What is an agentic CRM? A definition and four rules

An agentic CRM is a customer relationship management system with AI agents inside it. The agents read context, make a judgement and hand you the result: a researched lead, a drafted email, a reading of a reply. A classic CRM only stores what people type into it. An agentic one does part of the reading, thinking and writing itself.

That definition is easy. The hard part is deciding what the agents are allowed to do with what they produce. This article covers both.

What does an AI agent in a CRM actually do?

An AI agent is a loop. Your software sends a language model a task and a list of tools it may use. The model either answers or asks for a tool, your code runs the tool, sends back the result, and the loop repeats until the task is done.

In a CRM, the tasks are the slow parts of selling:

• Researching a new lead. Who is this company, how big is it, what are they doing right now, and is it a fit?
• Writing the follow-up. An email that opens the next step, based on that research.
• Reading a reply. Is this interest, an objection, an out-of-office or an opt-out?
• Noticing what went quiet. Which open deals have had no email, call or meeting for two weeks?
• Drafting a proposal from everything recorded on a deal.

In Tolva, each of these is its own agent, six in total, and each one ends by putting a proposal in front of a person.

How is an agentic CRM different from CRM automation?

Automation follows rules you write in advance: when a contact fills in this form, send that email. It never surprises you, and it never handles a case you did not foresee.

An agent decides its own path. It chooses which searches to run, what to write, how to read a reply it has never seen. That is what makes it useful, and it is also what makes it risky. You cannot review the path in advance, so you have to bound what the path can reach.

Where do agentic CRMs go wrong?

Three failures show up again and again:

1. Sending at scale without a human. One bad prompt reaches 400 people, and an email cannot be recalled.
2. Overwriting records with a guess. An agent believes a company has 200 employees and writes it into the record. Six months later nobody remembers it was a guess.
3. Inventing records. A deal that exists, with a value in your forecast, because a model inferred one from three sentences.

Each of these looks like the feature working. That is why they are dangerous.

What rules make an agentic CRM safe?

Tolva is built on four rules. An automated check tests them against the source code, so a change that breaks one fails that check.

1. No agent has a tool that contacts a person. Not disabled, not behind a permission: the tool does not exist. An agent cannot send an email or write a campaign. An approved draft becomes a draft campaign, and sending it is a separate human act.
2. A person approves every proposal before it changes data. Research, drafts, re-scores, reply readings and deal plans all wait on a review page. Approval is one database transaction, so a half-applied proposal cannot exist.
3. Every run is recorded, including failures. The ledger row opens before the model is called and records tokens, the number of web searches that actually returned results, the duration and the cost. A run killed halfway leaves a visible row, not an invisible bill.
4. The spend cap is checked before every run, not after. It counts failed runs too, because a failed call is still billed. When the monthly cap is reached, the agents stop and say so.

Two further habits have paid for themselves:

• An agent can raise, never blank. It can raise a score but not lower one, and it cannot mark a contact as unqualified, because that is a human judgement. It fills empty company fields and never overwrites what someone typed.
• Count what happened, not what was offered. Offering an agent web search is not the same as a search returning results. A brief written from zero retrieved pages is shown differently from one backed by four.

What happens when a prospect replies?

This is the moment most automated outreach gets wrong. Email services report opens, clicks, bounces and unsubscribes, but they do not report replies. A reply is just a new email. If the CRM never receives it, the sequence keeps mailing someone who has already answered.

In Tolva, the reply is received by the CRM and handled in two halves:

• The urgent half has no AI in it. In one transaction, every active sequence for that person stops, an explicit opt-out is honoured, and the reply is filed on their timeline. An out-of-office message is recognised from its email headers and does not stop anything.
• The judgement half waits for you. The reply-reading agent explains what the reply means and drafts an answer, and you decide whether to use it, and whether the reply deserves a deal.

The reply is also forwarded to your normal inbox, with the prospect as Reply-To, so you can answer directly as you always have.

What does running the agents cost?

Real numbers, measured on Tolva’s own runs:

• Lead research: USD 0.057 per contact, because it runs up to four web searches.
• Follow-up drafts, re-scores, reply readings, deal rescues: about USD 0.01 each, because they work from what is already in the CRM.
• Proposal drafts: about USD 0.06.

A hundred new contacts, researched and drafted, come to about USD 7. The point is not that the numbers are small. The point is that you can see them, per run, and that a cap stops the agents before the budget is gone.

Is an agentic CRM right for a small team?

It suits a small B2B team best when three things are true: you have more leads than hours to research them, your deals start with a conversation and end with a proposal, and you want to keep every customer-facing decision yourself.

The daily routine in Tolva is about five minutes: answer the replies waiting, then approve or reject the new briefs. Once a week, re-score whoever engaged and look at the deals that went quiet. Once a month, check the spend and read the rejected pile. If it is mostly “not a fit”, the agents are right and the lead source is the problem. If it is mostly “wrong”, the prompts need work.

The short version

An agentic CRM does the reading, the thinking and the writing. Whether it is safe depends on one question: what can the agents do without you? In Tolva the answer is nothing that reaches a customer, a deal or an inbox.

If you want to see that on a sample pipeline, request a demo.