01What an internal agent is, and what it is not
The difference between an agent, a chatbot and an automated workflow is in the guide on what an AI agent is. The focus here is different: the agent that works inside the business, with the team, rather than facing customers.
An internal agent receives a bounded task —prepare a reply to an email, summarise a meeting, update a record— and leaves it ready for review. It can query the tools it has been allowed to use and propose the next step. It is not another employee, nor a system that runs the business: it is a piece that saves repeated work and leaves the decision to a person.
02Examples of tasks it can prepare
| Task | What the agent prepares | What a person approves | What stays blocked |
|---|---|---|---|
| Reply to a customer email | A draft with the order details | The final text and sending | Sending without review |
| Summarise a meeting | Summary, decisions and tasks | What is shared and with whom | Sharing outside the team |
| Update the CRM | A proposed change to the record | The final change | Deleting records |
| Prepare an invoice | A draft with the lines | Issuing it | Issuing or charging |
The table is a generic example. In a real project, tasks and systems are written with the company’s own names after the diagnosis.
03The approval queue
Everything the agent prepares lands in a queue: an inbox, a channel or a CRM view where a person reviews it, corrects it if needed, and approves or discards it. It is defined up front what is always reviewed, who reviews it, and what happens if no one does so in time.
When a task falls outside the perimeter —missing data, a request beyond what was agreed, a doubtful model response— the agent does not improvise: it flags the case and leaves it for a person. That handoff is part of the design, not a failure.
04Permissions and data
The agent only accesses what its task needs, with read-only permissions when reading is enough. It is decided which data may reach the AI model, which must stay in the source system, and where each step is logged. Credentials are kept out of the instruction text. How the tools are connected is explained in the guide to connecting AI to n8n, Make and your CRM.
05How it is tested, monitored and corrected
Before it handles real work, the agent is tested with sample cases and edge cases. It then runs in parallel for a while: it prepares, but nothing leaves the queue without review. Which drafts are approved unchanged, which are corrected and which are discarded shows how to adjust the instructions or the perimeter.
If the starting point is sales follow-up, the guide to automating CRM and sales follow-up applies this pattern to leads and proposals. The AI integration service describes the full method.