AI that acts, not AI that demos well.
In August 2026 Salesforce and Anthropic announced Claudeforce, making Claude the platform's default reasoning engine. We were already working that way before the announcement: Nitzap has exposed an MCP server in production since launch.
The difference between an AI pilot that dies and one that sticks is always the same. It is not the model. It is whether the AI has real context and permission to act within the business rules.
Why most AI pilots stop at the pilot.
- The agent answers well in the demo and gets it wrong on real operational data.
- Nobody can say what the AI did, or under whose permission.
- The AI suggests, but somebody still has to do everything by hand afterwards.
- The project turns into endless integration because every system speaks its own language.
- Legal stops it at the first question about sensitive data.
- Token cost blows up before any return shows.
Three layers, in the order that works.
Starting with the good-looking agent is the fastest route to a pilot that goes nowhere.
1. Context
Data before the agent. Model, quality and governance, so the AI answers from what is actually true in your operation.
See data consulting →2. Connection
An MCP server and APIs expose CRM, WhatsApp and ERP to the AI, under the permission of the user who asked.
See the MCP server →3. Action
Agentforce, Einstein Bot and Flows actually execute: they create records, schedule, reply and escalate to a human.
Talk about agents →What changes with Claude inside the platform.
The announcement brings the Atlas Reasoning Engine, Agentforce Vibes and Claude as the default in Slack. For anyone who already has Salesforce, that means new capability without changing platform. And for anyone building on the platform, it means the agent can finally call real tools instead of only generating text.
Because we are a Salesforce partner and keep MCP in production, we connect both ends without waiting for anyone's roadmap.
Real context
The AI reads live records and conversations, under the permission of whoever asked.
Auditable action
Every write passes through the org rules and is logged.
Where the team already is
Slack, WhatsApp, the CRM or whichever chat the company uses.
Seconds, not hours
The manual scramble becomes a single call.
What you can do today, with what you already have.
- Pull unanswered conversations and create an opportunity for whoever asked for a quote.
- Summarise the client history before the meeting, with what matters to the negotiation.
- Qualify a lead by score and suggest the next step to the rep.
- Generate an SLA report and clean the database on demand, in natural language.
- Answer repetitive questions on WhatsApp and escalate to a human when it goes off script.
- Suggest cross-sell and protect margin while building the quote.
The CRM stops being a screen and becomes a service.
That is what headless means: the business rules, the data and the permissions stay in Salesforce, but people no longer need to open it to work. Service happens where the team already is, and the record is written on the other side.
What makes that possible is the MCP server and the Nitzap APIs. It is the same mechanism that lets the AI operate: if an agent can read and write in the CRM, a custom interface can too.
- Service outside Salesforce, with queues, connections and history tied to it.
- No license for people who do not use the CRM, such as partners, contractors or field operators.
- A single database, because the data never leaves Salesforce.
- Three API modes living together: official WABA, coexistence and Web API.
What the board and legal ask.
Want to see it running?
We show MCP executing inside Salesforce, on real data from an operation.
WhatsApp +55 27 99997-0276Do we need an Agentforce license to start?
It depends on what you want to do. Flow automation and the Nitzap MCP server work without it. Salesforce conversational agents require the license, and we help you size it before you buy.
Will the AI access data it should not?
No. Access uses the permission of the user who asked, with the profiles and sharing rules your org already has. The AI does not see what the person would not see.
Can we use a model other than Claude?
Yes. The MCP server is open: Claude, OpenAI, Gemini, Bedrock or Copilot. Claudeforce makes the path inside Salesforce easier, but does not force the choice.
How do we control the cost?
With context caching and a well defined scope. In Nitzap, for example, the insight for a record is reused while the record does not change, instead of reprocessing on every open.
How long until a first case in production?
A well defined case, with the data already organised, usually goes live in weeks. What stretches an AI project is almost always the data layer, not the agent.
Bring a case, not a topic.
AI solves specific problems. Tell us which process eats your team's time and we will say whether it can be automated today.
Rather talk now?
Sales team on WhatsApp, Monday to Friday, 9am to 6pm (BRT).
WhatsApp +55 27 99997-0276