Is AI in your CRM the same as an agent connected to it?
No. Many CRMs now include their own AI features, such as summaries or suggested replies. Those features work inside that one system. An agent connected through an integration can carry out one task across several systems. It can look up a customer in your CRM, prepare a quote in a document, and update the status in your project tool. Each system gets its own permissions, so you decide per action what the agent may read or change. If the AI built into your CRM already handles the task, you may not need an integration at all.
How does an AI integration with your CRM work?
An integration usually runs through your software’s API: the gateway programs use to request or change data. You create a separate key for the agent, apart from your own account. For that key, you choose which data the agent may read and what it may change. The agent then sees only what the task requires, and you can revoke access without changing your own login.
Hermes, the agent Voltti works with, uses that API when a task calls for it. For example, when a quote request comes in, the agent looks up the customer in your CRM and reads earlier agreements. That is read access. Whether the agent may then update a note or status is a separate decision. To see how that quote process works step by step, read automating quotes with an AI agent.
Some integrations run through MCP, the Model Context Protocol. Its official documentation describes it as "an open-source standard for connecting AI applications to external systems." Hermes supports MCP, and our own agent works with Storyblok and Search Console through such connections, each with its own access. Whatever the route, give the agent its own key for each system. Revoke or replace that key when the pilot ends.
Which integrations should you avoid?
Do not connect everything you can. Avoid integrations with more permissions than the task needs, systems you rarely use, and connections without logging or oversight. If you are unsure, ask one question: does the benefit outweigh the added risk?
That risk has a name. OWASP, an independent software security organization, calls it excessive agency in its list of risks for language model applications. An AI system with too much functionality, too many permissions, or too much autonomy can cause damage when it misunderstands an instruction or is manipulated. OWASP recommends keeping functionality and permissions to the minimum necessary. For high-impact actions, a person approves first.
How do you limit permissions in your CMS and CRM?
Give each integration only the permissions it needs. Reading, editing, and publishing are three separate permissions. A CRM connection that only looks up contact information does not get write access. For a CMS, agree who may edit and publish, and which changes need approval in advance. A price change or a legal text usually calls for more oversight than a typo in a blog post.
This is how we work ourselves. Our own Hermes installation at Voltti edits copy directly in Storyblok, the CMS behind this website, and sometimes publishes it as well. Every change starts as a ticket and ends with a verification report in that ticket. We review the work afterward and correct it where needed. We connect Search Console as read-only: the agent has nothing to change there, so it gets no write access. To see how an agent makes targeted improvements to a website, read SEO with an AI agent.
Checklist: six questions before you turn on an integration
Go through these questions for every new integration. If you cannot answer one, wait before you connect.
What work does the integration save? Choose work that comes back every week. Our article on automating recurring tasks explains how to spot it.
Which data does the agent need, and which does it not need?
Which permissions are needed per action: read, edit, or publish?
Is every action by the agent logged, so you can check afterward what happened?
Who reviews the work, and does that happen before or after?
How do you pause the integration, and how do you roll back a change?
How do you test an integration in a pilot?
Start with one integration for one task. Write down how much time the task takes now and how often something goes wrong. Check the data, too: an agent works only as well as the fields in your CRM, so tidy up outdated statuses first. Then let the agent run alongside your current process and compare the same numbers at the end. A pilot usually takes 2 to 4 weeks, from the first conversation to the evaluation, as described in our approach. Also look at what the reviewer had to fix: that tells you whether the permissions and agreements are right. To see why one workflow is a better starting point than a big AI plan, read start with one frustrating workflow.
What about personal data in your CRM?
A CRM holds customer names, email addresses, and history. That is personal data, and the GDPR applies. One of its principles is data minimization: according to the European Commission, personal data should be adequate, relevant, and limited to what is necessary for the purpose. For an integration, that means: do not give the agent access to fields or customer groups the task does not use. To see how we protect data and access, read our page on security and privacy.
For more on how we set up integrations, see our page on AI automation. To find suitable processes, read our article on automating business processes.
