What problem does it solve?
Government information is spread over thousands of pages written by different departments, and people rarely know which agency owns their problem. They phone or visit because they cannot find or trust the answer online, so contact centres answer many questions the website already covers: which form, which deadline, which office, what does this letter mean. Queues grow at exactly the moments demand spikes, such as tax deadlines, benefit changes or an emergency.
Keyword search and scripted FAQ bots only help people who already know the official term. Generative assistants can read a question in everyday words and synthesise an answer across pages, but a wrong answer from a government channel carries more weight than a wrong answer from a shop. The work is in grounding, refusal, privacy and a clean route to a human.
How does it work?
- Understand the question. The assistant reads the question in the resident's own words and language, and classifies it (information request, personal case question, urgent need, complaint, out of scope).
- Retrieve from official content only. It searches an index of approved guidance pages and answers from the retrieved passages, with links to every source, and refuses when the content does not cover the question.
- Protect personal data. Personal data in the question is detected and masked or the question is rejected; nothing is used to profile the resident.
- Check the answer. A second check screens the draft for unsupported claims, advice the government cannot give, tone and personal data before it is shown.
- Route what it cannot answer. Case specific questions go to the authenticated service or a human adviser with the conversation attached; urgent needs are pointed to the right phone line.
- Audience
- Customer facing
- Autonomy
- Supervised agent
- Adoption
- Mainstream
- Channels
- Web chat, Mobile app, WhatsApp, Phone and voice
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Interactions handled | Not pooled | 20,000 to 2 million | 2 | 2 vendor |
| Accuracy | Too few to pool | at least 76% | 1 | 1 organization |
| Containment rate | Too few to pool | about 20% | 1 | 1 organization |
| Customer satisfaction | Too few to pool | 50% | 1 | 1 vendor |
| Time saved per task | Too few to pool | about 2 minutes | 1 | 1 organization |
| Users served | Not pooled | about 300,000 | 1 | 1 organization |
Value drivers: Inclusion and access, Customer experience, Lower cost to serve, Speed and cycle time.
Indicative value
A national agency that receives 2 million phone and chat contacts a year
USD 480,000 to USD 3.2 million
Human handled contact cost avoided per year
How this is calculated
Formula: contacts * infoShare * containment * costPerContact. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Assisted phone and chat contacts per year contacts, contacts per year | 2,000,000 | 2,000,000 | The reference agency. |
| Share of contacts that are general information questions infoShare, fraction of contacts | 0.3 | 0.5 | Editorial assumption. Replace with your own contact reason analysis. |
| Share of those questions the assistant resolves without a human containment, fraction of information contacts | 0.2 | 0.4 | The low end follows the benchmark on this page (DVLA reports around 20% of web chat enquiries automated by its scripted, non generative bot); the high end is an editorial assumption for a grounded generative assistant. DVLA's figure is a share of all web chat enquiries, not of information questions only, so the two bases differ. Replace with your own pilot data. |
| Cost of a human handled contact costPerContact, USD per contact | 4 | 8 | Editorial assumption for a blended phone and chat contact in the public sector. Replace with your own fully loaded cost. |
What it leaves out: Gross avoided cost only. It leaves out the cost of building and running the assistant, content maintenance, the value of 24/7 access and the extra demand an easier channel can create.
Who already uses it?
9 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
Estonian Information System Authority (RIA)
Estonia · Government and public sector · 2026
Bürokratt is a virtual assistant for citizens that understands everyday Estonian and is available around the clock. Its development is managed by the Information System Authority, and it is built for local authorities and government agencies to manage enquiries, automate frequently asked questions and provide customer support. The authority lists about 20 organizations and websites that use it, including the state portal eesti.ee, the Estonian Tax and Customs Board, the Estonian Health Insurance Fund, the Police and Border Guard Board, the Consumer Protection and Technical Regulatory Authority and a legal information bot of the Ministry of Justice and Digital Affairs.
No outcome disclosed.
