How to use AI for sales prospecting: start with the data, not the prompts

Ask ChatGPT or Claude for a list of 200 Finnish machine shops with 50 to 200 employees and you’ll get one. It will contain real companies that fit, real companies that don’t and companies that don’t exist. All three look identical on the list. That’s the core problem with AI prospecting, and it’s why the useful question is less about which AI tool to pick and more about what data you let it touch.

I run my own marketing on AI agents, and there the agent gets away with a lot because it works on material that already exists: my channels, my data, my voice. Prospecting is a different animal. The agent needs to know something about a world it has never seen.

I learned this the hard way

In my own prospecting work I’ve built prospect lists with Clay paired with ChatGPT or Claude. Out of a typical 200-company list, somewhere between 10% and 30% of the rows were bad: companies that no longer existed, subsidiaries of bigger groups that didn’t actually fit the criteria, or plain errors. Even paid enrichment data was sometimes simply wrong. I still had to prune the lists by hand, and bad rows leaked into the CRM anyway. They tend to stay there.

Ten to thirty percent doesn’t sound catastrophic until you follow those rows downstream.

Why a 30% error rate breaks the whole system

Agentic sales systems are chains, and every step trusts the one before it. My company lists synced automatically into LinkedIn ad targeting, so every out-of-ICP giant that slipped onto a list quietly burned ad spend on its own. Wrong contacts went into sequences and came back as bounces, and bounces damage your sender reputation. Rep time went to accounts nobody should have touched in the first place. A bad row doesn’t stay put. Every downstream step pushes it forward.

The research matches what I saw. B2B contact data decays at about 2.1% a month, roughly 22% a year, in the classic HubSpot and MarketingSherpa figure, and Gartner puts business-data decay near 3% a month. Even a clean list doesn’t stay clean. And a language model with no data source is its own problem: without external grounding, models fabricate a meaningful share of their factual answers, which the HALoGEN study measured across task types. My observed 10 to 30% bad rows sits right inside those ranges.

The most expensive loss doesn’t show up in those numbers, though. The companies that are hardest to find data on also get the least outreach. They’re often the segment most worth reaching, precisely because nobody else reaches them, and they’re the first ones bad data drops from your list.

How I run prospecting with AI now

My current system is built on one idea: get the base data right, then enrich on top of it.

• A script scores and ranks the prospect pool, and the CRM is the single system of record. A row either lives there or it doesn’t exist.
• Agents I’ve built gather the enrichment signals: who visited the website (seen through tools like Leadfeeder or Dreamdata), who liked or commented on my LinkedIn content, and, the most interesting one, who’s active on competitors’ posts.
• I send every message myself. The agents read, score and draft. Nothing goes out without me.

Give the agent a data source, not a memory

When I wrote about what an agent needs (a goal, context, tools and loops with checkpoints), tools was the part I defined too narrowly myself: publishing APIs, analytics, inboxes. Data sources belong on that list, and in prospecting they’re the item that decides everything else. With MCP, connecting a source is configuration rather than an integration project: Clevenio’s MCP server, for example, exposes Finnish company and decision-maker data to the agent, so it can query by industry, size and location and get back rows that come from the business registry instead of the model’s memory.

The difference sounds technical but it’s bigger than that. Without a source, the agent produces text that looks like data. With one, it retrieves data and can tell you where it came from.

Where to start

This order has worked for me, and starting small is the cheapest way to learn which of your data you can trust.

1. Connect one source before you build anything else. Ask it a few questions you already know the answers to. You’ll see immediately whether it answers correctly.
2. Give the agent permission to say no. If a search returns nothing, the agent has to report that. Otherwise it fills the gap itself, and that’s exactly the failure that gets these experiments cancelled.
3. Ask for a count before the list. “How many companies match these criteria” is a cheap question, and the answer shows right away whether your targeting makes sense. Three matches and nine thousand matches both mean the criteria are off.
4. Automate last. Once the search works by hand, you can build a loop around it.

What this doesn’t solve

An agent that finds the right companies still hasn’t sold anything. The list is the part of prospecting that genuinely automates, because underneath it’s a search problem. The message, the timing and the conversation itself are still human work, and I haven’t seen evidence that’s changing any time soon. But when the list no longer eats days of your time, that time moves to exactly those things.