Running a Five-Person Agency With Twenty Agents: A Day in a Governed CRM
This is an illustrative scenario, not a customer story. The agency, names and timings are invented to show how the product works today.
The setup
Five people: two founders, two developers, one account manager. Twenty agents, each doing one narrow job: account research, meeting prep, call summaries, pipeline checks, renewal watching. The agents are the agency's own; AI PRO CRM does not run them, it gives each an identity and controlled access.
Agent capacity is set per plan (Free 1, Starter 3, Team 10, Business 30, with extra agents purchasable), so this agency is on Business. There are no credits or usage meters.
08:30 — reading the overnight notes
The account manager opens their companies. Overnight, agents have read the CRM through the MCP endpoint and left internal notes: a summary of last week's meetings, a flag that an opportunity has not moved stage in a month, a list of contacts with no recent activity.
Each note shows the agent as author. The agents changed nothing else, because they cannot: the only agent write is an internal note on an existing company. The account manager moves the stalled opportunity, completes two tasks and logs a call. The agents did the reading; the person made the changes.
10:00 — a new agent
A developer adds a renewal-watching agent with a name, handle, purpose and themselves as owner, and picks the Assistant preset so it can leave notes. They issue a 90-day credential. The token is shown once and stored only as a hash. No shared keys, no borrowed human login.
12:00 — something odd
After a prompt change, a pipeline agent calls the same search tool over and over. The call ledger shows a run of rate-limited calls against its credential, with tool, status and duration but never arguments or results. Its owner pauses the agent, which stops its credentials at once, fixes the prompt and reactivates it.
Even a manipulated agent could only write notes here, up to its daily write cap.
14:00 — a client question
"Who changed our primary contact?" The audit trail shows the account manager edited the person record on Tuesday. The call summary that prompted it sits on the company as a note from the call-summary agent. One clear answer.
16:00 — housekeeping
The founders rotate credentials near expiry, disable an unused agent and scan the ledger for errors.
What the humans actually did
They read the notes, made the changes that mattered and kept the agents tidy. There is no approval inbox and no agent write beyond notes, which keeps the model simple: agents advise, people act.
FAQ
Is twenty agents realistic for five people?
Yes, when each agent does one narrow thing. Narrow agents are easier to name, own and review.
Can the agents update deals or create tasks?
Not today. Agents read the CRM through MCP and can add internal notes to companies. Stage moves, tasks and field changes are made by people.