Why CRM is a good place for a first agent
A CRM already contains deals, interaction history, tasks, and response templates. That makes it convenient for a pilot: the agent does not need access to every internal data source from day one. It can receive one deal record, a short interaction history, and the rule for the next step, then prepare an email draft or a task for the account manager.
But the word “agent” can quietly turn an assistant into a user with nearly the same rights as an employee. Today it suggests wording; tomorrow it may send the email itself, change a deal stage, and apply a discount. An integration does not become safe just because the initiative came from a model. The application must constrain actions independently of how convincingly the agent explains its choice.
A practical first objective is not “automate sales.” It is to reduce time spent preparing a routine next step: gather facts from the record, compare them with an approved playbook, and show the employee a short, verifiable draft.
The flow: read, propose, validate, act
Make the process explicit, even if the user sees only one button:
1. **Fetch context.** The server verifies that the current user can access the deal and gives the agent only the needed fields: company, product, stage, last agreed action, and relevant correspondence.
2. **Prepare a proposal.** The model returns a structured result: proposed next-step type, message draft, facts from the record, and questions that must be answered before sending.
3. **Check business rules.** Ordinary application code—not the language model—checks required fields, allowed customer segments, prohibited promises, discount limits, and valid deal-stage transitions.
4. **Show it to a person.** The account manager sees the source facts beside the draft, edits the text, and explicitly approves the action.
5. **Execute one bounded command.** The backend checks authorization again and sends the CRM one specific operation for one specific record.
The boundary matters: a model can propose calling a tool, but the application decides whether to run it. A function call is a request to your code, not evidence that the action is permitted.
For an initial pilot, complex RAG may not be necessary. If the record contains current information, use it directly. If rules live in policy documents and price lists, retrieve only from an approved collection, return references to the matched passages, and show them to the manager. A customer email or retrieved web page is input data, not a system instruction that can change the agent’s permissions.
A permission matrix, not a promise of “human in the loop”
“An employee will check everything” does not describe a control. Decide in advance which actions the agent may perform and at what level:
| Action | At launch | After a successful pilot |
|---|---|---|
| Read permitted fields from one deal | Yes, as the current user | Yes |
| Prepare an email or note draft | Yes | Yes |
| Create a task of a fixed type and due date | After approval | Possibly automatic if reversible and tested |
| Send a customer message | Explicit approval for every send | A narrow template and segment, with limits |
| Change a discount, owner, or deal stage | No | Separate policy and approval |
| Delete records, export data in bulk, or run arbitrary queries | No | Do not expose as agent tools |
Approval must be tied to the content of the action, not to general consent to start the workflow. Before the send button, show the recipient, subject, final text, and related deal fields. If the text or recipient changes after approval, ask for approval again.
How to design the tools
Avoid a universal function that can update arbitrary fields or run arbitrary queries. Give the model a few narrow operations with typed parameters: get a deal summary, prepare a follow-up, and create a task after approval. The server should validate each operation separately.
For every change, useful checks include:
- the user still has access to the record;
- the customer and deal IDs match the open record;
- the email recipient came from the CRM, not from a model guess;
- changes fit the allowed fields and enumerated values;
- a repeated call will not create a duplicate task or email;
- the operation is audited with user, time, object, changed fields, and approval status.
Validate model arguments against a schema, but do not confuse valid JSON with a valid business action. A schema can ensure a discount is a number; only your rule can limit it to an approved range. Authorization must be checked on every call by the backend or CRM, not inferred from prompt instructions.
Where a local model helps—and where it does not
Local deployment may fit when deal and correspondence data cannot be sent to an external service, suitable infrastructure exists, and a team can update the model, monitor quality, and keep the integration operating. But running inside your perimeter does not make an agent safe by itself: excessive permissions, prompt injection, and tool-handling errors remain.
An external API can be simpler for a limited test with anonymized or synthetic data. For a production environment, the choice depends on data requirements, contractual terms, latency, cost, available expertise, and the model’s Russian-language quality. Compare the same scenario, including integration, evaluation, support, and human approval costs.
In either deployment, separate the layers: the model proposes; the orchestrator manages stages; the backend authorizes; and the CRM executes only permitted operations. If all of this is bundled into one long system prompt, control relies on the model always interpreting the boundaries correctly. That is a bit like leaving the warehouse keys with a note saying “please be careful.”
A two-week pilot without permission to send email
Start with one team and one request type, such as preparing a follow-up after a product demo. For the first days, the agent only reads selected fields and drafts a message; the employee sends it through the usual process. Compare against the current workflow, not against a perfect answer in a product demonstration.
Before the pilot, define 30–50 anonymized examples or scenarios with an expected next action. Measure separately:
- whether facts were extracted correctly from the CRM;
- whether the draft matches the business tone and approved commitments;
- whether required clarifications were missed;
- how many drafts managers accept unchanged, substantially rewrite, or reject;
- how many minutes are saved after reading, checking, and editing;
- whether the agent attempted a forbidden action when given mistaken or provocative email content.
After that, enable one reversible operation, such as creating a task after approval. Do not expand permissions at the same time as the number of scenarios: otherwise it will be hard to identify which new capability caused an incident.
Economics and a decision to continue
Do not count only token costs. A pilot includes CRM connector setup, field cleanup, identifier protection, model evaluation, an approval interface, action auditing, monitoring, a fallback process, and employee training. Human review adds time, but may cost less than correcting a mistake and restoring customer trust.
A useful model metric is net time per request:
Time saved = old-process time − context preparation − draft review − edits − error handling.
Multiply median time saved by the actual number of eligible monthly requests and the cost of an employee hour. This is not a forecast; it is a rough boundary for deciding whether the pilot deserves to continue. Separately count implementation and support costs, as well as the cost of a wrong message or a missed inquiry.
Continue only if quality is acceptable on the test set, employees trust the process, the audit confirms correct permissions, and savings remain after review. If the model writes well but the manager spends the same amount of time rechecking everything, it is too early to automate sending.
What a manager can do this week
Choose one recurring CRM operation and describe it in five lines: which fields to read, what to prepare, which rules to check, who approves, and what the agent must never do. Start in “draft only” mode, with no writes or sending. Measure time and keep an explicit approval log.
This tests usefulness before granting rights to take external actions. The agent can gradually receive exactly the operation that proved useful—like a new employee, except it is better not to leave its permission list in the welcome email.
