CRM Solid AI Agents is the autonomous reply engine that runs on your live Telegram, X, email, and social inbox channels. When a new DM lands, the agent reads it, consults its knowledge base, applies your rules, and either sends a reply automatically, queues one for your approval, or saves a draft - depending on the reply mode you chose.
This guide is about configuring and deploying that engine. If you want to go deep on writing a persona prompt itself - the 200-word system-prompt structure, the voice rules, the example responses - read the AI sales bot training guide first, then come back here to deploy it as a live agent.
Step 1: Create an agent and write the persona
Budget: 5 minutes. In CRM Solid, go to AI Agents > New Agent. You will see three starter templates:
- Sales SDR - pipeline-focused outreach and demo booking.
- Customer Support - grounded answers with human escalation.
- Lead Qualifier - two-question scoring and auto-tagging.
Pick the template closest to your use case and customize the persona field. The persona is free text in the agent editor - it sets identity, goal, and behavioral rules. A working example for a Sales SDR agent:
You are a sharp B2B sales rep for CRM Solid. Be concise,
confident, and helpful. Always move toward a demo or signup.
Never quote a discount not in the knowledge base.
If you do not know the answer, say so and trigger handoff.For deeper persona-writing guidance - how to match tone, handle objections, and test voice consistency - see the AI sales bot training guide. That guide covers the full 200-word persona template; use it as the source for the persona text you paste into the agent editor here.
Also set per-agent model and tuning:
- Model: GPT-4o, GPT-4o mini, Claude, Gemini, or a self-hosted endpoint. GPT-4o is the default for Sales SDR work; GPT-4o mini is a cost-effective option for high-volume support routing.
- Temperature: 0.3-0.5 for support (consistent, low-variability answers); 0.5-0.7 for sales (natural-sounding variation).
- Max tokens: 400-600 keeps replies concise and platform-appropriate. DMs are not essays.
Step 2: Ground the agent with a knowledge base
Budget: 5 minutes. The knowledge base is the single most effective way to prevent the agent from inventing facts. Items are appended to the agent context in the order you add them.
In the agent editor, open the Knowledge base tab. Add items in priority order:
1. Pricing facts
"Pro plan: $99/month, up to 20 seats, 14-day free trial,
no credit card required."
2. Demo / signup link
"Demo link: app.crmsolid.com/demo
Signup link: app.crmsolid.com/register"
3. Top FAQ answers (one per item)
"Annual billing: 20% discount vs monthly."
"Cancel anytime; no contracts or exit fees."
"Data is stored in EU-West-1 (Ireland)."
4. Key objection responses
"Compared to Intercom: CRM Solid unifies Telegram, X, email,
and live chat in one inbox. Intercom does not do DM channels."
5. Escalation note (last item)
"If the question is outside the above, say 'I will get a
human on this' and trigger handoff."Keep each item short and factual. The agent reads them in order, so put the most-frequently-needed facts at the top. A 5-10 item knowledge base for a focused agent outperforms a 100-item one because retrieval stays sharp.
Accuracy rule: never put a claim in the knowledge base that you would not put on your pricing page. The agent quotes the knowledge base verbatim; any inaccuracy reaches real customers.
Step 3: Connect the agent to live channels
Budget: 2 minutes. Under the Channels section of the agent editor, select the platforms this agent covers:
- Telegram: linked Telegram accounts are listed by username. Pick the account(s) this agent should watch.
- X (Twitter): same - pick the @handles this agent covers for incoming DMs.
- Email: pick the inbox address (e.g., [email protected]) you want the agent to reply from.
- Social Inbox: the unified inbox catches messages from any connected social channel; the agent can cover this as a catch-all.
Scope each agent narrowly. A Sales SDR agent on Telegram and X performs better than one agent trying to cover all four channels with a single persona. Channel-scoping also makes the Activity Log much easier to review.
Also set the Trigger mode - when the agent fires on each channel:
- All messages: agent replies to every incoming DM on its channels. Full hands-off mode.
- Only keyword: agent fires only when the message contains a keyword you configure. Good for triage agents.
- Only first contact: agent replies only to the very first message from a new contact, then steps back. Good for welcome / qualification flows.
Step 4: Write rules and guardrails
Budget: 10 minutes. Rules run on every incoming message before the agent replies. Each rule is an IF-condition THEN-action pair. Open the Rules Engine tab in the agent editor:
IF message contains "price"
THEN reply with pricing (sends a canned pricing reply)
IF matches /refund|cancel/i
THEN handoff to human (stops agent, routes to inbox)
IF contact stage = New
THEN set stage: Qualifying (moves the CRM stage)
IF platform = Telegram
THEN add tag: telegram-lead (tags the contact)
IF message contains "demo"
THEN assign sequence: Onboarding (starts an outreach sequence)
IF matches /unsubscribe/i
THEN block (stops all messages to this contact)Rules run top to bottom on every message. Order matters: put your handoff rules above your reply rules so an escalation trigger is never overridden by a canned reply.
