Finding a prospect meant opening websites one by one, judging by feel, writing an email from scratch, creating the CRM record, chasing the payment and finally typing the invoice into sevDesk.
SUCCESS CASE 01 · OUR OWN SALES PROCESS
From a website finding to a paid appointment.
We built this for our own business: discover websites, investigate real customer tasks, and turn supported findings into a tailored offer, CRM records and an Outlook draft. A person reviews the offer before sending. The next part connects payment, booking and a sevDesk invoice draft, verified in a separate sandbox test.
CRM & Outlook drafts verified live · Payment & invoice drafts verified in sandbox
OUR OWN BUSINESS · HOW IT CONNECTS
The workflow at a glance
- DiscoverLocal businesses with a website
- InvestigateSaved evidence from real customer tasks
- QualifyA supported finding and a feasible fix
No supported finding? No offer. Uncertain evidence? Human review.
Ready for your review
- CRM opportunity
- Personal offer page
- Outlook draft
NEXT PART · VERIFIED IN A SEPARATE SANDBOX TEST
Confirmed Stripe payment- sevDesk invoice draft
- Customer-confirmed Teams booking
No real payment or invoice sending. Automatic invoice drafts apply to the tested pricing path.
- websites reviewed
- 4
- qualified preview offer
- 1
- extra drafts on publication replay
- 0
- separate sandbox payment test
- 1
THE STARTING POINT
Prospecting by hand does not scale, and guessing is not fair to anyone.
A discovered business is investigated with saved evidence, qualified against fixed rules, and prepared as a CRM opportunity, a personal offer page and an email draft. In the separate sandbox payment test, an invoice draft is prepared and the customer chooses and confirms an available appointment.
Our own process, verified on 18 September 2026. Every offer is still read by a person before it is sent.
01 / THE FLOW
From discovery to a draft. Then payment and booking.
n8n orchestrates. A small set of private services does the durable work. The person stays where judgement matters.
Discover
Local businesses are collected from OpenStreetMap by city and area and stored once, without duplicates.
N8N · OPENSTREETMAP · POSTGRESInvestigate
An AI agent with a fixed set of tools reads the website the way a customer would: can I find the opening hours, the menu, a phone number? Findings must cite saved observations; a person still checks the conclusions.
CLAUDE · SCOPED TOOLS · ISOLATED BROWSERQualify
A rule-based gate decides. It needs a verified finding, a real failed customer task, a feasible fix and a published contact email. A clean website or insufficient evidence produces no offer.
N8N · POSTGRESPrepare
For qualified cases only: German offer copy, a priced offer with a private link, a company, opportunity and note in the CRM, and an Outlook draft. Every write is checkpointed.
TWENTY CRM · MICROSOFT 365 · OFFER PAGEReview and send
A person reads the draft, edits it where needed and sends it. The workflow never sends mail on its own.
OUTLOOKPay and book
The customer pays on the offer page. In the tested sandbox path, Stripe’s signed webhook confirms the amount paid. sevDesk receives a test invoice draft, and the customer selects and confirms a Teams slot from our live calendar.
STRIPE · SEVDESK · MICROSOFT TEAMS02 / WHAT IT TOOK
Getting to a flow we trust.
The first versions produced confident text with thin evidence. Most of the effort went into making the process honest and safe to rerun.
Evidence before conclusions
The investigator must cite saved observations. Submissions without evidence, or with “coverage complete” after a failed fetch, are rejected and sent back for another attempt.
Rejections are results
A healthy website produces no offer. A finding without a verified contact email is rejected even when the problem is real. The pipeline is tuned for defensible offers, not volume.
Safe to repeat
Replaying a publication, a payment webhook or a booking returns the same CRM records, draft and invoice. We tested this deliberately, not by accident.
Least privilege
Investigation, offers, billing, mail and voice run as separate services with scoped credentials. The browser is configured to block private network access and has no CRM or accounting credentials.
A clear next step for the customer
Payment confirmation comes from the server before booking opens. The customer sees available times, confirms a slot and can return to the page without starting over. An invoice issue does not block appointment booking.
Failures stay visible
Accepting a queued job is not the same as completing it. The workflows wait for investigation and qualification results, and report failures. Uncertain external writes pause for reconciliation.
03 / WHAT THE NUMBERS SAY SO FAR
Real checks. Clearly bounded results.
Recorded on 18 September 2026. CRM and mailbox publication were tested live; payment, invoice drafting and booking were tested separately with a sandbox offer. These are technical acceptance results, not customer revenue or conversion claims.
- Four websites in the latest quality review
- One qualified preview offer for a specific missing-information fix. Two websites with no significant issue. One held for review because the available evidence could not establish the site’s condition. Nothing from this sample was published or emailed.
- One publication, one replay
- Company, opportunity, note, offer page and Outlook draft were created once. The replay created nothing new and returned the same draft.
- One separate sandbox payment journey
- Stripe sandbox checkout, signed webhook, sevDesk draft invoice with net, VAT and gross verified independently, and a Teams meeting booked for the customer.
- Over 180 automated tests
- Queue, fencing, idempotency and recovery checks against a real database, plus service and boundary tests for offers, billing, mail and voice.
What we are still improving
- Human review remains part of the process: validate the observation, scope, wording and required customer access before sending.
- Automatic invoice drafts are currently limited to the tested sandbox pricing path. Existing pipeline offers do not automatically enter it. Production billing, invoice sending and accounting payment posting remain separate work.
- The generated email is still too long and explains too much of its method. We are shortening it to observation, consequence, fix and next step.
- The latest review confirmed that linked menu PDFs were read and uncertain findings stopped qualification. Some narrative claims still need editing; evidence checks cannot validate every sentence.
Could this work for you?
Not every business needs a prospecting pipeline. The building blocks transfer: an evidence-backed check, a reviewable draft, replay-safe CRM writes, and a payment step that leads to an invoice and an appointment. We start with the one that removes the most copying.
Discuss a similar workflowLET’S FIND YOUR STARTING POINT
What could work better for you?
Tell us where things get stuck. We’ll help you identify a practical next step.
A few useful answers.
Does the AI send emails to businesses?
No. It prepares an Outlook draft. A person reads, edits and sends it. The same applies to invoices, which are created as drafts.
How do you avoid inventing problems?
The investigator must reference saved page evidence for every statement. A rule-based gate then requires a real failed customer task and a feasible fix. Without sufficient evidence, no offer is prepared.
What happens if a step fails halfway?
Completed writes are recorded and reused on replay. If a remote write has an unknown outcome, the process pauses for reconciliation before trying again. The acceptance tests verified reuse of CRM records, an Outlook draft, a test invoice and a calendar booking.
Which tools do I need for something similar?
Your existing CRM, mailbox and invoicing tool are usually enough. n8n and a small private service connect them. We check the integration options for your tools first.