Account 41 gets a worse write-up than account 3, and nothing about account 41 earned that: by then you have pasted the same prompt forty times. Most advice on how to use AI for prospecting demos one account and never the list. What holds up on a whole list is five steps run in order: write the ICP, research every account the same way, score each against the ICP, write the first line from that research, and send it.
Each step, as the rest of this post walks it:
Write the ICP down where the AI can read it.
Research every account with the same prompt, so account 400 gets the pass account 1 got.
Score each account against the ICP, with a written reason, and sort by the score.
Write the first line from the research, so the opener is about that company.
Send it through an email and LinkedIn sequence, and read the replies in one place.
The chat-window version of this works, by the way. For fifty accounts, a chat window plus some patience produces research a rep can use, at no cost beyond a subscription you already pay for.
What breaks is the list, not any single answer. Research done one account at a time never turns into a view of all of them, so you finish with four hundred paragraphs and no answer to the question that matters on Monday: which twelve to work first, and why. That gap is why people and company search that arrives enriched and scored is the shape we built rather than another research box.
If you would rather see the ranked version of your own list, research and score your first accounts free.
How to use AI for prospecting: start from the ICP
Step one has no AI in it. Before a model researches anything, write down who you sell to: the industry, the company size, the roles that buy, and the signals that tend to show up before a deal.
A model with no ICP to judge against still produces confident paragraphs, and confident paragraphs about the wrong companies cost a rep a morning before anyone notices.
"Low reply rates drag your deliverability down, and poor targeting is what causes low reply rates."
Utku Zihnioglu, CEO of Oneprofile
In Oneprofile the description lives in your business context, one profile for the workspace with sections such as Ideal customer profile and Ideal buyer personas. The ICP scorer reads it on every run. With no business context it returns no score at all, on purpose: a score against a guessed ICP is worse than no score.
A quick test: hand the ICP to someone who has never seen your product and ask them to sort ten companies. If they can't, a model can't either.
ChatGPT for sales prospecting: where AI lead research breaks
ChatGPT for sales prospecting is the entry point almost everyone uses, and the ways it comes apart are consistent enough to list:
Nothing accumulates. The answer lives in a conversation rather than on the account record, so the next person to touch that account starts from zero.
Nothing re-runs. An account that raised a round in March is still described by the paragraph you wrote in January.
Answers are not comparable. Forty companies produce forty differently shaped paragraphs, because the prompt drifted and so did your attention.
Sourcing is uneven. Some claims come from a page the model just read. Others come from training data, which is where hallucination in AI systems stops being a research-paper concern and becomes a wrong headcount sitting in your CRM.
How to use AI for prospecting with account research at scale
Step two is where most of the tooling lives. Every serious tool in the category now ships a research agent, and they are good. You write a prompt once, point it at a domain or a profile, and it runs per account: a few searches, a few page reads, then a structured answer with the sources attached. Better ones say they found nothing instead of filling the gap with a plausible sentence.
Treat that as table stakes. It is also where the category has stopped. Run the agent over four hundred accounts and you have four hundred good, well-sourced write-ups, which is a real improvement and still not a ranking.
Stage | Unit of work | Where it stops |
|---|---|---|
Chat window and a spreadsheet | One account, by hand | Around 50 accounts, then drift |
Per-account research agent | One account, automated | Write-ups, not a score you can sort on |
Scoring against one ICP | Every account, one scale | Needs all the accounts in one place |
A small complaint about the genre, since we are here. Nearly every AI research demo runs on one beautiful account, usually a Series B company with a tidy website. Account 173 never gets demoed: a holding company with a one-page site and a Gmail address on the contact form. What a research pass does with account 173 is the only interesting question about it.
What feeds AI account research: data sources and your own systems
The quality of the answer is decided before any model runs, by what you put in front of it, and inputs come in three kinds.
Third-party data covers facts about strangers: headcount, industry, funding, a work email you can then verify. No single data source knows every company, so these get chained into a waterfall that tries data sources in order until one answers.
Web research covers what no database sells: whether the careers page lists three sales roles, or what the pricing page says now rather than eleven months ago.
