Service · MCP & LLM Integration
Your customers ask ChatGPT, Claude, Copilot, Gemini or another AI assistant and get answers straight from your SaaS application. Or they use an AI assistant we build right into your product. That takes load off your support and opens up a new revenue stream for you.
In brief
More and more teams get their work done in an AI assistant. They ask it for figures and have it write reports. An application that is not connected gets bypassed. Its data is exported and rebuilt in Excel.
The bridge is the Model Context Protocol (MCP), an open standard through which AI assistants access applications. Originally developed by Anthropic, it is now supported by nearly every major AI platform, from ChatGPT and Claude to Microsoft Copilot and Gemini, as well as by countless developer tools and AI agents.
Way 1 · MCP server
Anyone already working with ChatGPT, Claude, Copilot or another AI assistant connects your SaaS application there in a few clicks. From then on, the assistant they know answers questions directly with data from your software.
Ideal for teams that already work with AI every day.
Way 2 · In-product AI assistant
A chat window right inside your SaaS application. Your users ask questions without leaving the product and without an AI subscription of their own. Answers appear as a table or as a link to the right view.
Ideal for every user, and sellable as a paid add-on.
Both ways share the same foundation. We build it once, your customers benefit twice.
Case study
A software company from Vienna with a SaaS application for property managers: properties, tenants, payments, maintenance and documents in one place. An established product whose customers increasingly ask whether it works with their AI.
The company is fictional and the figures are illustrative. The approach and the lessons learned come from our work with MCP servers and AI assistants.
The result up front
The path to the result
Step 01 / 08 · First conversation
The managing director of Hausfeld Software GmbH gets in touch with two observations: in sales, prospects increasingly ask whether the software works with ChatGPT or Copilot. And support answers questions every day whose answer is already in the data. A competitor is already advertising AI features. The company wants to follow suit, but carefully, because tenant data is sensitive and must not end up with an AI provider unchecked.
From practice
Many start by asking for the best language model. What matters more is which questions your customers ask and which data they need for them.
Call notes
First call · 60 min
“Our customers want to know whether we work with their AI. And our support answers questions every day whose answer is already in our system.”
Our notes
Next step: analyse three months of support requests
Step 02 / 08 · Use cases
We analyse three months of support requests and talk to six property managers about their day-to-day work. The result is clear. 45 percent of all requests are lookups whose answer is already in the software. Five questions make up most of them. This becomes the first list of tools the AI gets.
From practice
The first ten tools decide success. Starting with your customers’ most frequent questions delivers measurable benefit from day one.
“Which tenants are behind on rent?”
14 %
“Where do I find the statement or document for …?”
11 %
“When is the next maintenance due at property …?”
9 %
“Which leases are about to expire?”
6 %
“How many units are vacant?”
5 %
Share of all support requests
45 %
Basis: 2,700 support requests over three months, six customer interviews.
Step 03 / 08 · Data & permissions
Before we build the first tool, we clarify which data exists and who may see it. In the software of Hausfeld Software GmbH, every role sees only what it needs for its work. These rules apply unchanged to the AI. It only answers with data the person asking could open themselves.
From practice
Data protection is your customers’ most common objection to AI. A clear permission concept turns it into a selling point.
| Properties | Tenants | Payments | Maintenance | |
|---|---|---|---|---|
| Management | assigned properties only | assigned properties only | assigned properties only | assigned properties only |
| Accounting | all properties of the company | all properties of the company | all properties of the company | no access |
| Caretaker | assigned properties only | no access | no access | assigned properties only |
| Owner | assigned properties only | no access | assigned properties only | assigned properties only |
Every AI call runs through the same permission rules as the interface.
Step 04 / 08 · Concept
The use cases become a catalogue of tools. Each tool does exactly one business task, such as analysing arrears or creating a case. Templates for recurring analyses come on top. Read tools come first. Write tools follow only once reading works reliably, and they only create drafts that a person approves.
From practice
If your application already has a REST API, the tools build on the same business logic. That keeps application and AI access in sync, and every new tool is immediately available to both ways.
