What is AI prospecting, and which jobs should AI do for you?
AI prospecting is using AI to do the repeatable work between "we sell to companies like this" and "a person is talking to a person": building the list, finding the right people, researching accounts, spotting triggers, drafting outreach, sorting replies and logging it all. Our rule: hand AI any job that is repeated, checkable in under a minute and cheap to get wrong, let it draft anything that fails one of those tests, and keep the jobs that fail two or more (choosing the segment and every live conversation) plus the final pick on the one line that only fits this company. Done that way, our invented worked example below frees about 63 hours a month for one rep; done the other way, it just sends the wrong email faster.
AI makes you faster at whatever you are already doing. Pointed at the wrong job, it makes you wrong faster.
How we research this: facts come from each vendor’s public pages and named studies, dated where shown; the math is ours and labeled. Vendor prices and limits are as of October 9, 2026.
Updated
What counts as prospecting, and where does AI fit?
Prospecting is the pipe between a segment and a conversation: twelve jobs, in order, in the table below. AI touches every stage except the first and the live reply.
An AI SDR is a product that bundles five or six of those jobs and sends for you (see what an AI SDR is). This page is the job list behind any bundle, including the cheap one: a general assistant plus a spreadsheet. And no AI step rescues a missing ideal customer profile.
What is the three-question test for handing a job to AI?
Our guidance, not a vendor standard. Ask it of any job:
- Repeated? Does it happen the same way 50 or more times a month?
- Checkable? Can a human tell in under a minute whether the output is right?
- Cheap if wrong? Does a mistake cost a credit or a correction, or does it cost you a prospect?
Three yeses: AI runs it, you spot-check. One no: AI drafts, you check every output. Two or three nos: keep it human.
"Checkable" is the hinge. OpenAI's research post of September 5, 2025 says that when models are graded only on accuracy, they are "encouraged to guess rather than say 'I don't know.'" A confident brief is not evidence; demand a source you can open.
"Cheap if wrong" is about buyers. A wrong CRM field costs you a two-second fix. A wrong first line costs you the prospect, and no follow-up buys that back.
Which prospecting jobs should AI do, and which should you keep?
Twelve jobs, one table. Verdicts follow our test.
| Job | Repeated? | Checkable fast? | Cheap if wrong? | Verdict | You still |
|---|---|---|---|---|---|
| 1. Choose segment and disqualifiers | No | No | No | Human | Decide |
| 2. Build the account list | Yes | Yes (sample 30 rows) | Yes | AI runs it | Read 30 random rows |
| 3. Find the people who own the problem | Yes | Yes | Mostly | AI runs it | Keep a title map |
| 4. Verify emails and phones | Yes | Yes | No (bounces hit your domain) | Verifier tool | Verify before every send (options) |
| 5. Research the account | Yes | Yes, if every fact cites a source | No | AI drafts, you check | Open one source per fact |
| 6. Watch for triggers | Yes | Yes (events are public) | Yes | AI runs it | Judge if the trigger is a reason to write (for a hiring push, see how to read a job posting as a sales signal) |
| 7. Rank who gets worked today | Yes | Yes (if the rule is written) | Yes | AI runs it | Write the scoring rule |
| 8. Write the first line | Yes | Yes | No | AI offers options | Pick, edit, then swap in another company name: if it still fits, rewrite |
| 9. Write follow-ups 2 to 5 | Yes | Yes | No (it lands in the prospect's inbox) | AI drafts | Approve once per sequence |
| 10. Send on schedule | Yes | Yes | Mostly | Software | Cap at 30 sends per inbox per day (our guidance) |
| 11. Sort replies | Yes | Yes | No (a misrouted warm reply goes cold) | AI sorts | Read every positive |
| 12. Answer replies, run the call | No | No | No | Human | This is the job |
Count it: AI runs four jobs, drafts or sorts four, and you keep two plus the final pick on every first line. Verifying and sending aren't judgment calls, so plain software beats any model there. The two human jobs decide whether anything else matters.
For bundles, try Pick the Stack.
Where does AI prospecting go wrong?
Volume without targeting. An AI list plus autopilot sending is the outbound activity trap with a faster engine. Google's sender guidelines say to keep your "spam rate below 0.1%" and prevent it "from ever reaching 0.3% or higher" (bulk sender: close to 5,000 messages a day to personal Gmail accounts). That is Google's rule, not an industry average; see spam complaint rate guide.
Confident research nobody checked. Invented illustration: the brief says "they opened a Dallas office," with no source. The prospect never had one.
Trivia posing as relevance. Ask a model to "personalize" and it digs up an alma mater or an old podcast. Fun facts, zero buying reasons.
Letting AI choose the segment. A vague prompt returns 2,000 "lookalikes" nobody qualified.
Worked example: how many hours does AI prospecting give back to one rep?
