The Cheapest Way to Qualify 10,000 Leads With AI

The cheapest way to qualify 10,000 leads with AI is to ask one or two yes/no questions per row. On Columns that is 10,000 to 20,000 row-questions, and both fit inside the Starter plan at $29 a month. By the end of this post you can price your own list in three lines and pick the plan that covers it.

Updated October 13, 2026. All prices are in US dollars.

The cost is rows times questions, nothing else

One row-question is one question asked about one row. That is the unit you pay for, so the math is a single multiplication:

row-questions needed = rows x questions per row

The plans on Columns are priced by usage, never per seat:

Cached answers are not billed again. If you re-run a list and nothing in a column has changed, you do not pay for those rows twice.

Price your list in three lines

Example: say you have an Apollo export of 10,000 companies and you want to know which are B2B SaaS and which sell to small businesses.

1. Rows: 10,000.
2. Questions: 2 ("Is this a B2B SaaS company?" and "Does it sell to small businesses?").
3. Row-questions: 10,000 x 2 = 20,000. That fits inside Starter's 25,000, so the bill is $29 for the month.

Divide the plan price by the rows and you get the cost per row: $29 / 10,000 = $0.0029. That is arithmetic on a plan price, not a measured result, but it is the number to hold up against any other option.

Here is what the column looks like, with invented values:

The third row is the point. A 58 is the AI telling you to look at that one yourself.

Add a third question and you change plans

The same list with three questions is 30,000 row-questions. That is over Starter's 25,000, so you are on Pro at $99 a month. Five questions per row is 50,000, still Pro.

The rule: when rows x questions is 25,000 or less, stay on Starter. Above that, and up to 150,000, go to Pro. Before you add a question, ask whether it changes who you contact. If the answer to a question would not move a single row from "call" to "skip", cut it. Two questions that decide fit beat five that describe the company.

If you are choosing those questions now, the yes/no format works best, and this guide to yes or no questions AI can answer about a website lists the ones that hold up.

Test 500 rows free before you spend $29

The free plan is 500 row-questions with no card. Use it on a real slice of your list, not a clean sample you built to look good.

1. Take 500 rows from the middle of your export, not the top. The top is usually the cleanest.
2. Write one question as the column header.
3. Read the confidence column. Sort ascending and read the bottom 50 rows yourself.
4. If the low-confidence rows are genuinely ambiguous (a holding company, a site with no copy), the question is fine. If they are rows where the answer was obvious to you, rewrite the question and run it again.

This costs nothing and tells you whether the question works before it touches the other 9,500 rows. For where to draw the cutoff once you have the numbers, see what confidence threshold to use to review AI answers.

Count the review time, because it is the real cost

The plan price is the small part. The part that grows is the rows a human still has to read.

Example: say 8% of your 10,000 rows land under your confidence cutoff. That is 800 rows. At 30 seconds each, that is 24,000 seconds, about 6.7 hours of checking. At $10 an hour that is roughly $67.

The formula to repeat with your own figures:

review hours = (rows x share under cutoff x seconds per row) / 3,600

Time yourself on 20 rows first and use that pace, not a guess. Then tighten or loosen the cutoff until the review pile is a size you will actually finish. Reviewing 800 rows is a Friday afternoon. Reviewing all 10,000 is the job you were trying to avoid.

Against a VA: weeks of calendar time, not just dollars

Published rates for lead generation virtual assistants run about $10 to $25 an hour, and the same page describes a lead qualification VA processing 50 to 100 or more prospects a day. At that pace 10,000 rows takes 100 to 200 working days.

Example: say your VA costs $10 an hour for an 8 hour day and gets through 75 rows a day. 10,000 / 75 is about 133 days, and 133 days x $80 a day is roughly $10,700. Your rate and pace will differ, so plug in your own.

The VA formula:

VA cost = (rows / rows per day) x hours per day x hourly rate

Two things to notice. First, the calendar time matters as much as the dollars, because a list that takes six months to qualify is stale before it is done. Second, judgment questions are slower than the headline pace. Time your VA on 20 rows of your list, not the vendor's average. If you want the longer version of this comparison at 5,000 rows, this article on qualifying 5,000 leads without reading every website walks through it.

Against a DIY API script: cheaper in tokens, more expensive in hours

If you can write code, a small model API is the lowest raw cost. GPT-5 nano is listed at $0.05 per million input tokens and $0.40 per million output tokens, and GPT-4o mini at $0.15 and $0.60.

Example: say each row needs 1,000 input tokens (the page text you feed in) and 20 output tokens (a yes or no with a reason).

• GPT-5 nano: 10 million input tokens x $0.05 = $0.50, plus 0.2 million output tokens x $0.40 = $0.08. About $0.58 for 10,000 rows.
• GPT-4o mini: 10 million x $0.15 = $1.50, plus 0.2 million x $0.60 = $0.12. About $1.62.

The formula: (rows x input tokens / 1,000,000 x input price) + (rows x output tokens / 1,000,000 x output price).

That number leaves out everything that actually costs you. You have to fetch each website, write and debug the prompt, parse the answers, handle failures and retries, and then work out which of 10,000 answers are wrong, because a plain API call gives you an answer and no signal about which ones to doubt. Sales-Mind states AI can cost 50 to 75 percent less per lead than manual qualification, but notice that a percentage like that says nothing about the hours you spend building the thing. For the DIY route and where it breaks, this piece on whether ChatGPT can qualify leads from a CSV covers it.

The rule: when you can code and a rough first pass is fine, the API is cheapest in dollars. When you need to know which answers to trust, or you do not want to maintain a pipeline, pay for a tool that gives a confidence number with every row. When both bills are a few dollars, decide on hours, not tokens.

Three mistakes that quietly double the bill

• Asking five questions when two decide fit. 10,000 rows x 5 is 50,000 row-questions, which moves you from $29 to $99.
• Skipping the 500 row test. You find out the question is vague after you spent the allowance on it.
• Reviewing every row anyway. Checking all 10,000 by hand erases the saving over a VA. The confidence number exists so you do not do that. It is also not a promise that high-confidence answers are always right, so spot-check a handful above your cutoff too.

Do this in the next ten minutes

1. Write down your row count and the two questions that decide whether a lead is worth a call.
2. Multiply: rows x questions. Find that number in the plan table above.
3. Pull 500 rows from the middle of your export into a CSV.

That is the whole setup for a free test.

Run the 500 row test on your own list

You now know the number you are buying and what the review will cost. The free plan gives you 500 row-questions with no card, enough to see the confidence column on your own data before you pay anything. Start free, no card.