries where it came from — the email, the call, the CV — so you can defend the shortlist to your client.

### It does the admin, as proposals

When the answer is a change, the assistant doesn't apply it. Every edit comes back as a card — sourced, reversible, applied only when you confirm.

**Changes arrive as cards** — Move a stage, set a field, create a task — each proposed on its own card, applied only when you confirm.

**@-mention any record** — Pull a person, job or company straight into the question. The mention resolves to the record — not to a guess.

**Written where it belongs** — Confirmed changes land as you, in the record's own history — not in a side channel nobody audits.

### Big asks, handled end to end

Real asks aren't one command — they're a morning's work. Long runs stay inspectable, results become reusable lists, and every answer inherits exactly what you're allowed to see.

**Runs you can pause** — Big asks — "re-score everyone against this spec" — progress step by step in the background, and stop the moment you want a look.

**Saves answers as lists** — Any result set becomes a reusable list in one click, so the work compounds instead of evaporating.

**Inherits your permissions** — It sees what you see and writes as you — never around you. Confidential records stay confidential.

## It reads what you can read. It writes nothing without you.

The assistant lives inside the permissions you already have. That's the whole trick — and the whole boundary.

- What it reads: the records, notes, CVs, emails and call transcripts you can already open. Nothing else, nothing further.

- What it never does: write without a card. Every change is a proposal you confirm — nothing lands on its own.

- What it never does: train on your data. Your database answers your questions. It doesn't become anyone else's.

## Questions, answered

### What can the AI assistant actually do in my CRM?

Two things. It answers questions about your database in plain English — across notes, CVs, call transcripts and emails, not just profile fields — and it does the admin those answers imply: moving a stage, setting a field, filing a note, creating a task. Answers cite their sources and changes arrive as cards you confirm.

### Does it search CVs and call transcripts, or only the fields someone filled in?

Everything you can already open. Kepler uses hybrid retrieval over the unstructured material, so a phrase like "led a team through a migration" surfaces the CV that says exactly that — and the ones that say it differently in a note or a transcript.

### Can it change records without me?

No. Every edit comes back as a card — sourced, reversible, and applied only when you confirm it. Confirmed changes are written as you, in the record’s own history, not in a side channel nobody audits.

### Can it see records I’m not allowed to see?

No. The assistant inherits your permissions exactly: it reads what you can read and writes as you, never around you. Confidential records stay confidential, and your data is never used to train AI models.