Salesforce plans to bill Flex Credits for third-party AI agents accessing the platform via the Model Context Protocol (MCP) or direct APIs. Every successful call by an external agent will then cost you money. That creates a dangerous incentive: to save costs, companies might keep agents away from Salesforce and build error-prone data silos instead.
Security vs. pricing model
For months Salesforce has stressed how important secure interaction between AI agents and the platform is. Tools like the Headless Toolkit and Agentic Identity give external models such as Claude or ChatGPT controlled access to data and workflows. They separate permissions and make actions traceable.
Billing via Flex Credits torpedoes this approach. If the most secure way into the platform is the most expensive one, you will inevitably look for ways around it. Many customers already complain that they are suddenly supposed to pay for standard tasks that used to run free of charge via classic APIs, just because an AI agent triggers the call.
Why agents drive API costs
AI agents work differently from us humans. You usually know which record you need. An agent feels its way forward step by step: it discovers the data schema, searches for the right record, reads supporting data and only then executes the action.
Each of these steps potentially counts as a separate interaction. The more autonomously your agent acts, the more sub-tasks it generates and the faster your Flex Credits melt away. In the end you pay twice: the LLM provider for the language model and Salesforce for the access.
What is the Salesforce architecture?
The Salesforce architecture is based on a multi-tenant cloud structure. All customers use the same infrastructure, while the data stays strictly separated. The central principle is a single data model, the “single source of truth”. External applications and AI tools interact with this database directly via secure APIs instead of copying data. Exactly this principle comes under pressure from the new pricing model, because direct interaction becomes expensive for autonomous agents.
The danger of data silos
If every agent call costs money, you will try to minimise those calls. That provokes risky architecture decisions. Teams might create secondary copies of their Salesforce data for the agent to access. Or developers move the logic into external tools and use Salesforce only as plain storage for the final result.
Such workarounds bypass Salesforce’s security mechanisms. Agent capabilities end up in GitHub repositories where business users can no longer manage them. Add to this the problem of data quality: industry reports show that 77 percent of B2B Agentforce implementations fall short of expectations because the CRM data is dirty. If you copy this data into external silos, your agents will soon decide on a completely outdated basis.
The real cost of AI in CRM
Introducing Agentforce costs serious money. An example from manufacturing: a team of 50 users paid 416,500 US dollars in the first year for implementing Agentforce Industries. That included 75,000 US dollars for technical setup and 100,000 US dollars for training. Recurring costs are almost 7,500 US dollars per user and year. Possible consumption overages through Flex Credits are not yet part of this calculation.
Hybrid architectures as a way out
You need to weigh carefully which processes you map natively via Agentforce or MCP and which you move elsewhere. For complex customer service cases, where the agent reaches deep into the Salesforce data structure and needs context, the official route pays off. Native integration cuts handling time and justifies the price.
For simple workflows with a clearly defined outcome, classic APIs or middleware solutions are the better fit. That way you prevent your budget from evaporating on simple data queries, while keeping data sovereignty in the CRM.
We at Aviando analyse your processes and design an AI architecture that protects your Salesforce data and avoids unnecessary API costs. Let’s discuss how you integrate agents efficiently.
