evidoria

← Back to browse

Good practice Imported

AI-Assisted Parliamentary Reporting — The Flemish Parliament's Whisper-Based Transcription Pipeline

Belgium · Brussels · See the Belgium profile · See the Brussels profile

Evidence: Descriptive / self-reported Top 66% 53/100 · Ask Evidence Copilot about this practice

The Flemish Parliament's IT unit built a pipeline combining Pyannote, OpenAI Whisper and ChatGPT to speed up plenary reporting; by October 2025 nearly 80% of editors used it regularly, ahead of a 2026 full rollout.

~80%
Editors using the tool regularly (October 2025)

Details

Maturity
Scaling
Promoter
Vlaams Parlement (Flemish Parliament) IT Department, with VLAIO's Innovative Public Procurement Programme (PIO)
Period
2022–2026
Region (NUTS)
BE10
Keywords
parliamentary administration, legislative documentation, public sector IT

Context

Written reporting in the Flemish Parliament is a labour-intensive, time-pressured process handled by two editorial services that transcribe and summarise plenary and committee sessions. Following an early 'Q-Pilot' proof-of-technology experiment shortly after ChatGPT's public release, the Parliament's IT department built a proof-of-concept combining Pyannote for speaker attribution, OpenAI's Whisper for transcription, and ChatGPT for post-processing of livestreamed plenary video, partly funded through VLAIO's Programma Innovatieve Overheidsopdrachten (PIO) innovative-procurement programme.

Activities

After roughly four months of testing, the proof-of-concept was rolled out to the full editorial team in May 2025. Full integration into the Parliament's editorial application began in January 2026, following iterative prompt refinement and additional training driven by an internal AI task force of editors-in-chief.

Results

By October 2025, nearly 80% of editors were using the tool regularly, and most reported it had become a mainstay that reduces their workload, though editors still manually verify all output against the original audio and net time savings are, in the developers' own words, 'difficult to quantify at this stage' — some editors finish faster while others spend more time carefully revising.

Implementation
Indicative cost
Medium (€50k–€500k) — Funded in part through VLAIO's Programma Innovatieve Overheidsopdrachten (PIO), the Flemish Innovation and Entrepreneurship Agency's innovative-procurement programme; no total budget figure is disclosed.
Time to results
Long (> 3 years) — Q-Pilot proof-of-technology experiment shortly after ChatGPT's public release (2022) → ~4-month IT-department proof-of-concept → full editorial team rollout May 2025 → ~80% regular use by October 2025 → full editorial-application integration beginning January 2026.
Staffing & skills
Flemish Parliament IT department, Two parliamentary editorial services, Internal AI task force of editors-in-chief for prompt refinement and staff training

Conditions for success

  • Built from off-the-shelf, well-documented components (Pyannote, Whisper, ChatGPT)
  • ~4-month proof-of-concept testing period before full rollout
  • Mandatory manual verification of every AI output against the original audio
  • Iterative prompt refinement driven by an internal task force

Common failure modes

  • Net time savings are hard to quantify — some editors finish faster while others spend more time revising
  • Full application integration was still only starting as of January 2026, roughly four years after the initial Q-Pilot experiment

Commonly funded by

National / regional programmes

Indicative funding routes for practices of this type — always check each programme's current calls and eligibility rules.

Do you run this practice? Claim it — verified implementers get a public contact pathway and can propose corrections.

Data sources

Where this practice's information was retrieved from, and when.

Similar practices you may find useful