Invented example, our math. One rep on our house plan: 4 inboxes x 30 sends a day x 21 sending days = 2,520 emails a month, and across 5 touches that is 2,520 / 5 = 504 new prospects. Minutes per prospect, before and after handing AI the jobs the table says it can have:
| Job (from the decision table) | Manual, min per prospect | With AI on the right jobs, min per prospect |
|---|---|---|
| Build list + find people (jobs 2 and 3) | 2 | 0.5 |
| Verify (job 4) | 0.5 | 0.5 |
| Research brief (job 5) | 3 | 1 |
| First line (job 8) | 3 | 1.5 |
| Follow-ups 2 to 5 (job 9) | 2 | 0.5 |
| Sort replies + CRM logging (job 11 plus admin) | 1.5 | 0.5 |
| Total | 12 | 4.5 |
The arithmetic:
- Manual: 12 x 504 = 6,048 minutes, or 100.8 hours a month. Which is why reps under time pressure quietly skip research and the first line.
- With AI: 4.5 x 504 = 2,268 minutes, or 37.8 hours.
- Difference: 6,048 - 2,268 = 3,780 minutes, or 63 hours a month. Over 21 sending days, 3 hours a day (invented example).
- First lines still cost 1.5 x 504 = 756 minutes (12.6 hours) and source checks 1 x 504 = 504 minutes (8.4 hours). Protect that judgment; spend the freed hours on calls.
Our 3-minute research figure matches how much time to research a prospect; research only your top tier deeply and totals land lower.
Counter-example, also invented: the same rep hands AI job 8 unedited and job 1 ("find me 2,000 lookalikes"). Time per prospect falls to about 2 minutes (0.5 + 0.5 + 0 + 0 + 0.5 + 0.5, our estimate, with zero minutes on research and first lines), but 2,520 emails now hit unqualified accounts with first lines that fit any company. We won't invent a reply rate for that; you already know it.
How do you set this up in one week?
Our guidance, as a one-page plan:
- Day 1: name the segment in one sentence and list five reasons an account is out.
- Day 2: pull a small list, around 200 accounts, with filters. Let Prompt A grade a random 30.
- Day 3: map titles to people, run every address through a verifier, and cut anything that fails.
- Day 4: run Prompt B on your top 20 accounts. Open one source per fact.
- Day 5: AI gives five first-line options per account; you pick and edit. AI drafts follow-ups 2 to 5; you approve once. A human answers every reply.
- Week 2 onward: compare positive replies and meetings held per 100 prospects against your last month without AI, using the calculator, which estimates sends, replies and meetings.
Two prompts of ours; five more are in AI prompts for sales prospecting.
Prompt A, list audit (job 2):
You are auditing a prospect list. Our target: [SEGMENT IN ONE SENTENCE]. Disqualifiers: [LIST 3 TO 5]. Here are 30 rows: [PASTE COMPANY, INDUSTRY, SIZE, TITLE]. For each row return: fit (yes / no / unsure), the one reason, and whether the title is likely to own [PROBLEM]. End with two counts: no, and unsure. If a field is blank, mark unsure.
Prompt B, research brief with a source rule (job 5):
Build a brief of up to 5 bullets on [COMPANY] for a first touch about [YOUR PROBLEM AREA]. Use only this material: [PASTE HOMEPAGE TEXT, 2 JOB POSTS, 1 NEWS ITEM]. Each bullet: the fact, why it matters to a [TITLE] (label this part "inference"), and which pasted item it came from. If a bullet has no pasted source, write "no source found" instead of the bullet. Rules: no compliments, no adjectives about the company.
Which AI prospecting tools do which jobs?
Vendor facts from each vendor's own page, read 2026-10-06.
- General AI assistant: jobs 5, 7, 8, 9. No list access of its own: it only knows what you paste in.
- Clay (Claygent): jobs 5 and 7 inside Clay's tables. Clay's page says Claygents "Research target companies and people with AI" with web research, lists lead scoring as a use case, and charges "1 action per AI prompt" plus data credits (Clay review).
- HubSpot prospecting agent: jobs 3, 6, 8, 10. HubSpot's help page says it detects companies matching your buying signals, identifies contacts and generates outreach, with "Review before sending" and "Send automatically" modes, on Starter, Professional or Enterprise tiers, using HubSpot Credits per personalized email.
When should you skip each AI prospecting tool?
- General assistant: skip it as your list builder. It has no database, and pasting 500 rows into a chat is how data goes stale and leaks.
- Clay: skip it if your monthly list is small enough to research by hand; every AI prompt costs an action, so a 50-account month does not need a credit meter.
- HubSpot prospecting agent: skip it if HubSpot is not your CRM, and skip "Send automatically" until "Review before sending" has earned it.
- Any AI SDR bundle: skip it until jobs 1 and 8 are written down. A bundle automates whatever it is handed (comparison).
The sending days, sends per inbox, inboxes per rep and touches per contact on this page are our house assumptions, with one line on why we use each.
Sources
Research and vendor pages accessed October 6, 2026.