How to Build an AI Sales Prospecting Agent with MCP and B2B Data

An AI sales prospecting agent connects to B2B data through MCP so it can search for your ideal customer profile, narrow the results, request available work contact fields, and format a shortlist. You describe who you want to reach, review the result, and decide which profiles to reveal. You can build this workflow today with a compatible MCP client and a B2B data provider like Descovo.

What an AI sales prospecting agent actually does

Let’s be specific, because “AI sales agent” means different things to different vendors. Some companies use the term to describe full-stack platforms that handle everything from list building to email sending to reply management. That’s not what we’re talking about here.

An AI sales prospecting agent, in the way we mean it, is focused on the data layer. It is an autonomous workflow where the agent:

1. Takes your ICP criteria as natural language input
2. Translates those criteria into structured database queries
3. Searches a B2B database of 700M+ profiles
4. Filters and ranks the results based on relevance
5. Requests available work email and direct phone fields for approved matches
6. Outputs a structured prospect list for review

The agent handles repetitive research steps that would otherwise happen across filter menus, result pages, exports, and spreadsheets. You can refine the request conversationally, inspect the shortlist, and keep control over when reveal credits are used.

The key distinction: this is the data acquisition layer, not the outreach layer. The agent finds the people. What you do with that list — email sequences, LinkedIn touches, cold calls — is up to you and your existing tools. More on that later.

The workflow: from ICP to reviewed shortlist

Here is what the workflow looks like, step by step, using a Descovo MCP connection inside a compatible client such as Claude Code or Cursor.

Step 1: Define your ICP

You start by telling the agent who you’re looking for. Natural language works:

The agent understands what you mean and translates it into structured search parameters. No filter dropdowns, no boolean strings.

Step 2: Agent dispatches peopleSearch through call_operation

Behind the scenes, the agent uses call_operation with operationId: peopleSearch. The request looks like this:

{
  "tool": "call_operation",
  "arguments": {
    "operationId": "peopleSearch",
    "params": {
      "body": {
        "searchParams": {
          "jobTitleV3": {
            "include": [
              {
                "type": "functional",
                "seniority": ["vp", "director"],
                "keywords": ["sales", "revenue"]
              }
            ]
          },
          "industry": {
            "anyOf": ["Software Development"]
          },
          "location": {
            "include": ["United States"]
          }
        },
        "pageSize": 25
      }
    }
  }
}

The MCP server exposes the operation through the call_operation wrapper. The agent constructs the supported request body from your prompt and gets back matching profiles. You can continue refining supported filters such as title, seniority, industry, and location before revealing any contact fields.

Each search costs 0 credits. The agent can refine the list without using reveal credits.

Step 3: Agent filters and ranks results

The agent reviews the results and applies your criteria. It can filter out people who don’t match your seniority requirements, remove duplicates across searches, and rank by relevance. You can also guide it: “Prioritize profiles with revenue leadership titles” or “Put VPs first.”

This is where the agent helps most: it can apply your instructions consistently across the returned profiles, while you keep final control over the shortlist.

Step 4: Agent requests available contact fields

Once you approve the shortlist, the agent makes a reveal attempt for each selected person:

{
  "tool": "call_operation",
  "arguments": {
    "operationId": "syncQuickContactReveal",
    "params": {
      "body": {
        "linkedinUrl": "https://www.linkedin.com/in/example-profile",
        "enrichmentType": {
          "getWorkEmails": true,
          "getPersonalEmails": false,
          "getPhoneNumbers": true
        },
        "validateEmails": true
      }
    }
  }
}

Each reveal attempt costs 1 credit, including not-found results. When available, work email and direct phone are returned together under that credit.

Lite uses a lower-yield lookup. Standard balances speed and coverage. Turbo prioritizes speed, while Exhaustive trades latency for the broadest available provider coverage.

Step 5: Agent outputs a structured list

The agent compiles everything into a clean, structured output. You can ask for it as a table, as CSV, as JSON, or in whatever format your downstream tool expects. No manual formatting. No copy-paste gymnastics.

