The Audit Meeting Context
A Geneva audit and accounting firm employing 35 professionals held about fifty weekly client meetings to discuss annual accounts, tax issues, management recommendations, and action plans. Each meeting required a staff member to take detailed notes to document exchanges, decisions made, and agreed actions. Formal report writing then took 45 minutes to an hour per meeting.
This process represented a considerable burden, mobilizing about 50 weekly hours for note-taking and writing. Moreover, quality varied by writer, some important information was sometimes omitted, and clients had to wait several days before receiving the report. The firm sought a modern solution to automate this repetitive task while improving quality and speed.
The Intelligent Transcription Solution
We developed a complete meeting transcription and analysis system based on Azure Speech Services, Azure OpenAI, and Microsoft Teams. The architecture integrates naturally into the firm's daily tools.
All client meetings take place in Microsoft Teams, which natively offers a recording function with automatic transcription. We configured a compliance policy enabling systematic recording of client meetings after explicit consent at session start. Recording and transcription are stored in SharePoint with permissions restricted to participants.
Teams' raw transcription correctly identifies different speakers but lacks structure and synthesis. This is where our intelligent analysis system intervenes. At each meeting's end, a Power Automate flow automatically triggers and retrieves the complete transcription plus metadata (participants, duration, date).
This transcription is sent to a GPT-4 model deployed in Azure OpenAI Service with very precise instructions to structure information. The system prompt asks the model to analyze the conversation and produce a structured document in several sections: executive summary of 3-4 lines capturing the meeting's essence, context and meeting objectives, key points discussed organized by theme, decisions made with decision-maker's name, actions to undertake with assigned responsible party and agreed deadline, pending points requiring later follow-up, and next steps.
For identified actions, the system automatically creates tasks in Microsoft Planner assigned to the right people with deadlines mentioned during the meeting. A follow-up task is also created for the file manager if points remained pending. This automation guarantees no commitment is forgotten.
The generated report is cleanly formatted in a Word document stored in the client's SharePoint folder, and a summary email is automatically sent to all participants within the hour following meeting end. Clients appreciate this responsiveness whereas they previously had to wait two to three days.
An additional feature analyzes conversation sentiment and tone. The model detects if the client expresses concerns, dissatisfaction, or conversely enthusiasm. These indicators are useful for relationship management, enabling identification of clients requiring particular attention or conversely very satisfied ones who might recommend the firm.
For complex technical discussions involving amounts, dates or regulatory references, the system automatically extracts these factual elements and presents them in a separate table, facilitating verification and later follow-up.
Measured Results
After eight months of use, the system transformed the meeting documentation process. Time devoted to note-taking and report writing dropped 85%, from 50 weekly hours to 7.5 hours devoted only to reviewing and adjusting generated documents. This savings represents the equivalent of one full-time position freed for value-added activities.
Report quality and comprehensiveness improved. AI-generated reports are more complete because they omit no discussion element, unlike a human who can miss information while trying to note and listen simultaneously. Standardized structuring facilitates reading and information search.
Client satisfaction significantly progressed, measured at +32 NPS points. Clients greatly appreciate receiving the report within the hour following the meeting, demonstrating the firm's professionalism and responsiveness. Several clients spontaneously mentioned this efficiency as a positive differentiator.
Action tracking considerably improved. Action completion rate within agreed deadlines rose from 72% to 94%. Automatic Planner task creation with notifications guarantees nothing is forgotten. File managers have complete visibility on ongoing commitments via their Planner boards.
An unexpected benefit is the archive value of complete transcriptions. In case of later dispute or disagreement about what was said or agreed, the firm can retrieve exactly the original conversation, offering important legal protection. This traceability already helped clarify several misunderstandings.
Compliance and Data Protection
Confidentiality was a major concern for an audit firm handling sensitive financial information. The architecture guarantees maximum protection. All recordings and transcriptions are stored in SharePoint with encryption at rest and in transit, accessible only to meeting participants and file managers. The GPT-4 model used is deployed in private mode with Microsoft contractual commitment that no data is used to train other models.
Explicit consent is requested at the start of each client meeting, explaining the conversation will be recorded and transcribed to document exchanges. This consent is timestamped and preserved. Participants can request certain sensitive passages be removed from the final transcription.
Automatic retention policies delete recordings after 7 years in compliance with accounting archiving legal obligations, while keeping structured reports indefinitely.
Technical Architecture
The system relies on native Microsoft Teams integration with SharePoint for initial recording and transcription, Power Automate Premium for analysis workflow orchestration, Azure OpenAI Service for report analysis and generation, Microsoft Planner for action task management, and Word Online for formatted document generation.
Power Automate flows are organized in reusable components: a master flow triggered at each meeting end calls specialized sub-flows for transcription retrieval, AI analysis, task creation, and document generation. This modular architecture facilitates maintenance and evolutions.
Monthly cost including Power Automate Premium, Azure OpenAI consumption and storage represents approximately 450 CHF. The 42.5 weekly hours saved (170 monthly hours) generate very comfortable ROI.
Future Evolutions
The firm plans to extend the system to internal team meetings, facilitating project tracking and strategic decisions. An automatic translation feature could enable generating reports in multiple languages for international clients. Finally, longitudinal transcription analysis could identify recurring topics or emerging concerns across the client portfolio.
Conclusion
This automatic transcription and analysis system demonstrates how AI can eliminate a tedious administrative task while improving result quality. By freeing professionals from note-taking, we enable them to focus fully on listening to the client and their expertise value. The firm has a clear competitive advantage through its responsiveness and professionalism, while optimizing operational costs.