evidoria

← Back to browse

Good practice Imported

JuntaGPT — Andalusia's internal generative-AI assistant for civil servants

Spain · Sevilla · See the Spain profile

Evidence: Descriptive / self-reported Top 75% 47/100 · Ask Evidence Copilot about this practice

Andalusia's regional government launched JuntaGPT, an internal generative-AI assistant giving up to 240,000 public employees access to a Gemma 3-based tool covering 400+ procedures, though its efficiency claims are self-reported and unaudited.

240,000 employees
Public employees with projected access at launch (June 2025)
400+ procedures
Administrative procedures covered (June 2025)
<300 milliseconds
Target response latency
2 years to 15 days
Projected case-file resolution time (government projection, unaudited)
1,000,000+ documents/year
Projected annual document savings (government projection, unaudited)
JuntaGPT — Andalusia's internal generative-AI assistant for civil servants

Details

Maturity
Scaling
Promoter
Agencia Digital de Andalucía (ADA), Junta de Andalucía
Period
{'end': '', 'start': 'June 2025'}
Region (NUTS)
ES61
Keywords
generative AI, digital government, civil service productivity, document processing

Context

Andalusia's Agencia Digital de Andalucía (ADA) launched JuntaGPT in June 2025 as an internal generative-AI assistant for the Junta de Andalucía's civil service, running Google's open-weight Gemma 3 model (27 billion parameters) on the Hércules supercomputer at the CICA centre in Sevilla. The tool is designed to respond in under 300 milliseconds and cover more than 400 administrative procedures, from drafting correspondence to processing case files, and was projected at launch to reach up to 240,000 Andalusian public employees. JuntaGPT is strictly internal and does not replace human-staffed citizen service channels.

Objectives

The regional government's stated aim is to speed up internal administrative work: officials projected that case files previously taking up to two years could be resolved in fifteen days, with savings across more than one million documents annually.

Results

These efficiency figures are launch-day government projections, not independent, audited evaluation results. No post-implementation usage logs or before/after processing-time data have been published as of this review.

Conclusions

JuntaGPT demonstrates a large-scale, open-weight generative-AI deployment reusing existing public compute infrastructure, but its impact claims remain unverified pending independent evaluation.

Implementation
Indicative cost
Medium (€50k–€500k) — Deployed on existing public-sector compute (Hércules supercomputer, CICA), which may lower relative capital cost, but no published budget or licensing figures.
Time to results
Medium (1–3 years) — Launched June 2025 as an operational service, not framed as a phased pilot; no rollout timeline beyond launch was published.
Staffing & skills
Agencia Digital de Andalucía (ADA), CICA (Centro Informático Científico de Andalucía) technical/compute team operating the Hércules supercomputer

Conditions for success

  • Reuse of existing public supercomputing infrastructure (Hércules at CICA) rather than new build
  • Use of an open-weight model (Gemma 3) to avoid vendor lock-in
  • Clearly scoped set of 400+ administrative procedures

Common failure modes

  • Efficiency claims are self-reported government projections with no independent audit or published before/after data
  • No disclosed bias/fairness or ethics-review mechanism

Where it fits

Governance type
regional/sub-national government
Scale
large sub-national civil service (up to 240,000 employees)
Income level
high income (Spain)

Commonly funded by

National / regional programmes Digital Europe Programme

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