Claude for lead generation: what it does, and what it will not do

You have a Claude subscription and ten hours a week for outbound. Somewhere between those two facts is the question of whether you still need to pay a data vendor.

The answer is in Anthropic's own documentation, and it is more useful than anyone's opinion about it. Anthropic publishes what each surface reaches, what it cannot reach, and where the accuracy is your problem rather than theirs. Most of what gets written about Claude for lead generation is somebody's workflow. What follows is what the vendor says, with the page it says it on.

What Claude can actually reach

Three surfaces matter for this job, and they are not the same thing.

Web search runs inside a conversation. Claude decides when to search and can run several searches for one question. Anthropic's documentation says a simple query typically uses one to three searches and comparative research can use ten or more. Results come back with citations.

Research is the agentic version. Anthropic's support page describes it running multiple searches that build on each other, drawing on web search and on Google Workspace. It is a paid-plan feature, and the same page warns that research sessions "can use up your limits faster due to Claude retrieving multiple sources."

Connectors are the interesting one for anybody with data sitting in tools. Through the Model Context Protocol, Anthropic documents that Claude can read, create, modify and delete data in connected apps and take actions on your behalf. Available across Free through Enterprise, on the web, desktop, Cowork and mobile.

That last line comes with a cap most people find at the worst moment. On the free plan you get one custom connector. One.

Cowork, and what it does with a file

Cowork is where the actual work happens if your leads live in spreadsheets, and it is close to unwritten territory in the coverage of this topic.

Anthropic describes it as agentic, multi-step work on knowledge tasks rather than code. With permission it reads and writes local files, generates Excel workbooks with working formulas, formatting and multiple tabs, creates documents, and organises files into folders. It cannot delete a file without you saying so.

For a list you already have, that is a real capability. Deduplicating four exports, normalising job titles across them, splitting a 2,000-row file into segments with a written rationale per segment: this is the kind of work Cowork does well, and doing it by hand is where most of the ten hours a week goes.

Concretely, the shape of the instruction matters more than the tool. Connect the folder the exports are in, then ask for the whole job at once rather than step by step: merge these four files on email address, keep the most recent row where they conflict, collapse job titles into five buckets and tell me the rule you used, drop anyone whose company has under twenty staff, and give me back one file per bucket plus a note on what you threw away and why. The last clause is the one people leave out, and it is the one that lets you check the work.

Three limits, all from Anthropic's own pages. Cowork is paid plans only, so Pro at minimum. Cloud sessions run in isolated environments that cannot reach your home or company network directly, and local file access exists only while the desktop app is open with the relevant folders connected. And Anthropic states that working on tasks in Cowork consumes more of your usage allocation than chatting with Claude, because multi-step tasks are compute-intensive and need more tokens to run. That is the sentence that explains why your limits vanish faster than you expected.

Anthropic is also blunt about supervision in a way that is worth quoting rather than paraphrasing: "Human oversight is recommended in high stakes scenarios since Claude can sometimes make mistakes." On the risk of instructions hidden in content Claude reads, the same page says that despite the safety measures in place, "the chances of an attack are still non-zero."

Connectors and MCP: what Claude reaches, and what it will not

The MCP documentation contains one requirement that quietly rules out a lot of setups. Remote MCP servers must be reachable over the public internet from Anthropic's cloud. Private networks, VPNs and firewalled servers will not connect. At the API level, only public HTTPS servers work, and local STDIO servers cannot be connected through the API connector at all.

If your CRM sits behind a VPN, this is the end of that idea.

Two more things from the same pages, both worth knowing before you connect anything sensitive. MCP tool-use data is not covered by Zero Data Retention and is retained under the standard policy. Agent Skills are not covered by it either.

On Google Workspace, where a lot of small-team lead data actually lives, the connector documentation is specific about the edges. Claude can search and read Gmail, but for attachments it gets metadata only. It can search and retrieve Drive files including Sheets, but it "extracts text content only," so anything living in an embedded image does not reach it. Sending mail and sharing files require your explicit approval by default.

Skills, and the security line Anthropic writes itself

Agent Skills are packaged instructions Claude loads when they are relevant. Anthropic's overview describes a three-tier loading system: about 100 tokens of metadata always in context, the SKILL.md file under 5,000 tokens loaded when triggered, and bundled resources pulled in on demand.

