AI assistant#

The AI Assistant lets you ask your CRM questions in plain language — "How many leads came in this week?", "Show me my overdue follow-ups", "Draft a WhatsApp follow-up for this lead." Behind the scenes it does not guess: it reads live numbers from your workspace's data through controlled tools, so the answer is the same number a report would show.

There are two AI features in this area:

1. AI Assistant (this page) — you ask, it answers/does.
2. AI Follow-up Agents — autonomous agents you enroll leads into (see below).

Who this is for#

• Managers and admins — this menu is shown to manager-and-above roles.

• Sales teams — benefit from the AI-drafted emails and WhatsApp messages on Lead 360.

• Difficulty: Medium

• Time: 10 minutes

• Prerequisites: The AI Assistant menu is visible to managers and above. Using a real LLM needs an AI provider key (see "Setup" below) — without one, the assistant degrades honestly instead of pretending.

Steps#

1. Ask a question#

1. Open AI Assistant in the sidebar (/admin/ai-assistant).
2. Type a question, for example:
   • "How many leads do we have and how many are new?"
   • "Which leads haven't been contacted yet?"
   • "What are my overdue follow-ups this week?"
   • "How is the team performing this month?"
   • "What did Meta ads produce last month?"

Expected result: an answer grounded in your data. Numbers come from live queries against your workspace — never a guess.

2. Ask it to do something#

The assistant can also act, within your permissions:

• Create a follow-up — "Create a follow-up for Ramesh tomorrow at 10am to discuss pricing."
• Draft a WhatsApp message — it writes the draft; you approve and send it from the inbox.

Expected result: the follow-up appears in Follow-ups with the details you gave; drafts land ready to review, never sent behind your back.

3. Know what it can do today#

The assistant has a fixed set of skills (tools):

Anything outside this list — the assistant says so rather than inventing an answer.

AI Follow-up Agents (autonomous mode)#

An AI SDR agent is a configurable persona that follows up enrolled leads step by step: it decides what to do next for each enrolled lead and does one thing per cycle (for example, send a follow-up email), within hard guardrails.

4. Create an agent#

1. Open AI Follow-up Agents (same group in the sidebar, /admin/ai-sdr-agents).
2. Click New agent and set:
   • Goal — Qualify (find out if they fit) or Book (get a meeting).
   • Channel — email, SMS or WhatsApp.
   • Persona and offer context — how it should sound and what it may offer.
   • Guardrails: maximum touches (1–20, default 4), minimum hours between touches (default 24), and whether it must respect business hours.
   • Active — off until you switch it on.
3. Save.

Expected result: the agent appears in the list, inactive. Nothing happens until you enroll leads.

5. Enroll a lead#

1. Open a lead's Lead 360 view and use Enroll AI SDR (or bulk-enroll from All Leads).

Expected result: the enrollment is created. Every 15 minutes a worker lets the agent take one step per enrolled lead — a message, a wait, or a stop — always inside its guardrails. Leads who opted out or that the agent's rules say are unreachable are skipped, and enrolling twice is a no-op.

6. Setup: connect an AI provider#

1. Go to Settings → AI.
2. Choose the provider (OpenAI or DeepSeek), paste the API key, and pick the model (default: DeepSeek deepseek-v4-flash). Keys are stored encrypted and shown masked.

Expected result: the assistant upgrades from deterministic summaries to a conversational model. Note that even with a model connected, every number still comes from live tool queries — the model never invents figures.

Expected result#

You can interrogate your pipeline in plain language and get real numbers; and AI follow-up agents can take over repetitive touch sequences without ever sending something a human cannot audit (all activity is visible on the lead's timeline).

Common mistakes#

• Expecting the LLM to compute numbers. It does not; tools query the database. If a number looks wrong, check the underlying report/data, not the prose.
• Turning an agent on with no guardrails. Start with the defaults (4 touches / 24 hours); raise limits deliberately.
• Enrolling opted-out leads. The system blocks it by design — respect the signal.
• Forgetting billing. AI calls consume your provider credits; usage is visible in the platform's AI settings for admins.

Troubleshooting#

Technical details#

• Assistant page: App\Filament\Pages\AiAssistant (slug ai-assistant, manager-or-above). Services: App\Services\Ai\Assistant\TenantAssistantService and TenantAssistantTools (the tools listed above). Provider resolution via AiChatClient::resolveProvider — tenant settings.ai.* → platform AiSettings → environment.
• Agents: App\Filament\Resources\AiSdrAgentResource (plan feature ai_sdr); model enrollment via AiSdrEnrollment; processing ProcessAiSdrEnrollments runs every 15 minutes (one decision/action per enrollment per tick). Guards: ai_sdr_opted_out_at, AiSdrAgent::canReach, idempotent enrollment.
• Email drafting on leads: AiEmailComposerService; enrichment uses the EnrichLead job (needs the lead's email).

Related guides#

• Lead 360 view — where drafts, enrichment and enrollment live.
• Follow-ups & tasks — where created tasks appear.
• Reports — verify any number the assistant gives you.
• Meta overview — the ads data the assistant can read.