Gemeente Tilburg
Netherlands · Government and public sector · 2026
The municipality of Tilburg uses Vragen.AI, an AI search and answer function supplied by Swis, on its websites such as Tilburg Helpt, registered in the Dutch national algorithm register with a start date of September 2026. Answers come only from information already on the connected sites and always cite their source; personal data that residents type by mistake is detected and removed, and questions are not stored to help returning visitors or passed on to the AI companies. Staff do not watch conversations live but review the answers to improve the website. The register names the spreading of wrong information as a major risk. Several other Dutch municipalities (for example Doetinchem and Wijchen) register similar website chatbots grounded in their own pages.
No outcome disclosed.
Driver and Vehicle Licensing Agency
United Kingdom · Government and public sector · 2025
The Driver and Vehicle Licensing Agency (DVLA), which handles driving licence and vehicle enquiries such as vehicle tax, asks phone callers to say what their enquiry is about and routes them by intent to a recorded answer, an SMS link to a GOV.UK service or the right adviser. On web chat, a bot answers general enquiries with scripted responses and gathers diagnostic information before handing over to an adviser. Both run on Google Dialogflow inside the Content Guru Storm contact centre platform. The records state that neither uses generative AI to write answers (every response is configured by DVLA staff) and that neither makes significant decisions about customers. The natural language IVR cut the average time a caller spends in the IVR menus by half (90 seconds), a navigation time rather than an adviser handling time.
- Containment rate: about 20%, web chat enquiries
"Automation of around 20% of customer enquiries on the web chat channel through bot generated auto responses."
Claimed by: organization - Users served: about 300,000, per month, web chat bot
"Around 300k customers access the chat bot every month."
Claimed by: organization - Time saved per task: about 2 minutes, adviser chat handling time per handed over chat
"Reduction in an advisor's chat handling time of around 2 minutes linked to chat bot data capture (minimising the advisor time asking for this)."
Claimed by: organization - Users served: about 900,000, per month, natural language IVR
"Around 900k customers access the IVR every month."
Claimed by: organization
Government Digital Service
United Kingdom · Government and public sector · 2025
GOV.UK Chat is a retrieval augmented generation assistant in the GOV.UK app that answers questions from GOV.UK guidance only, with links to the source pages under every answer. It retrieves from a vector index of roughly 100,000 GOV.UK pages, uses Claude Sonnet 4 on Amazon Bedrock in the EU, rejects questions that contain common personal data patterns, and runs every answer through a second model check on advice, language, tone and quality before the user sees it. The transparency record describes a limited test with up to 2,000 users over four weeks; it makes no decisions about users.
No outcome disclosed.
Foreign, Commonwealth and Development Office
United Kingdom · Government and public sector · 2024
British nationals abroad who write to the FCDO first see a triage tool that matches their question to approved guidance templates, such as how to replace a lost passport or what documents are needed to marry abroad. Personal data is removed with Azure services, a model splits the message into component questions, an embedding search shortlists templates and a second model picks the best template identifier, so the user never sees generated text. Users can still send the written enquiry to staff, and the FCDO says it will review each such case against the tool's answer; urgent cases are pointed to the consular contact centre. The record lists the tool's phase as beta or pilot.
- Accuracy: at least 76%, offline test on batches of historic anonymised enquiries (76% to 81%)
"Depending on the batch, the tool provided the correct response for 76% to 81% of enquiries."
Claimed by: organization
Abu Dhabi Government (TAMM)
United Arab Emirates · Government and public sector · 2025
TAMM is Abu Dhabi's single platform for about 950 government services from many entities, from car registration and visa renewals to traffic fines. The platform, including its AI assistant, is powered by Azure OpenAI Service and G42 Compass 2.0 (which also gives access to the Arabic JAIS model). The assistant answers questions about processes, shows the status of a user's requests and speaks several languages. A photo reporting feature lets residents photograph a problem such as a pothole or a broken traffic light; the assistant helps fill in the report and updates the reporter on the repair.
No outcome disclosed.
