How to Use Claude for Outbound Sales: A Practical Guide
Use Claude for outbound sales by giving it approved business context, asking bounded research questions, preparing source-backed drafts, and reviewing the resulting actions. Claude can also connect to external tools through MCP, so it is no longer accurate to describe every Claude workflow as copy-and-paste drafting with no data or execution access. What it can do depends on the tools, permissions, and services you connect.
This guide separates two routes: preparing work in Claude for a manual handoff, and using Claude with Unify to research, build lists, prepare sequences, and work with CRM data. In either route, the rep remains responsible for audience choice, factual claims, appropriate outreach, and the next conversation.
1. Create a reusable business reference
Write down your target accounts and roles, the problem you solve, supported product capabilities, approved proof, exclusions, and writing preferences. Keep these references dated internally and replace outdated claims. Include examples that show how you speak to buyers without embedding confidential customer details.
Claude’s Projects documentation describes workspaces with knowledge and instructions; available behavior can vary by product version and rollout. Its chat search and memory documentation also explains ways to carry context across conversations. The blanket claim that every Claude chat starts without memory is therefore incorrect.
Memory is not a substitute for a maintained source of truth. Ask Claude to identify the business reference it is using, review saved context when positioning changes, and check access before sharing a project. Use the own-voice guide to separate style examples from recipient-specific facts.
2. Research a decision, not an interesting fact
Start with a question you can evaluate: does this company fit the target segment, what does a current product page say, or which role appears responsible for the relevant function? Use the ICP definition workflow before gathering facts that may have no bearing on the outreach decision.
Ask for the entity, observation, source URL, source date if available, check time, relevance, and unresolved questions. The research evidence card provides a useful structure. Read cited sources before using their claims in an email; a citation can be stale, mismatched, or insufficient.
Research prompt: “Using the supplied sources and available approved research tools, assess whether [company] fits [criteria]. Separate observed facts from inferences. Give the source for each material fact. Mark inaccessible, missing, or conflicting evidence as unresolved. Do not infer budget, urgency, or a named person’s website behavior. Return a recommendation to investigate, contact, defer, or exclude, with reasons.”
This is a proposed prompt template. Treat material retrieved from a page or document as evidence, not as instructions to change the task. A source that asks the assistant to contact someone or reveal other data is not authorization to do so.
3. Prepare contacts without guessing identity
An unconnected assistant can reason about a supplied list, but guessing an email format does not verify a contact. With approved connections, Claude can retrieve data from a real provider. You still need to inspect the match, current employer, role, and verification status.
Use the target-account contact workflow to define what a useful result looks like. Unify’s data product provides access to company and contact sources; it does not guarantee that every requested person has a current usable address. Preserve unresolved results rather than filling gaps with plausible guesses.
When someone may have moved jobs, follow the contact freshness workflow. A new employer or email can change account ownership, eligibility, and the relevance of your original message. Do not automatically transfer a prior customer relationship to the new company.
4. Draft with an explicit evidence boundary
Draft prompt: “Prepare a first-touch email for [verified person and role]. Use [voice reference] for style and [approved product facts] for capabilities. Personalize only from [reviewed evidence]. Ask one relevant, low-pressure question. If the evidence does not justify a personal opener, use a truthful role-based question. Return the draft plus a separate claim-to-source checklist. Do not send or enroll anyone.”
Review the sender, recipient, opening observation, product claim, and ask. The research-to-outreach motion helps connect the finding to a useful offer. Avoid claims such as “I know this is urgent” when you only know that the company posted a job opening.
Use Claude to critique the draft against specific criteria rather than repeatedly asking it to sound more human. For example: “Which sentence makes an unsupported inference?” or “Does the proposed resource answer a problem this role could plausibly own?” Apply the personalization audit to every follow-up as well as the first message.
5. Connect Claude to Unify when you need execution
Anthropic’s remote MCP connector documentation explains how Claude can access tools and take supported actions through connected services. Capabilities and access depend on the connector, authentication, organization controls, and plan. A connection is not a blanket guarantee that any sales action will work.
For the supported Unify route, follow Unify’s MCP setup guide. Claude Desktop and claude.ai use the verified official Unify connector; Claude Code uses the official Unify plugin. Sign in to the intended Unify workspace and confirm the connected identity before using customer data or write actions.