What the output looks like

A formatted result can look like the illustrative table below. A work email or direct phone field may be unavailable for an individual profile.

The agent can format returned fields as a table, CSV, or JSON for your next step. Review the output before importing it into a CRM or outreach workflow.

How credits work in this flow

Descovo separates free search from metered reveals, so the agent can narrow the list before it spends credits:

• Search costs 0 credits. Calls to peopleSearch and companySearch do not consume reveal credits.
• Reveal attempts cost 1 credit. When the agent calls syncQuickContactReveal, the attempt costs 1 credit even if no contact field is found.
• Available fields are bundled. When available, work email and direct phone are returned together under that credit.

The Free plan includes 500 credits each month. The Starter plan includes 5,000 credits each month at $99 per month standard pricing. A $49 per month launch price is available through August 20, 2026.

Plan your usage in reveal attempts, not returned fields. For example, 50 reveal attempts use 50 credits whether every attempt returns a field or some return not found.

Compare the workflow: dashboard research vs. agent-assisted research

The agent-assisted version reduces context switching and makes the search-and-reveal sequence explicit. It does not remove the need to review returned fields, handle not-found results, and follow your own outreach quality controls.

What Descovo does not do (and what to pair it with)

Honesty about scope matters. Descovo is a B2B data layer, not an all-in-one sales platform. Here is what it does not do:

• No email sequences. Descovo searches profiles and requests available contact fields. It does not send emails on your behalf.
• No CRM integration. The agent outputs data in whatever format you need, but there is no native Salesforce or HubSpot sync.
• No intent data. Descovo tells you who someone is and can request available contact fields. It does not tell you if they are in-market right now.
• No email sending or tracking. Open rates, click tracking, and reply management live in your outreach tool, not here.

This scope keeps Descovo focused on B2B search and reveal operations through MCP. You can pair its structured output with the tools you already use:

• CRM: Salesforce, HubSpot, Pipedrive — import the agent’s output as new leads or contacts.
• Outreach tools: Pass reviewed, available contact fields into your existing campaigns.
• Custom workflows: If you’re building your own AI-powered outbound system, the agent’s JSON output slots directly into your pipeline.

The agent handles data. Your tools handle delivery. Your team handles relationships.

The SDR role isn’t dying — it’s being augmented

Every conversation about AI in sales eventually lands on the same question: “Will AI replace SDRs?” This workflow does not answer that question, but it can change which research steps an SDR coordinates manually.

Research and list building can pull SDRs across databases, profile pages, spreadsheets, and enrichment tools. An agent can coordinate the repeatable search and formatting steps while the SDR reviews the shortlist and chooses what happens next.

This leaves more attention available for the work that still needs human judgment:

• Personalization that goes beyond “I saw you posted on LinkedIn.” Use the reviewed shortlist as a starting point for research into a prospect’s actual challenges.
• Timing. Following up at the right moment matters more than having perfect copy. Time saved on research is time available for strategic sequencing.
• Conversation. The phone call, the demo, the negotiation — these are human skills that AI augments but does not replace.

Effective users give the agent clear, supported criteria — for example, “Find sales directors in New York” or “Find CTOs at software companies in the United States” — then inspect the matches before authorizing reveals.

The data workflow becomes easier to coordinate. The relationship layer still depends on human judgment.

Key takeaways

• An AI sales prospecting agent connects to B2B search and reveal operations through MCP, turning a natural-language ICP into a reviewable shortlist.
• Search costs 0 credits. Each reveal attempt costs 1 credit, including not-found results. When available, work email and direct phone are returned together under that credit.
• The agent reduces context switching. You still review matches, authorize reveals, and handle unavailable fields before outreach.
• Descovo is the data layer, not the whole stack. Pair it with your existing CRM, outreach tools, and custom workflows. It does one thing well.
• Plans provide monthly reveal credits. Free includes 500 credits per month. Starter includes 5,000 per month at $99 standard pricing, with a $49 launch price through August 20, 2026.
• SDRs get augmented, not replaced. The agent can coordinate repeatable research steps. The human focuses on personalization, timing, and closing.