AI in Salesforce without surprises
Agentforce, Einstein, connecting Claude or ChatGPT to your CRM data. It works well on one condition: AI gets only the permissions it needs. That is why I start with permissions, not with the agent.
One sentence that explains everything
AI and MCP servers do not add new permissions. They use the ones already there. If an agent runs on an administrator account, it can do anything an administrator can.
An Agentforce agent runs as a designated user with that user’s permission set. Calls through an MCP server run as the logged-in user, respecting object, field and record permissions. That is the good news: the Salesforce security model still applies. The bad news: if permissions are too broad, AI does damage faster than a person, because it does not hesitate before clicking “delete”.
Hence the order: first tidy up administrators, accounts of former employees and apps connected from outside. Then a separate user for the agent with a minimal permission set. Only then the agent.
What I implement
Agentforce and Einstein
Set-up and configuration within what your licence allows. I check the licence before I promise anything.
An agent for customers or staff
Answers questions from CRM data and the documents you point it to. Actions that write data require a person to confirm.
Less typing
Case and call summaries, draft email replies, automatic sorting and assignment of leads. Prompt Builder with templates in your company’s voice.
Salesforce MCP
Safely connecting Claude, ChatGPT and other assistants to the CRM: a separate user, a minimal permission set, read-only wherever that is enough.
An external AI model
When built-in features are not enough or the licence is too expensive, I connect Claude or OpenAI through the API. Personal data masked before it is sent.
Permissions for AI
A dedicated user for the agent, a review of administrators, Connected Apps, OAuth tokens and MCP servers. A log of the agent’s actions.
A CRM report written by AI
A question in plain language, AI writes the query, the result lands in your inbox on Monday. On a read-only account.
Team training
Sales reps learn to use AI in sales and reporting on their own examples. A separate site about training.
How I implement AI in Salesforce
Permissions audit
Who can do what, which accounts the integrations run on, what is connected from outside. Free, within a week.
One task
Not “AI in the company” but, say, case summaries for the service team. Measurable, checkable within a month.
A separate user for AI
A permission set tailored to the task. Read-only where that is enough. An activity log.
A prototype in a sandbox
On a copy of your data. You check the results before anything reaches production.
Go-live and observation
A month of measuring: time saved, errors, credit costs. Then a decision on what comes next.
Experience
I work with AI every day: Claude Code as a developer, Claude and OpenAI in production automations, Agentforce and Prompt Builder in Salesforce. I built my own MCP server with around 75 tools, exposed to the Claude assistant over OAuth 2.1, and a system where AI wrote nearly all of the code while I owned the specification, review and tests. At the company where I own Salesforce, I took part in rolling out AI in finance, accounting, administration and IT.
I read AI-generated code and configuration as carefully as a colleague’s work: platform limits, bulkification, security, tests.
Questions about AI in Salesforce
Do we need an Agentforce licence?
For agents and most Einstein features, yes. For connecting an external model through the API or an assistant through MCP, the licences you have are usually enough. I check this at the start.
Will our data be used to train models?
The Einstein Trust Layer and the API vendors’ terms say no. I check the settings (masking, zero retention, logging) and record them in the documentation.
What if the agent gets it wrong?
That is why actions that write data need a person to confirm, and the agent runs on an account with minimal permissions. A mistake should be noticed before anything comes of it.
An employee has already connected ChatGPT to Salesforce with their own login.
A common case, and the first thing I check in the audit. I replace it with a separate user with a limited scope and switch off unnecessary tokens.
Planning AI in Salesforce? First find out who in your org can do everything.
A free audit of permissions, access and AI settings. Report within a week.
Tell me what needs doing
Describe the problem or the list of things that are waiting. I reply within one business day with questions or straight away with a proposal.