# Enriching Contacts, Companies & Deals Karmaflow.ai can auto-fill custom fields on a record from your own data plus optional web search — no third-party enrichment account required. This article walks through the AI-research field workflow end to end. ## How AI-research fields work An **AI-research field** is a custom attribute you define once (per entity type) and the platform fills automatically on demand. The pipeline: 1. Pull the record's own card (name, email, company, tags, recent notes/activities). 2. Retrieve matching context from the per-tenant **vector store** (your calls, chats, emails, SMS, notes, activities), optionally from **web search**, optionally from **contact memories**. 3. Send the prompt plus the retrieved context to Gemini. 4. Persist the answer onto the record's `customFields`, along with the list of sources the model actually used. The sources are stored alongside the value so every enrichment is auditable — anyone reading the field can see exactly what context produced it. ## Defining an AI-research field 1. Open **CRM → Setup → Custom Fields** (`/settings/custom-fields`). 2. Pick the entity tab — **Contacts**, **Leads**, **Companies**, or **Deals**. 3. Click **+ New field**. 4. Set the **Friendly name** (e.g. "Industry"). The slug auto- fills from the name; override only if you need a specific `dbName`. 5. Change **Field kind** to **AI-research (auto-filled)**. An **AI research configuration** section appears below. 6. Fill in: - **Prompt** — the question for the LLM. Be specific. Examples: - *"What industry is this company in? Reply with one short phrase like 'B2B SaaS' or 'Healthcare provider'."* - *"Summarise this contact's role at their company in one sentence."* - *"List the top three product features they've discussed, comma-separated."* - **System prompt** *(optional)* — tone / format instructions that apply to every fill. Example: *"You are a sales-research assistant. Cite your sources. If you don't have enough information, reply 'unknown' instead of guessing."* - **Sources** — which retrieval channels to use: - **Vector store (own data)** — notes, activities, calls, chats, emails, SMS attached to the record. - **Web search** — public web results, summarised by the LLM. - **Contact memories** — persistent memory entries written by AI agents over time. - **Refresh interval (days)** — `0` for on-demand only; anything higher schedules an automatic refresh. - **Model** — defaults to `gemini-2.0-flash-001`. Leave this alone unless your workspace has a specific model preference. 7. Click **Save**. The field now appears in the **Custom Fields** card on every record of that entity type, with an **AI** chip next to its label and an **Enrich** button on the right. ## Enriching a record Open any contact, company, or deal. In the **Custom Fields** card on the right column: - The first time you see an AI-research field on a record, the value box says *"Not yet filled"* and the button reads **Enrich**. - Click **Enrich**. The button spins for a few seconds while the LLM runs. - The value appears in the box. A small **N src · view** link appears next to the button — click it to open a popover that shows: - When the value was generated and which model ran. - Each source the model pulled in, with its metadata (call ID, channel, timestamp, etc.). - Need to re-run? The button is now labelled **Refresh**. Click it any time to regenerate the value with the same prompt and sources. The new sources replace the old in the popover. ## Editing AI-research fields after the fact AI-research fields are read-only in the Custom Fields card — they only update via the **Refresh** button. If you want to override a value manually, the cleanest path is to: 1. Edit the field definition in **Settings → Custom Fields** and change the **Field kind** back to **Manual**. 2. Save. The existing value stays on every record. 3. Edit the value inline on each detail page as needed. 4. Optionally flip the field back to **AI-research** later — the manually-entered values remain unless someone clicks **Refresh**. ## Bulk enrichment The bulk-edit dialog on the spreadsheet view doesn't yet drive the AI pipeline — for now, enrichment happens one record at a time via the **Enrich** button. The scheduled refresh (`refreshIntervalDays`) handles bulk catch-up over time without manual clicking. ## Practical recipes A few prompts that work well in production: - **Company → Industry**: *"What industry is this company in? Reply with one short phrase."* Sources: web + vector store. - **Company → Recent news**: *"In one sentence, what's the most recent notable news about this company from the last 90 days? If none, reply 'no recent news'."* Sources: web only. - **Contact → Role summary**: *"Summarise this contact's role and seniority in one sentence based on conversations on file."* Sources: vector store + memory. - **Deal → Risk flag**: *"Based on the call transcripts and notes attached to this deal, list any concerns the buyer has expressed about pricing, timing, or alternatives. Reply 'none detected' if there's no signal."* Sources: vector store. - **Contact → Compliance flag**: *"Has this contact ever asked about HIPAA, SOC 2, GDPR, or another compliance topic? Reply with the specific term they mentioned, or 'no'."* Sources: vector store. ## When AI-research won't help (and what to use instead) - **Real-time data** (current price, today's weather, this morning's news) — refresh interval applies; the value reflects the last time the field was filled. - **Cross-record aggregates** ("how many deals does this contact have open?") — use a CRM dashboard widget instead. - **Public B2B data** (LinkedIn role, company headcount, funding round) — the LLM can fetch some of this via web search, though a dedicated data provider is usually denser. - **Hard facts you already capture** (email, phone) — leave those as manual fields. AI-research is for the *judgement calls* that humans usually make from context. ## Where it shows up next When you bring an AI-research field into a saved view as a column, list pages display the AI chip next to its column header so the user knows the value was auto-filled. The value still shows the most recent fill — no automatic refresh on page load to keep cost predictable.