Treukontax SteuerberatungTax advisory
Making tax knowledge accessible with AI. In seconds, not minutes.
Cloud Dream Team turned Treukontax's expensive, constrained AI initiative into a faster, more economical and controllable knowledge platform, engineered around the firm's business requirements.
Enterprise AI · Microsoft Foundry · Knowledge management
Business results
Faster answers, lower costs and greater control.
- 43 s → 4 sResponse time reduced by approximately 91%
- Lower costsLower AI consumption costs than under the original Microsoft Copilot Studio model
- Greater flexibilityCustom AI logic, adaptable knowledge retrieval and more control over the assistant
The business challenge
The knowledge was there. Reaching it took too long.
Tax professionals at Treukontax deal with demanding questions every day, on tax regulations, legislation, internal processes and the firm's own professional knowledge. Much of that knowledge was already available in SharePoint, but finding the right passage took time. Treukontax wanted an internal AI knowledge assistant that employees can ask in natural German and that answers from the firm's own sources, with every answer traceable to where it came from. The assistant was meant to support professional judgement, not to replace it.
The first approach
The first approach proved the idea. It also showed the limits.
Treukontax first built its own assistant in Microsoft Copilot Studio, with the firm's SharePoint knowledge as its source. The prototype showed what AI could do for daily work. In regular use, three limitations held back adoption.
Speed
Around 43 seconds
per answer. Too long for a question that comes up in the middle of client work.
Cost
High consumption costs
AI consumption under the Copilot Studio model made broad everyday use expensive.
Flexibility
Limited customisation
Few options to shape the AI logic, how knowledge is retrieved, how answers are written and how the application behaves.
The transformation
Four stages from prototype to platform.
The improvements did not come from switching platforms. They came from engineering: Cloud Dream Team designed the agent logic, the knowledge retrieval and the operating model around Treukontax's requirements, and keeps reviewing the architecture as Microsoft's capabilities evolve. The journey ran through four stages.
Starting point in Copilot Studio
Treukontax built its first assistant itself in Microsoft Copilot Studio, using the firm's SharePoint knowledge. Answers took around 43 seconds, consumption costs were high and customisation was limited, which held back adoption.
Exploring Microsoft Foundry
Cloud Dream Team explored a more flexible solution on Microsoft Foundry that answers from the firm's own documents (retrieval-augmented generation, RAG). Publishing it to Microsoft Teams with the early capabilities available at the time showed limitations in user experience, answer quality and integration maturity.
Engineering by Cloud Dream Team
Cloud Dream Team developed tailored agent logic, optimised knowledge retrieval and improved the way SharePoint knowledge reaches the assistant. The existing knowledge base in Microsoft Foundry was kept, and its ingestion and source management were improved, including control for administrators. Evaluation, versioned instructions, testing, rollback and stronger operational governance were introduced.
Next step: Microsoft Teams
Following its latest decision, Treukontax will bring the assistant to employees in Microsoft Teams, using Microsoft's emerging capability to publish RAG agents there. It is a deliberate, forward-looking choice. The integration is still evolving, and Cloud Dream Team's engineering, evaluation and governance practices support the next steps.
Measurable business impact
From a promising prototype to a knowledge platform designed for the whole firm.
Answers now arrive in seconds, so employees can use the assistant during client work rather than after it. AI consumption costs are lower than under the original Microsoft Copilot Studio model, making broad daily use more economical. These gains come from engineering decisions, not from a change of platform alone. Treukontax now controls its knowledge sources, AI instructions and answer behaviour, as well as the creation and further development of agents.
| Previous Copilot Studio agent | CDT's tailored AI solution | |
|---|---|---|
| Response time | Around 43 seconds | Around 4 seconds |
| AI consumption costs | High consumption costs | Lower consumption costs |
| Adaptability | Limited configuration options | Knowledge sources, AI instructions, answer behaviour and agent development under Treukontax's control |
Engineering & quality evidence
Measured, not assumed.
Before the decision for Microsoft Teams, Cloud Dream Team tested the tailored assistant against an earlier configuration on Microsoft Foundry. The figures below describe that evaluated configuration and the original design, not the experience in Teams.
- 16:0Preferred on 16 of 18 customer test scenarios, with two ties and no losses
AI-evaluated comparison with an earlier Microsoft Foundry configuration. Not a Microsoft Copilot Studio benchmark.
- 1 stageAnswer generation reduced from two model stages to one
Evaluated architecture
- ~1,000Daily users the solution was designed for
Original design capacity, not actual adoption
Governance
Control that pays off in daily work.
These capabilities keep the assistant trustworthy and adaptable as more people rely on it.
Answers employees can check
Every answer can be checked against the original knowledge source it came from.
Sources chosen by the firm
Administrators decide which knowledge sources the assistant uses.
Instructions under control
AI instructions are tested before release, versioned and can be rolled back.
Data handled as agreed
How data is handled and how long it is kept remain governed.
Improvement without a project
Treukontax can improve the assistant without a development change for every adjustment.
Customer voice
Our goal was to make specialist tax knowledge more accessible to our employees. With Copilot Studio, we encountered limitations in performance, cost and flexibility. Together with Cloud Dream Team, we developed a solution that is significantly faster, more cost-efficient and more flexible.
The next step
The best of both worlds.
Treukontax combines the familiar Copilot experience with the flexibility of an agentic platform provided by Cloud Dream Team. Its development team can integrate models from Microsoft Foundry together with its own tools, APIs and business-system interfaces.
In future, the agent is intended not only to answer questions but also to work with connected systems and initiate changes. This will allow Treukontax and Cloud Dream Team to implement functions and automations that go beyond the capabilities of a conventional Copilot agent.
Key technologies: Microsoft Foundry · Azure AI Search · SharePoint Online · Microsoft Teams
Are you facing a similar decision?
In 30 minutes, we'll clarify together:
- Clarify your current situation and constraints
- Discuss an approach that suits your environment
- Agree the next practical step
Prefer to write? hello@clouddream.team
30 minutes · Microsoft Teams · No obligation
