How do you build an AI sales agent?

Key facts

Core tools	CRM read/write, enrichment lookup, email draft, calendar, knowledge base
Default mode	Draft and queue for human approval; never auto-send cold outreach
Compliance	Consent checks, opt-out suppression lists and regional outreach rules
Grounding	Product claims come from an approved knowledge base, not model memory
Data handling	Scoped keys and audit logs per integration and per user
Models	30+ models on one OpenAI-compatible API for routing research and drafting
Human gate	Irreversible actions — sending, pricing, commitments — require approval
Roadmap	Assistants, files and batch endpoints are coming soon

TL;DR

• The agent drafts and prepares; a human approves and sends.
• Ground every product claim in an approved knowledge base.
• Enforce consent, opt-out suppression and regional outreach rules in code.
• Write to the CRM through validated schemas so records stay clean.
• Measure reply quality and meetings booked, not messages sent.

How it works, step by step

1. Map the workflow: account research, enrichment, qualification, drafting and follow-up.
2. List the tools with least privilege: CRM fields, enrichment source, email draft, calendar.
3. Build a knowledge base of approved product claims, pricing rules and objection handling.
4. Add approval gates for anything irreversible and suppression checks before any send.
5. Route research and enrichment to fast models and drafting to a stronger model.
6. Log every draft, approval, send and outcome with the account and user attached.
7. Review weekly: edit rates, reply quality and CRM field accuracy drive the next iteration.

Try it yourself

Open the AI agent builder →

What a sales agent should and should not do

The useful split is simple: the agent does preparation and drafting, humans do judgement and commitment. Research an account from public data and your own notes. Enrich the CRM record with validated fields. Score fit against your ideal customer profile. Draft a short, specific email or a call brief. Queue everything for review.

What it should not do: send cold outreach autonomously, invent product claims, promise pricing, or update fields it cannot verify. Every one of those actions either creates legal exposure or damages data quality, and both are expensive to unwind.

Compliance rules belong in code

Outreach is regulated, and the rules differ by region. Build the constraints into the tool layer so the model cannot bypass them: suppression lists checked before any draft is queued, consent status stored on the contact record, unsubscribe honoured immediately, and sending limits per domain. Keep a per-message audit trail linking account, template, approving user and timestamp.

• Consent state is data your system checks, not a sentence in the prompt.
• Suppression runs as a hard filter on every outbound action.
• Pricing and commitments are never generated by the model for external use.
• Auditability answers who approved what, when and on which data.

Measuring a sales agent honestly

Message volume is the wrong metric; it rewards spam. Track edit distance on drafts, reply quality, meetings booked, pipeline created and CRM field accuracy. A useful baseline is how much time a representative saves per account and whether the human edits get smaller over time.

Plugsky supplies the model layer rather than a sales product: 30+ models on one OpenAI-compatible key, live function calling and streaming for tool-driven workflows, embeddings and RAG for knowledge grounding, plus scoped keys, RBAC, SSO/SCIM and audit logs for controlled access. It does not include enrichment data, email sending or CRM connectors, so those remain your integrations. Self-serve plans are flat monthly; current plans and the free plugsky-micro and plugsky-lite models are on the live pricing page.

Honest comparison

Activity	Manual SDR	Prompt-only automation	Sales agent with gates
Account research	Slow, inconsistent	Generic summary	Structured brief with sources
CRM updates	Manual and patchy	Free-form text dumps	Schema-validated fields
Outreach	Personalised	Template blast	Drafted, human-approved
Compliance	Process and training	Prompt instructions	Hard filters in the tool layer
Measurement	Pipeline outcomes	Volume	Edit rate, replies and meetings

Frequently asked questions

Should a sales agent send email by itself?

Not for cold outreach. Draft and queue is the safe default; a person approves every external message. Transactional replies can be automated only where consent and content rules are already satisfied.

How do I stop it making up product claims?

Ground drafting in a retrieval step over an approved knowledge base, and validate that generated claims map to a retrieved passage before the draft enters the review queue.

Can it update the CRM directly?

Read access is safe. Writes should go through narrow, schema-validated tools, one field group per tool, with an audit log and an easy rollback path.

What about GDPR and outreach law?

You stay responsible as the sender. Enforce lawful basis, consent and opt-out handling in your own systems, and keep regional rules per market rather than one global policy.

Which model should do what?

Fast models handle enrichment extraction, classification and summarisation. A stronger model drafts complex, high-value outreach. Route per task on one API key.

How do I know it is improving?

Track human edit distance on drafts, reply and meeting rates, and CRM field accuracy. If edits do not shrink over time, the knowledge base or prompts need work.

Does Plugsky include enrichment or email tools?

No. Plugsky provides models and the API layer with function calling, embeddings and RAG. Enrichment providers, CRM connectors and email sending are integrations you choose and operate.