13% → 86% — the jump in AI-generated Salesforce answer accuracy after rolling out Credal's MCP
Checkr is a trust and safety platform that powers hiring for businesses ranging from SMBs to Fortune 500 enterprises, offering tools like background checks, identity verification and compliance screening. With a mission to create a safer, fairer world by helping organizations make trusted hiring decisions, Checkr delivers fast, accurate screening at scale. Through its suite of products, the company supports thousands of employers and millions of workers, helping to streamline hiring and protect workplace communities.
As businesses scale their use of AI models, pointing those models directly at raw Salesforce data becomes an increasingly attractive idea. The reality is less appealing: give a model direct access to raw CRM data, and it will produce incorrect business answers with total confidence.
As more of the business starts making decisions on that data, a harder question follows: how do you make sure people are deciding on accurate, high-quality information? What separates a raw LLM output from a genuinely trustworthy answer?
What's missing from the raw data is exactly the context that makes an answer correct: definitions, workflows, SOPs, when a case should escalate, why routing works the way it does, and what actually counts as an MQL.
Checkr ran an AI eval, pointing Claude directly at raw Salesforce data. The result: accurate only 13% of the time. The model pulled from the wrong fields, queried the wrong objects, and misread the segment definitions.
Checkr runs three separate Salesforce instances, one per product line, each with its own data model and its own segment definitions. Context is splintered across the business: a "strategic segment" in Product Line 1 isn't a "strategic segment" in Product Line 2, and what one instance calls a Sales Qualified Lead (SQL), another calls an opportunity.
What Checkr wanted was simple to state and hard to build: let the entire Revenue org query Salesforce in natural language, across all three instances, and get back answers they could trust inside Claude — whether they were researching an account before a call, analyzing pipeline, or building a pricing quote.
“The goal was to empower business users to use Salesforce data without needing Salesforce architecture expertise. When there are ambiguous fields, custom objects, legacy data — AI will silently choose for you and often get it wrong, leading to risk for our business in time and accuracy.”
Salesforce's API exposes far more than what any user actually sees in the interface. Checkr's team deliberately curates the UI, leaving many fields off-screen even though users technically retain access to them — fields that exist only to support backend automations. When an AI tool queries the raw API, it sees all of it, with no way to tell curated, authoritative data from backend data. That made bad answers very hard to debug.
This ambiguity had always been a RevOps headache, but Claude's rollout turned it into an urgent one. Far more employees were suddenly querying Salesforce through AI tools, and getting back answers that sounded authoritative but were often wrong.
Working with Credal, Zoë built three separate MCP servers, one per Salesforce instance.
That's because the same term can mean completely different things depending on the business unit: a "strategic segment" in Product Line 1 is not a "strategic segment" in Product Line 2, and what counts as a Sales Qualified Lead (SQL) in one instance is an "opportunity" in another.
Each server ships with its own instructions, tuned to give the right data for that business unit.
Checkr's RevOps team saw an immediate jump in the reliability of AI-driven Salesforce answers, from 13% to 86%. By grounding every query in business-unit-specific context, the MCP combats the "correct-sounding but wrong" answers that AI tools produced before.
Checkr's team now queries Salesforce in natural language, across all three instances, directly inside Claude and Cursor — for account research ahead of calls, pipeline analysis, and pricing quotes — and gets back answers grounded in the same business definitions, SOPs, and escalation logic that used to live only in people's heads.
Before building with Credal, Checkr weighed the obvious paths. Each addressed part of the problem; none fit a company running three Salesforce instances and a fast-growing mix of AI tools.
Before partnering with Credal, Checkr explored Claude's native plugins and skills. Limited collaboration on skill and plugin development out of box was a key blocker, as the entire RevOps team needed to contribute critical knowledge.
A skill combined with an MCP wouldn't have solved it — ensuring consistent activation of a skill for every user is nearly impossible. A plugin felt like the wrong fit too: if a user invoked the connector independently of the plugin, they'd lose the critical context that made the answers trustworthy.
Second, the tool-level controls weren't granular enough. With three separate Salesforce instances, Checkr needed to specify exactly which instance each plugin pointed to, and Claude's plugin model didn't support that kind of routing. There was also no way to enforce how skills were read, so the carefully curated instructions about which fields were authoritative and how segment definitions differed couldn't be guaranteed to actually shape the AI's behavior.
Finally, plugins only work in Claude. Checkr needed the same governed, context-rich experience whether someone was working in Claude, Cursor, Lovable, or Slack.
Salesforce now ships hosted MCP servers and an Agent Fabric registry as part of Agentforce, letting developers expose an org's APIs, Flows, and prompt templates to AI agents. But Salesforce's native tooling also only governs within a single org; it doesn't provide a governed, cross-instance routing layer for external AI surfaces. Salesforce's hosted MCP exposes the org's API surface, but it doesn't carry curated instructions about which fields are authoritative or how segment definitions differ across business units. Credal's editable MCP instructions save that knowledge directly.
A basic in-house server connecting one tool to one instance is achievable. The trap is maintenance, which scales non-linearly against Checkr's reality of three instances, multiple AI surfaces, and constantly evolving tooling. Governance — audit logging, human-in-the-loop approval, permissions mirroring, rate limiting — would all have to be built from scratch and maintained indefinitely.
The observability Checkr relied on to debug bad answers and iterate on instructions — logging pipelines, dashboards, alerting — would itself have been a bespoke build.
Reach out to sales@credal.ai to see how Credal's MCP can bring the same accuracy gains to your team.
One platform for all agents.
Full visibility for admins, full access for teams.
