Five things to check before you switch on Agentforce in a small company
Agentforce is sold as “an agent that works for you”. In a large company a team rolls it out. In a small company one person switches it on on a Friday afternoon. This list is for that person.
1. Licence and cost, before you fall in love with the demo
Agentforce requires a licence and is billed by usage (conversations, actions). The demo does not show you the bill. Before switching on: which Salesforce edition, which add-ons, what an agent conversation costs, what the limit is and who gets an alert when it runs out. In a company where the agent answers customers on the website, the cost depends on traffic, not on you.
An alternative worth knowing about: for many tasks (summaries, draft emails, questions about your data) an external model connected through the API, or an assistant through MCP, on the licences you already have, is enough. Cheaper, though with fewer ready-made building blocks.
2. The data the agent will work on
The agent answers from what is in the system. If half the accounts have an empty “industry” field, the agent will not guess. If the knowledge base has three articles from 2021, the agent will quote them with confidence. Before switching on: which objects and fields the agent will read, whether they are filled in, whether they are current, whether duplicates will make it answer the same question two different ways.
This is usually the biggest piece of work and at the same time the biggest value of the rollout. Companies that tidy up their data “for AI” gain even if they never switch the agent on in the end.
3. Permissions: which account the agent runs on
The agent runs as a designated user and has that user’s permissions. Not the administrator account. Not an integration account with full access. A separate user with a permission set for one task: read what is needed, write only where it is planned, sensitive fields switched off. More in the piece on agent permissions.
On top of that, the Einstein Trust Layer settings: masking personal data before it goes to the model, zero retention, a conversation log. Switched on by default does not mean checked.
4. One task, not “AI in the company”
An agent that “helps sales” cannot be evaluated. An agent that “writes a summary of every closed case and proposes the next two steps” can. Pick one task that takes someone’s time today, measure that time, switch the agent on there and nowhere else. Actions that write data get a human confirmation at the start. After a month you know whether it works, what it costs and whether it is worth extending.
5. Who looks after it once it is on
An agent is not a project that ends. Models change, Salesforce ships three releases a year, data goes stale, employees find new ways of asking questions. Someone has to go through the conversation log once a month, check the costs and adjust the agent’s instructions. If the company has no administrator, that is a job for whoever looks after the system. If nobody looks after it, after six months the agent will answer worse than it did on day one and nobody will know why.
The list in brief
| Point | Question | Who answers |
|---|---|---|
| Licence | What does a conversation cost and who gets the limit alert? | Owner, administrator |
| Data | Are the fields the agent reads filled in and current? | Head of sales, administrator |
| Permissions | Which account does the agent run on and what can it write? | Administrator |
| Task | One, measurable, with a “before” time? | Owner |
| Care | Who reviews the log and the costs every month? | Whoever looks after the system |
Point 3 and part of point 1 are what I check in the free audit. The rest is a conversation I am happy to have before you switch anything on.