evidoria

← Back to browse

Good practice

AMA virtual assistant — Portugal's AI chatbot for public services

Portugal · Lisbon · See the Portugal profile

Portugal's AMA launched an AI chatbot on ePortugal.gov.pt in May 2023. By December 2023, 23,780 conversations (101/day) and a 10% rise in Digital Mobile Key activations. Expanded by end-2024 to cover 2,300+ public services in 12 languages via ChatGPT.

23,780 conversations
Chatbot conversations handled (May–Dec 2023)
101 conversations/day
Average daily conversations (May–Dec 2023)
10 %
Increase in Digital Mobile Key (CMD) activations (following May 2023 deployment)
2,300+ services
Public services covered after expansion (end of 2024)

Details

Maturity
Scaling
Promoter
Agência para a Modernização Administrativa (AMA)
Period
2023-present
Keywords
virtual assistant, generative AI, chatbot, citizen services, digital identity, Azure OpenAI, multilingual, eGovernment

Context

Portugal's Agência para a Modernização Administrativa (AMA) launched an AI-powered chatbot on the ePortugal.gov.pt portal in May 2023, piloting citizen support for the Digital Mobile Key (Chave Móvel Digital), Portugal's digital identity and e-signature service.

Objectives

The initiative aimed to help citizens navigate CMD activation, digital signature use and expiry-date queries through a 24/7 automated assistant, while testing generative-AI infrastructure ahead of a broader rollout across public services.

Activities

The chatbot was built through a public-private partnership: Microsoft supplied Azure Cloud and OpenAI Service infrastructure, Defined.AI optimised the Portuguese-language voice model, and DareData handled integration. AMA invested in knowledge-base curation and quality-assurance processes specifically to mitigate AI hallucination risk. By late 2024 the assistant was rebuilt into a full generative-AI assistant powered by ChatGPT with a specialised knowledge base covering over 2,300 public services in 12 languages, with transactional capabilities (service execution, process tracking, appointment scheduling) planned for 2025.

Results

Between May and December 2023 the chatbot handled 23,780 conversations (about 101 per day), covering mainly CMD activation, digital signature and expiry-date topics. A 10% increase in CMD activations was observed following deployment, credited in part to the assistant's guidance role. The case was recognised by OECD OPSI as a public-sector AI innovation.

Conclusions

Phase 1 evidence is limited to one service type over seven months with no independent evaluation of causal attribution; Phase 2 outcomes (2,300+ services, 12 languages) have not yet been publicly reported.

Implementation
Indicative cost
Medium (€50k–€500k) — Multi-vendor cloud/AI partnership (Microsoft Azure/OpenAI Service, Defined.AI, DareData); no public cost figures found in source.
Time to results
Medium (1–3 years) — Piloted May 2023 on one service; expanded to a full generative-AI assistant covering 2,300+ services in 12 languages by end 2024; transactional features planned for 2025.
Staffing & skills
AMA (Agência para a Modernização Administrativa) as lead government agency, Microsoft providing Azure Cloud and OpenAI Service infrastructure, Defined.AI for Portuguese-language voice model optimisation, DareData for systems integration and implementation

Conditions for success

  • Extensive knowledge-base curation and quality-assurance processes to reduce AI hallucination risk
  • Multi-vendor public-private partnership combining cloud, language and integration expertise
  • Phased scope expansion from a single pilot service to 2,300+ services

Common failure modes

  • AMA itself flags hallucination risk as an active, acknowledged challenge
  • No independent impact evaluation has been published to confirm the pilot's CMD-activation effect is causal

Where it fits

Governance type
national government agency
Scale
national (phased from single-service pilot to national rollout)
Income level
high-income (Portugal/EU)

Data sources

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

Similar practices you may find useful