Read tools
Phase 1
search_properties
Search properties and units
analyze_arrears
Analyse rent arrears
find_documents
Find statements and documents
list_maintenance
List due maintenance
list_expiring_leases
Expiring leases
analyze_vacancy
Analyse vacancy
Write tools
Phase 2, with approval
create_ticket
Create a case or damage report
draft_payment_reminder
Payment reminder as a draft
Analysis templates
Step 05 / 08 · Technology & model choice
The tools build directly on the existing business logic of the SaaS application. The language model is swappable and can be changed later without rebuilding anything. A test with real customer questions decides which model it will be. Because tenant data is involved, the choice falls on a model hosted in an EU data centre.
From practice
Which language model fits best depends on your questions and on your data location requirements. For most applications, a good model hosted in the EU is the better choice than the strongest model in the US.
Your customers’ AI assistant
ChatGPT, Claude, Copilot and more, connected via MCP
In-product AI assistant
Chat window right inside the SaaS application
Tools
One tool per task, shared by both ways
SaaS application data
Database and business logic stay unchanged
Language model
swappable by configuration
Model test with real customer questions
| Model | Correct tool calls | Data location |
|---|---|---|
| Local model, 8 bn parameters own server |
71 %
|
own server |
| EU cloud model chosen EU |
94 %
|
EU |
| US frontier model USA |
97 %
|
USA |
Step 06 / 08 · Implementation
In the first three weeks, we build the MCP server with the read tools and access management for your customers. First customers can already connect from their own AI assistant. Then comes the in-product AI assistant, a chat window in the SaaS application that uses the same tools and shows results as tables and links to the right view.
From practice
After just three weeks, the first customers can connect their own AI assistant. The benefit becomes visible early, and their feedback flows straight into the next stages.
Three stages
Weeks 1–3
MCP server, read tools, access, documentation
Weeks 4–6
In-product assistant, system prompt, result display
Weeks 7–8
Write tools with approval, tests with real questions
Step 07 / 08 · Pilot & rollout
Twelve property managers get the assistant switched on for a month of testing. We review the conversations anonymously and see where the AI misunderstood a question or where a tool was missing. Then the assistant launches as a paid add-on. The MCP server is open to all customers.
From practice
Product documentation written for AI assistants improves answers more than any instruction in the system prompt. Your customers will read it too.
Weeks 1–4
Pilot
Twelve property managers work with the assistant.
Week 5
Review
Conversations reviewed anonymously, tools and documentation sharpened.
Week 6
Plan launch
AI assistant as an add-on for €12 per user and month.
Ongoing
Expansion
Every new feature gets its tool right away.
Adopted from the pilot
Step 08 / 08 · Measurement & result
Six months after the plan launch, we measure against the baseline from the analysis. Support gets noticeably fewer lookup questions. 42 percent of users have booked the add-on, and more than half of the customers have connected their own AI assistant. In sales, “works with ChatGPT, Claude and Copilot” has become a visible argument.
From practice
The add-on pays for itself from day one, because language model costs are only a fraction of the add-on price and can be planned precisely per customer.
| Metric | Before | After |
|---|---|---|
| Support requests per month | 900 | 680 |
| Arrears report | 25 min | 30 sec |
| Users with AI add-on | – | 42 % |
| Customers with own AI assistant | – | 57 % |
Economics per year
From our own experience
In our own company, we use three MCP servers with around 75 tools, among them for CRM and finance. So we also know MCP from the user’s side. We know what makes a good answer and where a tool does too much or too little.
Transferable
Wherever people search or analyse data from an application, AI can shorten the work. Typical starting points:
MCP server and assistant as an offering for your customers, optionally as a paid plan.
Employees ask their ERP or ticket system in plain language instead of building reports.
Customers find invoices and contracts themselves, around the clock, with real data.
Manuals and project files become searchable and answer questions with sources.
One AI assistant works across several applications without a dedicated integration project.
When data must not leave the building, we run the language model in your infrastructure.
If you nod at two or more of these, a conversation is worth it.