Then there are your own systems, the half almost nobody researches. An account that pays you $740 a month, filed two tickets about API rate limits, and has nine of fifteen seats active is the strongest research subject on your list, and no data source sells any of those facts. A Stripe subscription record knows the plan and whether the last charge failed. Your support desk holds the complaint, in the customer's own words.
Oneprofile connects to 142 integrations that read data in, write it back, or both. Billing, support desks, product analytics, and your own Postgres database feed the research the same way a data source does, and the result goes back to the CRM as fields your reps can filter on.
How to use AI for prospecting on a whole list: score, then write
The question changes here. Not "what does this company do," which any agent answers well, but "which forty of these four hundred match the customers we keep."
A research agent can't answer that on its own. It hands back a paragraph, and four hundred paragraphs don't sort. Ask it for a number too and the number drifts with the prompt, so an 80 on account 200 and an 80 on account 3 are not the same 80.
Step three is scoring against your ICP, which holds every account to the same written description. It produces three things per account: a 0 to 100 score, a band (strong, possible, or poor), and a written reason. Sort by the score and the ranking falls out of it.
Account 173, with almost nothing on the record, comes back unscored rather than guessed at. This is where AI lead research across every account starts to pay, and the reason is the part reps care about: "84, paid plan since April, seats climbing" survives an argument.
Scoring has its own mechanics, covered in how to score accounts against your ICP and what lead scoring is.
Step four reads what the first three produced. Each sequence has an agent that writes every email and LinkedIn note from the contact's research, your business context, and the thread so far. You write what you're asking for in the step's prompt. The agent writes why you're reaching out to that person.
It opens each email on a different research angle, and each draft reads the ones before it, so a follow-up builds on the opener instead of repeating it. Skip the research and it still writes, but only from the contact's details and your business context.
Step five is email and LinkedIn outreach from the same list. Every reply lands in one inbox, and a reply stops that contact being emailed.
Where AI prospecting still needs a person
Three places, in practice:
Deciding who to call. The score sorts the list. It does not know that the founder of account 12 used to be your colleague.
Reading the first threads. The AI writing tab writes one contact's whole sequence and sends nothing. Mark it Looks good or Needs work, and a one-line reason gets the thread rewritten, with a learning proposed for the rest.
Approving what goes out. Switch on Needs approval before it sends on an email step and each message waits in Tasks, with the exact text that will go out, until someone approves it.
What AI prospecting costs per account
Per-token AI billing on a meter separate from data billing means nobody can predict the cost of a research run until it has finished. That is a strange thing to accept for a workload you want to run on your whole list every month.
Our prices, from the catalog as of October 15, 2026, so you can do the arithmetic yourself. Credits are $20 per 1,000 on Team and for extra credits on every paid plan, so a credit works out to $0.02.
Step | What you get per account | Price |
|---|---|---|
Company researcher, Medium effort | What they sell, who they sell to, their proof | 5 credits per account it answers |
Enrich person from email | The contact's profile | Up to 10 credits |
ICP scorer | A band, a 0 to 100 score, and the reasons | 3 credits |
Find work email | A work address, when you have none | Up to 10 credits |
Verify email | A deliverability check before you send | Up to 0.8 credits |
Research, a person profile, and a score on every account is up to 18 credits an account, when you already hold each contact's email. Finding and verifying the email first adds up to 10.8 credits. On 400 accounts that is up to 7,200 credits, or $144, and the estimate appears before the run rather than on the invoice.
A lookup that finds nothing costs nothing, so the person profiles usually cost less than that. The Free plan renews 200 credits a month, with up to 3 integrations, 3 tables, and 100 rows per table, which is enough to research 40 accounts before you decide anything. Team starts at $20 a month for 1,000 credits.
Whether frontier model prices keep falling at the rate they have, we genuinely don't know, and anyone who tells you they've modelled their 2027 AI research budget is guessing. What we'd rather not do is pass that uncertainty to you as a variable bill.
Anyway. If you want to test the premise without touching a tool, take twenty accounts you lost last quarter and write down the five facts that, in hindsight, predicted the loss. Count how many of those five your CRM can currently see. That count is the real limit on any AI lead research you run, whichever product does the running.
Then, if you want the version that doesn't stop at one account: Oneprofile researches every account from data sources and your own systems, scores each against your ICP with a written reason, and writes the first line from that research, on one credit meter that shows the price first. Research and rank your own list free.
