What can an AI sales agent do in your CRM, and what should it not do alone?
An AI sales agent uses a language model and CRM tools to do sales work on its own: research accounts, log buying signals, prepare next steps and draft outreach. It's safe when its tools are scoped, risky actions like enrollments or merges need a human confirmation, and every change shows up in the CRM history.
What is an AI sales agent?
Anthropic draws a useful line: workflows run a model along predefined code paths, agents decide themselves which tools to use and in which order [1]. An AI sales agent is the second kind, pointed at sales work, and it needs CRM tools to be more than a chatbot that writes emails.
A lot of what gets sold as an AI sales agent is outreach automation with a model writing the text. That's fine, but I think the interesting work is earlier, in research, signals and pipeline hygiene.
Which sales tasks can an AI agent do on its own?
| Task | Alone or with approval | Why |
|---|---|---|
| Research an account before a call | Alone | Reads a lot, writes one research note |
| Log a buying signal with its source | Alone | Adds data, a rerun does not duplicate it |
| Propose next steps on stale deals | Alone | Proposals only, a person decides |
| Draft a first message and save it on the account | Alone | Nothing is sent |
| Enroll contacts in a sequence | With approval | Reaches real people |
| Merge duplicate contacts | With approval | Hard to undo |
| Move deals to Lost | With approval | Changes reporting and notifies the team |
The pattern is simple. Reading and adding is cheap to get wrong, changing and contacting people is not.
What can go wrong with AI sales agents?
OWASP's excessive agency risk names three causes: excessive functionality, excessive permissions and excessive autonomy [2]. In sales that's an agent that can send email when it only needed to draft, a key that sees everybody's personal lists, or a routine that enrolls a whole segment without anyone looking.
The MCP specification puts it plainly: there should always be a human in the loop with the ability to deny tool invocations [3]. I'd build the approval into the CRM itself, not into the prompt, because prompts get edited and forgotten.
Where should an AI sales agent run?
There are two good options. The agent runs inside the CRM, on the vendor's platform and model access, or it runs in your own Claude Code and talks to the CRM over MCP. In Sigma CRM both exist: signal observers run on the platform, billed in credits in the cloud version, or they run on your own Claude Code and report their results back, without using platform AI credits for the model. More on the agents on the agentic CRM page.
How should approvals for AI sales agents work?
Approvals should sit in the CRM tool, so they hold no matter which agent calls it. At JadenX we built Sigma CRM so that bulk changes, merges, imports and sequence enrollments ask for confirmation first, new sequences start inactive and send nothing, and enrollment skips addresses on the do-not-contact list. Every write carries its source, UI, MCP, agent or observer, so a sales lead can see on the record what an agent did.
Frequently asked questions
Can an AI sales agent replace a sales rep?
No. It takes over research, logging and preparation. Conversations, judgement and decisions stay with people.
Should an AI sales agent send emails on its own?
I wouldn't start there. Let it draft and save, and let a person review and send or activate the sequence.
What does an AI sales agent need from the CRM?
Tools for every capability it should use, scoped access, confirmations for risky actions and a history of its changes.
Can we bring our own model?
In Sigma CRM you choose the provider and model per observer and can bring your own key.
References
- Anthropic, "Building effective agents", Anthropic Engineering, 19 Dec 2024. https://www.anthropic.com/engineering/building-effective-agents
- OWASP GenAI Security Project, "LLM06:2025 Excessive Agency", OWASP, 2025. https://genai.owasp.org/llmrisk/llm062025-excessive-agency/
- Model Context Protocol, "Specification 2025-06-18: Tools", modelcontextprotocol.io, 2025. https://modelcontextprotocol.io/specification/2025-06-18/server/tools
About this article: I work with a set of AI agents that research sources and turn my notes and the Sigma CRM documentation into a draft. The product facts come from our own documentation and code. The opinions are mine.
If you're thinking about AI agents for your sales team and want to start with the safe parts, talk to me at JadenX.