Guardrails belong in the persona text, not the rules engine. Add explicit "never" statements for the topics most likely to cause harm:
# Agent guardrails (add to persona instruction)
- Never quote a specific price not in the knowledge base.
- Never promise a delivery date, SLA, or uptime figure.
- Never give legal, medical, or tax advice. Say "I am not the
right resource for that" and trigger handoff.
- Never acknowledge competitors negatively. Point to the
comparison page instead.
- Never use absolute language: "always", "guaranteed",
"every customer". Use "typically" or "most teams".Test each guardrail deliberately: send the agent a message designed to trigger each restriction and confirm the response follows the rule. If it does not, rewrite the guardrail text and retest.
Step 5: Set rate limits and quiet hours
Budget: 2 minutes. Open the Guardrails panel (separate from the rules engine). The six settings that govern platform-safe sending:
- Rate limit: default 20 replies per hour, 80 per day. This is the safe ceiling for Telegram and X without risking account flags. Raise it only if you have confirmed your account tier supports higher volume.
- Reply delay: default 8-25 seconds. The agent waits a random interval in this range before sending. This pacing mirrors natural human typing time and avoids the instant-reply pattern that platforms flag.
- Quiet hours: default 22:00-07:00 UTC. The agent holds any replies that would go out in this window until quiet hours end. Adjust to your audience timezone.
- Stop keywords: "stop", "unsubscribe". When a contact sends either word, the agent stops all replies to that contact permanently. Never remove this; it is required for compliance on most messaging platforms.
- Handoff keywords: "human", "agent", "refund". When any of these appear in an incoming message, the agent immediately stops and routes to the unified inbox.
- Auto-handoff on uncertainty: when the model confidence is low - the agent is unsure how to reply - it escalates instead of guessing. Enable this; the cost of a missed escalation is higher than the cost of one extra handoff.
Step 6: Wire the human-handoff so context transfers cleanly
Handoff quality is what separates a frustrating experience from a good one. When the agent escalates, the human who picks up the thread should be able to continue without asking the contact to repeat themselves.
CRM Solid auto-generates a one-sentence summary of the conversation when a handoff fires. The thread appears in Unified Inbox with the summary pinned at the top and the full conversation visible below.
To configure the handoff experience:
- In the agent editor, open Handoff settings.
- Set Handoff keywords: the words that trigger immediate escalation (human, agent, person, refund, cancel, and any domain-specific terms).
- Set Handoff message: the reply the agent sends to the contact at the moment of escalation, e.g., "Let me get a team member on this - they will be with you shortly."
- Set Assign to: a specific team member, a team queue, or unassigned. Unassigned is fine if you check the inbox frequently.
- Enable Auto-handoff on uncertainty in the Guardrails panel (covered in Step 5).
Tip: review your handoff message quarterly. The message a contact gets the moment the agent steps back is often the one they remember most. Keep it warm and specific - "Let me get a team member on this" beats a generic "You have been transferred to support."
Step 7: Test in Suggest mode before going live on Auto-send
Budget: 7-14 days. The three reply modes are:
- Auto-send (hands off): the agent writes and sends the reply on its own, the moment a message lands.
- Suggest (you approve): a ready-to-send reply waits in your inbox. Send it with one click, or discard it.
- Draft (quietly saved): the agent saves a draft silently. Nothing is sent until you open and manually send it.
Start with Suggest mode for the first 7-14 days on every new agent. Open AI Agents > Activity Log daily and review every suggested reply. For each one, ask:
- Does it sound like the persona you wrote?
- Are the facts correct (no invented pricing, no fake features)?
- Would you be comfortable if this was sent to a real customer?
- Did it escalate correctly when it should have?
When at least 90% of suggested replies pass this review across two full weeks without a single factual error, switch to Auto-send. If quality is borderline, stay in Suggest mode and fix the persona or knowledge base first.
For agents that handle pricing questions or deal-stage conversations, consider staying in Suggest mode permanently. The approval click adds 30 seconds per reply but eliminates the risk of an auto-sent mistake on a high-stakes thread.
Step 8: Monitor weekly metrics and refine monthly
Budget: 20 minutes per week. Open AI Agents > Analytics each week and check four numbers:
- Handoff rate: target 20-40% for a well-tuned agent. Below 10% and the agent may be over-handling (not escalating when it should). Above 60% and the knowledge base or persona is too narrow.
- Reply quality score: CRM Solid surfaces a composite quality signal based on conversation outcome, CSAT (when collected), and handoff patterns. Target 4.0 or above.
- Conversation resolution rate: the share of agent-handled conversations that reach a clear outcome (lead qualified, demo booked, question answered, contact opted out). Target 65% or above.
- Downstream conversion: contacts the agent handled that became qualified leads or booked demos. Compare to human-handled conversations to see whether the agent is helping or hurting the pipeline.
Once a month, open the Activity Log, read 10 random conversation threads the agent handled, and make one improvement to the persona or knowledge base. One change per month, compounded, builds an agent that is measurably better every quarter.