Two constraints shape what a lead-generation skill can be. Custom Skills do not sync across surfaces, so a skill you upload to claude.ai is not available to the API and a Claude Code skill is a separate thing again. And on the API, Skills execution has no network access and no runtime package installation, which rules out a skill that calls an enrichment service.

Then there is the sentence Anthropic writes about installing other people's skills, which is the most sober thing on any of these pages. An untrusted skill "can direct Claude to invoke tools or execute code in harmful ways," so you should "treat Skills like installing software." The lead-generation skill repositories on GitHub are worth reading. They are also worth reading the way you would read a shell script somebody sent you.

What a CSV gets you

CSV is a supported upload. Anthropic's file documentation allows up to 500MB per file in a conversation with a maximum of 20 files, and 30MB per file in projects.

The constraint that decides everything is five words long: for documents other than PDFs, including CSV, "Claude extracts text only."

So the shape of a workable process is clear enough. Bring the list. Claude reads it, reasons over it, segments it, writes the messages, and hands the file back. The context windows are large enough that this is not the bottleneck: Opus 5 and Sonnet 5 carry one million tokens, with Sonnet 4.5 and Haiku 4.5 at 200,000. If your input exceeds the window the API returns an error rather than silently truncating, which is the behaviour you want.

What none of this does is produce the list

Here is the boundary, and it is not a criticism of the product. It is what the documentation says.

Anthropic's web search documentation states no accuracy guarantee and makes no hallucination-prevention claim for the tool. Its support page for web search tells you plainly to "cross-reference cited sources" and to "use authoritative sources for critical decisions." Verification is yours.

For lead generation that lands in two specific places.

Enumeration. Asking a model to list every company matching a set of criteria is a request for a complete set, and nobody has published a study of that exact task. The nearest academic work is not close enough to lean on: QAMPARI, a benchmark built specifically around questions with many correct answers, found the best retrieve-and-read systems reached an F1 of 32.8, but that measures 2022-era retrieval pipelines rather than a current model listing companies, and it should not be quoted as if it did. Separately, the HalluLens benchmark published at ACL 2025 includes a task specifically testing whether models fabricate details about entities that do not exist rather than refusing.

Treat a list of companies from a model as a starting hypothesis with unknown completeness.

Contact data. There is no credible published measurement comparing a model's email or phone output against a data provider's. We looked for one. What exists is vendor marketing and unrelated academic work on detecting AI-written email. So the honest statement is that the accuracy is unmeasured, and anyone quoting you a percentage here made it up. On why a verified-looking address still fails, our piece on catch-all domains covers the mechanism, and what list decay costs month by month covers the part that goes stale while you work.

Two claims in circulation are worth naming so you do not have to chase them. A widely shared Medium post promising 90% open rates from a Claude-based process reports a range of 75 to 90% with no sample size and no method stated, and names Perplexity and Clay as paid parts of the workflow. And a post titled around generating unlimited leads for free says in its own body that the framing is "a headline, not a literal promise," then lists Claude Pro or Max at minimum, plus Clay, Apollo or ZoomInfo, an automation tool and an email verifier as what it actually takes.

What it costs

Anthropic publishes plan prices and is deliberately vague about volume, which is itself a planning fact.

Pro is $17 a month billed annually or $20 monthly, described as at least five times more usage per five-hour session than Free. Max starts at $100 a month. Team is $20 per seat annually, $25 monthly.

Usage is where the estimating gets hard. Anthropic's pricing page states that every plan has limits that reset on a rolling five-hour session window, with weekly limits on top of that for paid plans. It does not publish a fixed message count; the support documentation says usage depends on the length and complexity of your conversations, the features you use, the model you chat with and the effort level you select. Add Cowork's heavier draw on top.

If you go the API route instead, web search is priced separately at $10 per 1,000 searches plus token costs. One search call counts as one use regardless of how many results come back, and failed searches are not billed.

What to do with this

If you have a list and cannot get through it, Claude is a genuine answer and Cowork is the part to learn. Connect the folder, bring the file, and get the segmentation and the messages done in an afternoon rather than a fortnight.

If you do not have a list, no amount of prompting produces one you can trust, and Anthropic's own documentation is the reason to believe that rather than our opinion. Finding the companies, and knowing whether the person named against one still works there, is a data problem. It is the one we work on: the accuracy claims data vendors publish about themselves is the right place to start if you are about to buy any of it. If those records already sit in Salesforce, what Salesforce and Anthropic actually shipped in August covers what Claude can now read there, and what it still cannot put in.

This piece describes Anthropic's published documentation as of 1 September 2026.