I am thoroughly familiar with ChatGPT, Claude, and other AI models. I bring AI to where it noticeably reduces your workload: into your apps, flows, and documents, with clear data protection rules and with a person who makes the final decision.
New inquiry in the mailbox"Tuesday's delivery is incomplete …"
Language model: understand the requestChatGPT or Claude · response as JSON
CategoryComplaint
Urgencyhigh
SummaryPartial delivery, 2 items missing
SuggestionTrigger replacement delivery
Review by the teamAccept suggestion?AcceptEdit
01Use cases
Where AI helps you
Orientation and use cases
Where does AI really help, and where does it not? I examine your processes for tasks that language models do well - understanding, summarizing, and classifying text, as well as extracting data - and assess benefits, costs, and risks.
AI in Power Automate
Integrate language models such as ChatGPT or Claude into flows through an API: with fixed instructions, a structured response, and a check of the result before it is processed further.
Analyzing documents and emails
Extract data from inquiries, invoices, delivery notes, and forms, summarize the content, and assign it to the right case - with AI Builder or a language model, whichever fits better.
AI in Power Apps
Suggest text, summarize long entries, or populate structured fields from free text - directly in the app, as a suggestion with a preview instead of silent automation.
Maintaining instructions properly
Instructions, examples, and the expected response format are stored centrally and versioned. This makes it possible to trace how the AI works and make targeted improvements.
Putting Microsoft Copilot in perspective
My focus is on ChatGPT and Claude. I give you an honest assessment of Microsoft 365 Copilot and Copilot Studio: What do they deliver for your case, what do they cost, and when is a custom flow the better choice?
02What I focus on
Quality you notice in everyday work.
Responsibility
The AI suggests. A human decides.
Language models are strong at understanding and organizing text - but they are occasionally wrong and still sound convincing. That is why I build solutions so that decisions with consequences are confirmed by a person: approvals, payments, and external replies.
Instructions and responses are logged, uncertain results go on a review list, and the process keeps working even if the AI service is temporarily unavailable.
Approval before sending
Log of instruction and response
Fallback without AI
Spot checks in operation
Teams · approval
Draft reply to the customerAI suggestion
"Thank you for your message. We will deliver the two missing items by Friday …"
Check the content before sendingApproval by a personSendEdit
LogInstruction, response, and reviewer
Approved14:32 · edited
Fallbackactive without AI
Schematic sample view with fictitious content. Not a measured project result.
Privacy
Your data stays under control.
Which data goes to which model, in which region, and under which contract? I clarify this with you in advance. I rely on business accounts instead of private accounts and leave out personal data where it is not needed for the task.
The legal assessment remains with those responsible for data protection in your organization - I provide the technical facts and implement the requirements.
Business accounts instead of private accounts
Data minimization
Data processing agreement
Costs with a cap
Data flow · language model
Name and emailremoved
Customer numberreplaced
Text of the inquirytransferred
Attachmentsnot needed
AccessBusiness
ProcessingEU, where available
Costswith a cap
Test first, then roll out
A pilot with real examples.
Before AI is used in a production workflow, I test it with real, anonymized examples from your everyday work. This way, you see in advance how often the results meet expectations, where rework is needed, and what a transaction costs.
The solution goes into production only when quality, effort, and data protection requirements are met - and it continues to be monitored there.
Real, anonymized examples
Measuring hit rate and rework
Cost per transaction
Decision on a solid basis
Pilot · sample inquiries
Test run with anonymized inquiriesAI result checked against expectations
Categorysuggested
Summaryfor review
Reworkmanually if needed
Costsdepending on model and volume
DecisionRollout with approval by the teamNext
Schematic sample view with fictitious content. Not a measured project result.
03Approach
From the first conversation to a solution in production.
Consulting and implementation from a single source: whoever writes the concept also builds it - with no friction between consultant and developer.
Initial consultation
Free and without obligation via Microsoft Teams: What is the challenge, what is the goal, and which licenses do you already have? Afterward, you will know whether and how the Power Platform fits your needs.
Process analysis
Together with the people who run the process every day: How does it work today, and where do manual handoffs between systems or media, delays, and errors occur?
Solution concept
Data model, screens, flows, permissions, and licensing - in writing, easy to understand, and with a clear effort estimate.
Implementation in stages
Short iterations, each with a working interim result. You see early on what is being built and can adjust course.
Rollout and handover
Acceptance, deployment to production, documentation, and handover to your team - so the solution does not depend on me.
Further development
If you wish, I will continue to support the solution afterward: enhancements, new requirements, and updates from Microsoft.
04Toolbox
Technologies in this area
ChatGPT
Claude
OpenAI API
Anthropic API
Azure OpenAI
AI Builder
Power Automate
Power Apps
HTTP and REST
JSON schema
Copilot Studio
Microsoft 365 Copilot
05Real-world projects
Projects in this area. No names, real substance.
As a rule, I do not name clients - their industry, size, and project requirements say enough.
11/2021 - 12/2021Implementation
Shading systems
more than 300 employees · international sales
Rollout of Microsoft Power Automate as the central automation platform and automation of invoice processing with Microsoft Power Automate (cloud and desktop flows), Microsoft AI Builder, and Microsoft Dynamics 365 Business Central. In addition, real-time visualization of presence status via the Microsoft Graph API in SharePoint and Microsoft Power BI, as well as training on Dataverse and canvas apps.
Which AI is the right one: ChatGPT, Claude, or Copilot?
That depends on the task, the data, and your licensing situation. ChatGPT and Claude are very strong at texts, analyses, and structured answers, whereas Microsoft 365 Copilot is tightly integrated into Microsoft 365. I test the candidates with real examples from your process before we decide.
May we pass customer data to an AI?
Not automatically. The contract, storage location, and purpose matter. Together, we determine what data is necessary, which service and contract are suitable, and how to avoid using personal data. The legal assessment is handled by whoever is responsible for data protection in your organization.
What does AI cost in operation?
Language models are usually billed by the amount of text processed. For typical tasks such as summarizing an email, this is often only a fraction of a euro per transaction. I calculate costs in advance based on your volumes and build in caps so that costs remain predictable.
Can AI take over our processes completely?
Rarely, and that is a good thing. AI takes a lot of reading, sorting, and writing work off your hands. Decisions with consequences - approvals, payments, and replies to customers - should be confirmed by a person. That is how I build the solutions.
Your process. My solution.
Tell me where things get stuck today - in a free, no-obligation initial consultation via Microsoft Teams, we will clarify what the Power Platform can do for your process and what a sensible first step looks like.