AI prospect list accuracy: measure fit, not list speed

Judge an agent on ai prospect list accuracy: the share of returned rows that truly fit your ICP, plus the share of good accounts it missed, measured on a labeled sample you control. Building a list from a plain-English prompt is no longer a differentiator, so the question for a vendor is how often the list is right and who fixes it when it isn't.

Prompt-built lists are everywhere now. A managed AI SDR vendor shipped one on November 4, 2026, and it was one launch among several that fortnight. When every vendor demos the same trick, the demo stops telling you anything. A fast list that is 60% wrong is a fast way to burn sequences and domain reputation.

Why prompt-built lists stopped being a differentiator

The prompt is the easy part. "B2B SaaS companies, 50 to 200 employees, hiring RevOps" produces a list in seconds, and the list looks plausible. Plausible is the problem. Nobody scrolls 800 rows and checks each one.

So the real work moves downstream, and ai prospect list accuracy becomes the thing you buy. Does each row meet every criterion you care about, and does the agent keep meeting them next month? The first is a measurement question. The second is an ownership question. I'll take them in that order.

How to measure ai prospect list accuracy on your own ICP

Vendor accuracy numbers aren't comparable, because each vendor picks its own prompt, rubric and grader. Your own ICP is the only fair benchmark. Here is what I do, and I'll call it the blind 200.

1. Write your ICP as hard criteria a person could check from public information: segment, size band, a required trigger, a disqualifier.
2. Have the agent return a list, then pull 200 random rows from it.
3. Have someone on your team label each row fit or not fit, without seeing the agent's confidence scores.
4. Count a row as correct only if every criterion holds.

The reason for 200 is arithmetic. If 90% of rows are truly correct, a sample of 200 gives a margin of about plus or minus 4 points (my own calculation, using the standard binomial approximation). At 50 rows the margin is closer to 8 points, too wide to tell a 90% agent from an 82% one.

One caveat: fuzzy criteria like "modern engineering culture" will make your labelers disagree. If two people label the same 30 rows and split on 6 of them, the criterion is the problem, not the agent. Rewrite it before you grade anything.

Precision, recall, and the misses nobody counts

Precision asks what share of returned rows are good. Recall asks what share of all good accounts in the market the agent found. Most evaluations check only precision, because those rows are right in front of you.

Recall is harder, and it's where I don't have a perfect answer. You can't label the whole market. What works in practice is a smaller second sample: take 100 accounts you already know fit (closed-won customers, a hand-built target list) and check how many the agent surfaces from the same prompt. If it finds 55 of 100, the list has a hole, even if every returned row is correct. Real ai prospect list accuracy has both halves.

Which one matters more depends on your motion. For a team with two AEs and a tight territory, precision wins, because every bad row costs a rep's attention. For a team covering a small named-account universe, recall wins. The scikit-learn documentation has the formal definitions if you want to hand them to an analyst.

What ai prospect list accuracy does to cost per qualified lead

Accuracy only matters to the business once it turns into money. Here is a worked example, and the numbers are assumptions, not benchmarks.

Say you spend $2 per row on enrichment and sequencing. The vendor claims 85% fit, and your blind 200 shows 62%. Your real cost per fitting lead is $2 divided by 0.62, about $3.23, instead of the $2.35 the claim implies. On 5,000 rows that gap is roughly $4,400 spent on accounts that were never going to buy, before you count rep time and sender reputation. Deliverability limits are their own subject, which I covered in four deliverability limits to demand from outbound agents.

Who corrects the agent when fit drifts

Fit drifts. Your ICP changes after a quarter of closed-lost data, a data source goes stale, or someone new edits the prompt. An agent that scored 90% in the pilot can sit at 70% by spring, and nobody notices because the list still looks fine. A pilot score of ai prospect list accuracy is a snapshot, not a guarantee.

So ask three plain questions before you sign:

• Who re-runs the blind 200 each month, you or the vendor?
• Who edits the criteria when a rule of thumb breaks, and how fast?
• Where do rejected rows go, so the correction sticks instead of repeating?

For better or worse, the answer is usually "you, but we have a dashboard." A dashboard of credits used isn't a measure of fit, which is the point of tying usage metrics to outputs rather than credits. It's also why bottom-up rollouts go sideways when thirty reps each run their own agent with their own prompt, a pattern I wrote about in what RevOps must own first.

At Coco, this is why we run on approval gates: a person reviews and signs off on the list before it reaches the CRM or a sequence, and every rejection teaches the next run. You can read how that works in AI with approval gates. It costs some speed. I'd take that trade.

What to put in a pilot

If you're evaluating agents now, treat ai prospect list accuracy as a pass or fail gate, and ask for the target in writing before the pilot starts. The same logic applies to the "agentic" label, and five tests for spotting agent washing covers how to tell a product change from a homepage change. Since most buyers won't rip out their stack, what a pilot must prove alongside your existing tools is the companion checklist.

My recommendation: set a precision floor (I'd start at 80% on your blind 200), set a recall check against known customers, name the person who owns monthly re-grading, and compare vendors on the same prompt and the same sample. I'm not certain 80% is right for every motion, and a high-volume, low-touch team might accept less. But a number you chose beats a number the vendor chose.

Ai prospect list accuracy is a measurement you own, not a feature you buy. If the vendor can't show you the blind 200 results and name who fixes the drift, the list builder is a demo, not a system.