SALES
Pipeline you can plan around
Outbound, inbound handoff and expansion, run as one motion.
GTM BrainFOR SALES
SEND 01SEND 05
It gets sharper with every reply and every call it reads.
CLAUDE FOR GTMPart of: Claude for GTM: how GTM teams run on Claude

Prospecting with Claude: what it does well, and where it fails on its own

The five prospecting jobs we hand Claude on every account, the three things it cannot do without help, and the order a real run follows.

Dark header with the theme Finding the right people above white tiles: the Claude logo, a people icon, and the Supabase and Prospeo logos.

Prospecting with Claude works best on the thinking around a list. Claude turns a buyer description into an ICP, cleans any list, researches each account, scores it with a reason and writes from the record. It holds no contact data. Every verified email and phone number comes from a data provider it calls, behind a cost gate a person approves.

We run this on every client account, as one part of a wider Claude for GTM system. Below are the five jobs Claude does well, the three places it breaks alone, and the order a real run follows.

What Claude does well in prospecting

Each of these jobs needs judgement applied the same way across hundreds of rows. That is where Claude earns its place. None of them needs Claude to know a fact it was never given.

The five prospecting jobs we hand Claude
JobWhat Claude doesWhat it works from
Define the ICPTurns a plain description of the buyer into titles, industries, headcount, stack and buying criteriaThe client's offer, proof and sales history
Normalise the listReads any export or CSV and maps it into one structureWhatever file arrives
Research each accountBuilds a snapshot of the company, the people and what changed recentlyThe domain, the website, LinkedIn and news
Score with a reasonApplies a locked model and writes one line on why each account landed in its tierThe scoring model and the enriched record
Write from the recordDrafts copy from what is true about that accountThe enriched record and the client's proof

Defining the ICP from a description

A founder can describe the buyer in two sentences. A data provider needs that turned into filters it accepts. Claude does the translation: titles, seniority, industries, company size, geography, technology and the criteria that make an account a fit.

When the brief leaves a gap, our build asks about three things only: titles, seniority and geography. Those decide who comes back, and a guess there wastes credits. Everything else gets a sensible default, shown on screen so a person can correct it.

The title list matters most, because it becomes the allowlist every lead is checked against. For a market outside the US and UK, Claude also translates the titles into the local language. A person scans those translations before the run, since a wrong one is a buyer the list quietly misses.

Normalising any list into one shape

Lists arrive in every format. An Apollo export, a client's CSV and a BlitzAPI pull each come with different columns. When we pushed lists into Clay for email finding, each one had to be reformatted to the exact column names first.

Claude reads whatever it is given and maps it into one structure. Three sources with three layouts make no difference to the next step. That one change removed a manual job from the start of every build.

Researching each account

Research is where Claude saves the most time on named accounts. The Prospect Researcher works outward from a domain: the company first, then the people in the relevant seats, then what moved recently.

We keep that order fixed. A hiring post means little until you know what the company sells and who would own the problem. The snapshot ends with the few angles a rep could open with, or says plainly that nothing is happening at the account. Our prospect research agent guide walks through the full setup.

Scoring with a reason line

The Account Scorer applies a model a person has already agreed (how to build an account scoring model covers that step). Disqualifiers run first. Fit and timing are scored apart. A strong fit with no trigger and an average fit with an urgent one need different handling.

Every account gets one line explaining its tier. When the data cannot fill a dimension, the line says so, and that dimension is left out of the score. The rep can then decide whether they agree. Here is an example of the shape:

An example reason line
Tier B. Strong fit on size, industry and stack. No hiring or funding data found, so timing was not scored.

Writing from the record

Once the list is clean, Clay layers hiring signals, firmographics and job changes onto each record inside outbound. Claude then reads that record and writes from what is true about the account. Each email starts from real context, and a template full of variables never enters it.

The same rule covers proof. Claude may only use the client's own cleared proof, read from their files. If a claim is missing from the record, it stays out of the email.

Where prospecting with Claude breaks on its own

It looks finished after the first run, which is where the demos stop. The failures show up later, at volume, on a real budget and with a second operator. We hit each of these three before we built around them.

It holds no contact data

Claude can guess that an address follows first.last at a domain. A guess still has to pass verification before it counts as an email. The same holds for phone numbers, and a model has no record of anyone's mobile.

So sourcing runs through data providers that Claude calls, in a fixed order. Leads we already hold in Supabase come first, because they cost nothing to reuse. BlitzAPI comes next on a flat rate. Metered providers only see what the cheaper sources missed.

Prospeo sits at the end of that order. It charges on the reveal, so a miss costs nothing, and what it returns is already verified. Its pricing page lists a verified email at 1 credit and a mobile number at 10. The Starter plan is $49 a month for 2,000 credits (checked 22 Oct 2026). Apify covers sources no database holds, such as Google Maps, and it bills for every row it returns.

Every address then goes through a verification waterfall. DeBounce checks the addresses already on file. Catch-all domains get a second check through BounceBan, which tests the mailbox without sending anything. A person with no address that passes comes back marked as exactly that.

