evidoria

← Back to browse

Good practice Imported

Decide Madrid's NLP Toolkit for Citizen Proposals on the Consul Platform

Spain · Madrid · See the Spain profile · See the Madrid profile

Evidence: Descriptive / self-reported Top 32% 67/100 · Ask Evidence Copilot about this practice

Warwick, QMUL and Alan Turing Institute researchers built four NLP/ML tools — recommendation, interest grouping, summarisation, aggregation — for Madrid's Decide Madrid Consul platform, scaled to 464,654 users and 2,400 daily proposals.

464,654 users
Registered platform users (2020)
800–2,400 proposals/day
Proposals received per day at peak (peak period)
135 institutions
Institutions running the Consul platform worldwide (current)
35 countries
Countries using the Consul platform (current)
90 million
Estimated citizens reached by the Consul platform globally (current)
182 proposals
Previously approved proposals discarded after the 2019 change of administration (2019)
Decide Madrid's NLP Toolkit for Citizen Proposals on the Consul Platform

Details

Maturity
Pilot
Promoter
Ayuntamiento de Madrid / Consul Democracy Foundation
Period
2015–ongoing (platform); NLP research 2018–2021
Region (NUTS)
ES30
Keywords
digital democracy, participatory budgeting, civic tech, open-source government software

Context

Decide Madrid launched in 2015 under Mayor Manuela Carmena on the open-source Consul platform, growing from 180,000 users in 2016 to 464,654 by 2020 and receiving between 800 and 2,400 proposals a day at peak. The Consul codebase is now stewarded independently by the Consul Democracy Foundation (established 2019) and runs in 135 institutions across 35 countries, reaching an estimated 90 million citizens.

Objectives

To help the platform cope with that volume, a research team applied NLP and machine learning directly to Decide Madrid/Consul data, aiming to ease the barriers citizens faced in navigating high volumes of proposals.

Activities

The team built four capabilities: recommending relevant proposals to citizens, grouping citizens by shared interests, summarising comments on proposals, and helping citizens aggregate and refine overlapping proposals. A related 'Decide ML' tool was separately built to route proposals to the correct government department.

Results

The published evaluation concludes that NLP and machine learning 'have a role to play in addressing some of the barriers' platform users faced. Governance proved fragile in practice: after Madrid's 2019 change of administration, the city discarded 182 previously approved proposals and halved participatory-budgeting resources.

Conclusions

The underlying Consul software is proven highly transferable at global scale, but the AI/NLP layer itself has so far been validated mainly in this one Madrid case study, and political discontinuity — not the technology — was the main risk observed in practice.

Implementation
Indicative cost
Medium (€50k–€500k) — The Consul platform itself is free, open-source software; specific budget figures for the NLP research programme or municipal deployment are not disclosed in the source material.
Time to results
Long (> 3 years) — Platform launched 2015; NLP research conducted and published 2018–2021; Consul Democracy Foundation established 2019 following a change of city administration that also affected PB budgets.
Staffing & skills
Ayuntamiento de Madrid (platform owner, until 2019 administration change), Consul Democracy Foundation (independent platform steward, established 2019), Academic research team: Miguel Arana-Catania, Felix-Anselm Van Lier, Rob Procter, Nataliya Tkachenko, Yulan He, Arkaitz Zubiaga, Maria Liakata

Conditions for success

  • An open-source civic-tech platform (Consul) with a large, active user base to generate training data
  • Independent, foundation-based stewardship insulated from single-administration political cycles
  • Sustained political commitment to participatory budgeting outcomes, not just the software

Common failure modes

  • Change of city administration in 2019 discarded 182 previously approved proposals and halved PB budget resources — AI tooling did not by itself guarantee political follow-through
  • The AI/NLP layer has been validated mainly in this one Madrid case study, so its generalisability elsewhere is not yet demonstrated

Where it fits

Governance type
municipal government, with platform stewardship now held by an independent foundation
Scale
city-wide (Madrid), platform used globally
Income level
high-income

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.

Attachments

Similar practices you may find useful