AI tools are great with language, but not your business. Ask them about your pipeline or lead routing, and they’ll confidently guess. Without access to your systems, objects, and logic, their answers fall short.
Users now expect natural language interfaces for business systems, but platforms like Salesforce weren’t built for that. Without live data and system access, AI cannot help your business run effectively.
Model Context Protocol (MCP) bridges that gap by connecting Large Language Models (LLMs) directly to systems like Salesforce. Sweep takes it further – translating messy metadata into structured, executable context. This article explores what MCP is, how leading platforms are adopting it, and how to make Salesforce a truly conversational, agentic system.
What Is Model Context Protocol (MCP)?
Model Context Protocol (MCP) is an open-source standard that enables AI models to interact securely and intelligently with external systems. First introduced by Anthropic and quickly adopted by OpenAI and Microsoft Copilot, MCP is gaining traction across the AI ecosystem. Today, leading AI platforms like Sweep, Asana, and Intercom are actively building MCP servers to make their systems AI-accessible.
Think of MCP as the “USB-C for AI,” a universal interface that allows large language models (LLMs) to connect with business tools and data without the need for custom integrations.
Unlike traditional APIs, which require rigid, service-specific calls, MCP provides a dynamic layer of abstraction that agents can interpret in real time. Rather than hardcoding every endpoint and field, AI agents can discover available capabilities, understand their requirements, and interact contextually, without being explicitly programmed for each case.
For example, without MCP, an agent integrating with an ERP system to update inventory must be manually wired to a specific endpoint like /inventory/update, with fragile assumptions about field formats and logic. With MCP, that same agent can query the system for available actions, interpret the required schema, and adapt to changes like a new inventory rule or an updated ERP entirely.
This shift enables AI agents to reason across systems, take action, and deliver real business outcomes. By accessing the right context and adhering to the right rules, AI can meaningfully contribute to business workflows and help teams move faster with less manual effort.
How Leading Platforms Are Adopting MCP
Salesforce
Salesforce has not yet released a general-purpose MCP server for its platform. Access to MCP remains restricted to controlled use within Agentforce, its internal AI agent framework. As of Agentforce 3.0 (July 2025), a native MCP client is in pilot, and Slack is developing its own MCP server. These integrations are premium-tier and largely limited to early access customers.
One example: an Agentforce-based AI agent is configured to pull order data from Salesforce, generate an invoice in PayPal, and trigger a Slack notification, all through MCP. Tool access is managed via certified servers in Salesforce’s AgentExchange, and business policies (e.g. flag invoices above $10K) are enforced through natural language instructions. The result is a cross-system workflow that doesn’t require custom integrations.
Outside of Salesforce’s ecosystem, independent developers have launched open-source MCP server implementations that expose capabilities like SOQL queries, metadata access, and object introspection, making it possible to experiment with Salesforce-integrated AI agents beyond the confines of Agentforce. However, these implementations often lack enterprise-grade security, access governance, and compliance safeguards.
Salesforce’s adoption of MCP, while promising, remains tightly scoped and proprietary in its implementation. Agentforce shows what’s possible when AI agents operate across tools with context and control, but the broader Salesforce platform remains closed to external builders. As open-source alternatives continue to mature, and as MCP gains traction as a shared standard, the pressure will grow for even dominant vendors to embrace more interoperable, agent-ready architectures. Whether Salesforce opens its ecosystem or doubles down on its embedded approach may define how accessible, flexible, and AI-compatible its platform truly becomes.
HubSpot + ChatGPT
HubSpot was the first major CRM to ship a production-grade MCP integration. In June 2025, it launched a “deep research” connector that allows users to ask ChatGPT natural language questions and receive live answers from their HubSpot data.
Once connected, users can query their CRM with prompts like: “Summarize my pipeline by region and flag high-risk accounts.” Behind the scenes, ChatGPT calls HubSpot’s MCP server, retrieves the necessary data, and generates a contextual response without the need for custom integrations or predefined logic.
In contrast to Salesforce’s more closed and premium-gated MCP rollout, HubSpot’s approach is built for open interoperability with a wide range of language models. By exposing a public MCP server, HubSpot positions itself as part of a broader AI ecosystem rather than locking users into a proprietary assistant.
This shift reflects a broader trend in CRM design, moving from static dashboards to conversational interfaces that combine live data, open standards, and flexible AI agents.
Challenges of Using MCP With Salesforce
Letting a language model query enterprise data directly sounds great in theory. In practice, it often breaks down. Data is fragmented, business logic is buried in automations, and queries are frequently slow or inaccurate.
Raw Metadata Is Not Real Understanding
MCP servers can expose objects and fields, but without semantic context, models often misinterpret what they’re querying. For example:
- Query: “How many demos were scheduled by Dan in the last 30 days?”
- Issue: Dan is auto-assigned every lead, including no-touch ones. Counting all demos under his name inflates attribution.
- Outcome: An out-of-the-box MCP server may return a number, but it won’t know to exclude
no_touch = true, leading to answers that sound correct but aren’t.
Volume and Complexity Overwhelm Models
Large Salesforce orgs often include thousands of flows, automations, and custom rules. Native MCP clients are not built to handle that volume. They lack prioritization and can’t infer which objects or logic actually matter in a business context.
No Agentic Reasoning or Memory
MCP clients treat every request as isolated. There’s no memory of past steps or ability to track multi-step processes. Without this, models can’t break down complex workflows, follow a chain of logic, or carry context from one action to the next. This makes it difficult to support real operational tasks like tracing flow logic or auditing permission settings.
Blind Spots in Business Semantics
Even with metadata access, models still lack understanding of how your business works. They don’t know what qualifies a lead, which processes are critical, or how internal teams interpret key terms. Without this semantic grounding, their outputs can be technically accurate but practically useless.
Sweep MCP: Bridging the Gap Between Natural Language and Business Systems
Where native MCP falls short, Sweep bridges the gap. Sweep does more than connect Salesforce to large language models. It transforms metadata into structured, actionable context, allowing models to interpret, reason, and take meaningful action with accuracy.
Context Layer
Sweep ingests metadata, system logs, and object relationships. It applies process mining and dependency mapping to surface critical information, including:
- Business logic and gating conditions.
- Routing rules and ownership assignments.
- Flow structures and automation triggers.
This metadata is pre-processed into a graph of logic and relationships. That gives AI the ability to navigate complex systems, reason across workflows, and return answers grounded in how your business actually operates.
Application Layer
Sweep layers on tools for change detection, QA, tech debt analysis, permissions, process design, and more. These are agent-ready and coordinate multi-step tasks, detect edge cases, and operate within enterprise guardrails.
Here’s how Sweep enhances the MCP experience compared to a native implementation:
| Native MCP alone | Sweep + MCP Layer |
|---|---|
| Operates on raw metadata | Interpreted and enriched |
| No understanding of GTM semantics | Context-aware and business-aligned |
| One-shot NL-to-query | Multi-step, agentic reasoning |
| Struggles with volume | Pre-processed and orchestrated |
| Blind to edge cases | Heuristic and pattern-driven analysis |
Final Thoughts: Why Context Is the Missing Layer
Natural language is becoming the default interface for systems of record. But metadata alone is not enough. Models need context – semantic, structural, and operational – to perform reliably at scale.
That context must be delivered through an intelligent middleware layer. Without it, AI is blind to enterprise logic. With it, users gain safe, intelligent, and seamless access to the full power of their systems.
MCP makes this possible. Sweep makes it reliable. This is the bridge between static dashboards and adaptive, conversational enterprise systems.
Curious how Sweep MCP works with your Salesforce org? Book a demo here.
