Chapter 1 of 7
Layer 1 · Data: Who do you want to reach?
Claude knows nothing about your customers until you tell it. So the start is not a prompt, but the question of who fits you.
Prompt to try
Here are our ten best customers with industry, size and region: [list]. What do at least seven of them have in common? Turn that into an ideal customer profile in three sentences and name two attributes you are unsure about.
Next, it matters where your data lives. In a Salesforce vendor study, 44 percent of German sales organizations say disconnected systems slow down their AI projects.
Do it today
Write down your ten best customers and try the prompt above.
Chapter 2 of 7
Layer 2 · Thinking: Claude with skills and MCP
This is where Claude works. There are four ways to use it. Each builds on the one before, and not every team needs all of them.
| Way | What happens | Example |
|---|---|---|
| Chat | You ask, Claude answers. | Summarize a company before the meeting. |
| Skill | Claude follows a written-down process. | Every research brief has the same structure. |
| MCP | Claude works directly in your software. | Find matching leads and create a list. |
| Routine | Claude handles tasks at set times. | Every morning: new replies, sorted. |
What is a skill?
A written work instruction for Claude, like the onboarding folder for a new colleague. Claude follows it the same way every time.
Example: the skill "Research before the first call"
- When
- "Prepare me for the meeting with company X."
- Steps
- Read the website and news, name three reasons to reach out, suggest two questions.
- Limits
- No figures without a source.
- Result
- One page, read in five minutes.
What is an MCP server?
An adapter that lets Claude work inside another application: search leads, create campaigns, read replies. The server decides what is allowed, not the AI.
Step 1
You ask
"Find twenty machine builders in Bavaria with a new head of sales."
Step 2
Claude plans
A skill sets out the steps.
Step 3
MCP server
Checks permissions, runs only what is allowed.
Step 4
Platform
Companies, contacts, campaigns, inbox.
Step 5
Result
A list, ready for approval.
Simplified. The MCP server decides which actions are allowed.

Key takeaway
A skill tells Claude how something is done. An MCP server lets Claude do it.
Chapter 3 of 7
Layer 3 · Channels: Why emails do not fail because of the text
Claude writes an email in seconds. Whether it arrives depends on domains, settings and limits.
If you cut corners here, you risk your own company domain. Then even proposals to existing customers land in spam.
| Rule | Yahoo | Microsoft | |
|---|---|---|---|
| SPF and DKIM | required | required | required |
| DMARC | at least p=none | at least p=none | at least p=none |
| One-click unsubscribe | required | required | recommended |
| Complaint rate | below 0.3% | below 0.3% | no number |
- 1
Separate sending domains
Never run cold outreach on your main domain.
- 2
Warm-up
New mailboxes start small. This takes weeks.
- 3
Clean addresses
Verify before sending. Every bounce hurts.
- 4
Limit volume
More volume means more mailboxes, not more emails per mailbox.
- 5
Monitor
Keep measuring complaints and spam placement.
Interactive checklist
Are you ready to send?
Tick off what you have in place.
What does that mean?
If a sending domain loses its reputation, your main domain stays clean.
What does that mean?
Without them, mailbox providers cannot check that the email comes from you.
What does that mean?
Google, Yahoo and Microsoft require at least p=none.
What does that mean?
A fresh mailbox sending lots of emails looks like spam.
What does that mean?
More volume through more mailboxes, not more emails per mailbox.
What does that mean?
Hard bounces show that the list is out of date.
What does that mean?
It also lowers the number of spam reports.
What does that mean?
Google names 0.3 percent as the upper limit, better below 0.1 percent.
What does that mean?
In Germany, the UWG and the GDPR apply. This item is not legal advice.
Basics first.
Before Claude writes the first email, the technology has to be in place.
Claude alone is not enough here
Domains, warm-up, limits and monitoring need a system that runs around the clock. In Germany, promotional emails fall under the UWG and the GDPR; clarify this with your legal advisor.
Chapter 4 of 7
Layer 4 · The person: Why it rarely fails because of the technology
The tools are there. What holds teams back is uncertainty, missing know-how and habit.
What holds companies back is law, cost and acceptance, not the technology.
Hurdles for companies that use AI · Multiple answers, 2026
- Data protection66 %
- Legal uncertainty56 %
- Costs hard to predict54 %
- Lack of team acceptance42 %
- Poor results26 %
Source: Bitkom press release of 14 Sep 2026 (603 companies with 20+ employees)
Companies say they train. Little of it reaches the people.
AI training from two perspectives · 2026
- Companies: "We train on AI"70 %
- Employees: "I was trained"21 %
- Employees: "There is no offer"37 %
Sources: Bitkom, 14 Sep 2026 and 28 Apr 2026
Our view: the salesperson becomes a conductor. They do not have to play every instrument, but they need to know when each one comes in. That takes good questions, analytical thinking and the will to keep learning.
- 1
Start with one task
Not "introduce AI" but: two weeks of call preparation with Claude.
- 2
Capture your best people
Write down your most successful colleague's process as a skill.
- 3
Train for real
Short and based on your own cases. The EU AI Act requires measures for AI literacy.
- 4
Talk about roles openly
What goes away, what becomes more important? Silence creates resistance.
Your question
If Claude took over all research tomorrow, what would your team do with the time?
Chapter 5 of 7
Layer 5 · Learning: measure and build routines
What works only shows in the numbers. Claude can read them for you every week.
- Every morning: sort new replies by interest, question and rejection.
- After every call: pull tasks and objections out of the transcript.
- Every week: which audience replies, which reason to reach out works?
The capabilities are here. How many routines run at once is, in the end, a question of computing power.
Chapter 6 of 7
Where sales automation comes in
Claude is strong at thinking. A sales automation platform is strong at doing: data, sending, inbox. The two work together through its MCP server.
| Task | Claude | Platform | Person |
|---|---|---|---|
| Target accounts | suggests a profile | provides companies and contacts | decides |
| Research | summarizes | provides signals | checks |
| Messages | writes drafts | sends safely | approves |
| Replies | sorts | collects all channels | holds the conversation |
| Learning | analyzes | measures | draws conclusions |

Strategy call
What would this look like for you?
We show you, using your own case, how data, Claude and sending work together. 30 minutes, no strings attached.
Chapter 7 of 7
Sources
All figures in this guide
- Haufe Akademie and HdWM, study "AI in Sales 2026" (542 sales professionals)
- Bitkom press release of 14 Sep 2026 (603 companies with 20+ employees)
- Bitkom press release of 28 Apr 2026 (1,003 people aged 16 and over)
- Salesforce, State of Sales, press release of 4 Feb 2026 (vendor study)
- Google, email sender guidelines
- Yahoo, sender best practices
- Microsoft, Outlook.com requirements for high-volume senders (2 Apr 2025)
- EU AI Act, Article 4 AI literacy (consolidated version)
- Model Context Protocol, official documentation

About the author
Otto Lang
Co-Founder and Managing Director, Sales Automation GmbH
Otto Lang is co-founder and managing director of Sales Automation GmbH in Munich. He sold his first B2B sales automation company in 2022 and founded Sales Automation GmbH with Anton Lang in 2024.
He builds the platform this guide is based on and is available to customers in day-to-day projects. The examples and steps come from working with mid-sized B2B companies.