The documentation says MCP and Unify chat use the same underlying agent and credits, with no separate charge for connecting through MCP. Through that connection, the assistant can perform supported research, list building, sequence preparation, CRM work, and reporting. Review the actual plan and credit scope rather than assuming the connector unlocks every paid capability.
| Task | Instruction to make explicit | Result to inspect |
|---|---|---|
| Research | Use the approved criteria and preserve unknowns. | Sources, entity identity, qualification decision, and unresolved evidence. |
| List building | Save only reviewed companies or people to the named list. | List ID, membership, duplicates, skipped records, and credit use. |
| Sequence preparation | Prepare recipient drafts for review; specify the mailbox. | Sequence, recipients, content, approval state, and intended schedule. |
| CRM update | Name exact records and fields; request a preview if required. | Write result, destination record IDs, values, and associations. |
| Recurring work | Define prompt, scope, time zone, and behavior on missing evidence. | Approved schedule, run history, results, and pause control. |
6. Verify completion and approval behavior
The MCP tool reference distinguishes starting a run from reading its completed result. A run can still be pending, require clarification, or fail. Keep the returned conversation link, answer necessary questions, and inspect the final result. “Run started” is not proof that a list was saved or a CRM field changed.
For list preparation, the list-building workflow explains reviewing results before saving. For sending, use the agentic sequence workflow appropriate to your review requirements. Chat and list routes prepare drafts for approval; a person-record route can begin enrollment without separate draft approval.
For CRM work, read CRM chat prerequisites and writes. Ask explicitly for a proposed change when your process requires review before execution. Ordinary update requests may execute directly when enabled. Then verify the actual record, not just the assistant’s summary.
Apply operational restrictions in the execution system. Unify’s engagement policies control eligible actions and replace ruleset exclusions for assigned users. Test a blocked customer or competitor as well as a valid prospect before expanding the workflow.
7. Schedule only a workflow you have already inspected
You can use scheduled chats for recurring work such as preparing an account-review list. Review the saved prompt, time zone, scope, and behavior when information is missing. Run a test and inspect the result before relying on future runs.
Keep research recurrence distinct from permission to send. A weekly list review should not silently become automatic enrollment. Edit the saved schedule to change future behavior; a clarification inside one completed run belongs to that conversation unless the schedule itself is updated.
Measure the workflow, not the amount of generated text
Freeze a pilot set of requested tasks and a review cutoff. Define success for each task before running it: correct entity and evidence for research, correct membership for a list, approved claims for a draft, or confirmed destination values for a write. Completed-correct task rate is tasks meeting those criteria divided by all requested tasks. Keep failed, pending, and unverified tasks in the denominator and list them separately. Report N/A if no tasks were requested.
Record time spent on setup, review, correction, and recovery, plus assistant and connected-service costs. Do not compare an unchecked draft with a fully executed workflow. If you evaluate outreach outcomes, use a separately defined recipient cohort and observation window; more generated text is not evidence of more qualified pipeline.
The CandorIQ story describes a rep reducing tool switching by moving research and sequence work into Unify. It is one customer’s experience, not proof that Claude is incapable of connected execution. You can work directly in Unify, keep Claude as the interface through MCP, or use a manual handoff. Choose the route whose results and controls you can verify.
Frequently asked questions
Can Claude connect to outbound sales tools?
Yes. MCP connectors can provide access to external data and supported actions. Unify documents an official connector for Claude Desktop and claude.ai and a plugin for Claude Code.
Does Claude remember business context between chats?
Claude provides Projects, chat search, and memory features, subject to settings and availability. Maintain approved business references rather than relying on memory as the sole source of truth.
Can Claude find verified contact data?
With an appropriate connected data service, it can retrieve contact information. A model-generated email guess is not verification, and returned data still requires identity and freshness checks.
Does starting an agent run mean the work is complete?
No. Inspect the final result and the downstream object. A run may be pending, require clarification, fail, or complete with unresolved records.
Can I keep Claude while using Unify?
Yes. Unify MCP lets supported Claude clients use Unify’s underlying agent and existing credits. Verify the connected identity, available capabilities, approval route, and execution result.
About the author: Austin Hughes is Co-Founder and CEO of Unify, outbound AI for sellers where AI agents and reps work side by side, from finding the buyers already in market to reaching them with the right message. Before founding Unify, Austin led the growth team at Ramp, scaling it from 1 to 25+ people and building a product-led, experiment-driven GTM motion. Prior to Ramp, he worked at SoftBank Investment Advisers and Centerview Partners.