It needs a cost gate

Connected to metered APIs with no gate, Claude spends whatever the task seems to need. On a large list that is a real bill, so nothing in our build spends before a person sees a number.

The Lead Sourcing Waterfall runs a preview first. It shows how many matching people we already own and the real match count from the first provider. It also shows the order credits will be spent in. If the match count is huge, the ICP is too broad, and we tighten the titles or geography before any spend.

A count still says nothing about quality. Keyword filters drift, so the List Build Run has a person read a sample of real rows before the full build is approved. On top of that, each provider has its own spend ceiling, and a larger job needs a second press.

Volume brought a failure we did not see coming. Our waterfall worked at 20 leads and came back worse at 500, because every provider call fired at once. The fix was pacing each request, which Claude Code for GTM covers in the section on loops. One more lesson cost us real credits: API keys belong in a vault the build reads from, never in the chat.

It needs the ICP from a file

Typed into a chat, the ICP shifts a little every time someone types it. One run asks for heads of operations, the next adds directors, and the lists stop being comparable.

We keep each client's ICP in a version-controlled file in their repository. Every build reads that file before it starts, and a change to the ICP is a dated edit with a reason beside it. Two operators running the same brief a month apart get the same people back.

Finding the right people at named accounts

A question we hear on sales calls is who to contact at a known list of targets. A supplier to airports put it simply: they had their list of target airports and needed to know who to contact at each one.

That calls for an account-first build. Each target company is resolved first, then a tiered title list finds the best contacts inside it. We always ask how many contacts per account to take, and the default is 5. A higher cap changes how deep we go into each account, while the title list still decides who qualifies.

A persona list across a whole market runs the other way: people first, filtered by company. Claude reads the ICP and picks the path, and a person confirms it before anything runs.

How a prospecting run fits together

This is the order a run follows on our accounts. Each stop exists because skipping it once cost us credits or a bad list.

  1. 01Read the ICP fileClaude loads the client's ICP and turns it into provider filters, with every default shown.
  2. 02Confirm titles, seniority and geographyA person approves the title list, since it decides both what is bought and what is kept.
  3. 03Preview with real countsOwned leads, match counts and the spend order come back before any credit moves.
  4. 04Approve the spendA person signs off the number. Larger jobs need a second confirmation.
  5. 05Source in orderOwned data first, then the flat-rate source, then the metered providers on the remainder.
  6. 06Verify every addressNothing reaches a sequence without passing verification, including catch-all domains.
  7. 07Research and scoreEach account gets a snapshot, a tier and one line explaining the tier.
  8. 08Write from the record, as draftsCopy lands as drafts, and a person approves every send.

It works when this order is kept. Claude does the reading, sorting and writing. The providers supply the data, and a person owns the spend and the send. If your team builds lists in Clay, our guide to cutting SDR prospecting time with Clay and AI covers the same flow there.

Where this sits in a team's wider use of AI is covered in the five levels of AI in GTM. Prospecting is usually the first job that moves from a chat tab into a system with memory. Each finished list then feeds the layers of cold email personalisation on top of it. The research snapshot later feeds call prep with Claude, once a reply becomes a meeting.

We build and run this sourcing system for clients, on their ICP and their budget.

Want this built for your team?One call with a founder. We tell you which parts apply to your GTM, and what we would build first.Request a call
Why is prospecting so hard?

It asks for three things at once: the right account, the right person inside it and a reason to write now. Each comes from a different source, and each goes stale at its own pace. A list that was accurate in spring can bounce in autumn, so the sourcing, verifying and timing all have to be repeated.

What is the best route to reach the right person at a target account?

Start from the account. Resolve each target company first, then search inside it with a tiered title list, so the most senior relevant buyer comes back first. Cap how many contacts you take per account, and verify every address before it is used.

Where does the data come from? Do you buy it?

Contact data comes from data providers, called in order of cost. Leads already owned for that client are reused first. A flat-rate source covers the bulk, and metered providers such as Prospeo fill what is left, charging per verified result. Every address is verified again before it reaches a campaign.

Which AI is best for prospecting?

For the reasoning work we use Claude: turning a brief into filters, cleaning lists, researching accounts, scoring and drafting copy. The contact data itself comes from data providers whichever model you choose, so pick those providers with the same care as the model.

FURTHER READING
David Turewicz
WRITTEN BYDavid TurewiczCo-founder and CEO, Kinetyca

Software, accelerator and startup background. Runs the GTM architecture side: how the system is designed, what it is built on, and where AI actually earns its place.

LinkedIn ↗YouTube ↗
KINETYCA GTM NEWSLETTERGet this a week early.Everything we publish here goes to newsletter readers first, along with the cases and numbers that never make it onto the site. Tuesdays and Thursdays, written by David and Matteo.
TWICE A WEEK · NO SPAM · ONE-CLICK UNSUBSCRIBE
Kinetyca GTM NewsletterNews, cases and what is actually working in GTM right now.
© 2026 KinetycaPrivacyTermsImprint
Follow the foundersDavidCEOMatteoCOO