Government of the City of Buenos Aires
Argentina · Government and public sector · 2024
Boti is the city's WhatsApp assistant for residents and visitors, launched on WhatsApp in 2019, which the city describes as the first government in the world to use WhatsApp as a contact channel with its citizens. Its services include appointments for procedures such as driver's licence renewal, public transport schedules, requests such as bulky waste collection and, during the pandemic, vaccination appointments and test results. In 2024 the city added a generative AI experience built on Azure OpenAI Service, scoped to tourism so the team could experiment without sensitive data. The city stresses tone (River Plate voseo, inclusive language) and a single central repository of government information as the basis for grounded answers.
- Interactions handled: at least 2 million, queries per month without human intervention
"Boti, the AI-powered chatbot, handles over 2 million queries per month without human intervention."
Claimed by: vendor
Madrid Destino
Spain · Government and public sector · 2024
Madrid Destino, the city's municipal tourism office, runs VisitMadridGPT, a virtual assistant that answers visitors in more than 95 languages from the city's official, expert curated tourism site, available when physical offices are closed. The city analyses the questions to find the most requested topics and adjust its website content. According to the story, Madrid attracted 10.6 million visitors in 2023.
No outcome disclosed.
Montgomery County Government
United States · Government and public sector · 2024
Montgomery County first launched Monty to relieve its 311 hotline during the pandemic, with 20 topics, and retired it when demand fell. Monty 2.0, built with Zammo.ai on Azure OpenAI Service and Azure AI Search, answers questions on more than 3,000 topics, with automatic translation into 140 languages, from the county's own knowledge base, and uses the county's geographic data to give address specific answers such as trash pickup days. It went through a seven month beta with a constituent focus group before the full launch in late 2024.
- Interactions handled: at least 20,000, since the beta deployment
"Since its beta deployment, Monty 2.0 has facilitated more than 20,000 constituent conversations, achieving a 50% customer satisfaction rate and reducing unanswered queries from 35%–45% to just 10%–15%."
Claimed by: vendor - Customer satisfaction: 50%, since the beta deployment
"Since its beta deployment, Monty 2.0 has facilitated more than 20,000 constituent conversations, achieving a 50% customer satisfaction rate and reducing unanswered queries from 35%–45% to just 10%–15%."
Claimed by: vendor
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- A current, owned corpus of official guidance with publication and review dates
- Contact reason data from the contact centre to choose the first topics
- A set of real questions with approved ideal answers for evaluation
Systems to integrate
- Content management system or publishing feed, so the index updates when guidance changes
- Contact centre or webchat platform for handover with the transcript
- Web, app, messaging and telephony channels
- Analytics for unanswered questions and feedback
Complexity: Low
A grounded information assistant touches no transactional systems, so the build is mostly content and evaluation work: cleaning the guidance corpus, deciding what it must refuse and testing accuracy by topic. Complexity rises once it reads personal case data or books services.
- 1
Choose topics by volume and harm
Start with high volume, low harm information topics (opening hours, which form, how to apply) and keep personal case questions, legal advice and urgent needs out of scope at first.
- 2
Clean and filter the content
Index only official pages, exclude documents likely to contain personal data, and chunk by heading so answers can cite the exact section. Connect the index to publishing so it updates daily.
- 3
Write the refusal and routing rules
Decide what the assistant must never do (give personal advice, guess eligibility, answer outside the content) and where each type of refused question goes.
- 4
Build an evaluation set before launch
Collect real questions with ideal answers per topic, add adversarial and jailbreak prompts, and score groundedness, accuracy and completeness on every change.
- 5
Pilot with a capped group
Run a limited test (the GOV.UK Chat transparency record describes a test with up to 2,000 users over four weeks), review transcripts, then widen channel by channel.
- 6
Measure unanswered questions, not only usage
Track refusals, handovers and repeat contacts per topic and feed gaps back to the content owners.
Guardrails
- Answers only from retrieved official content, with a link to every source and a refusal when retrieval finds nothing relevant
- Personal data detection on input, with masking or rejection before text reaches a model
- A second check on every answer for unsupported claims, advice and tone before display
- Clear AI disclosure and a visible route to a human or the official phone line
- No decisions about eligibility, entitlement or enforcement
KPIs to instrument
- Share of questions answered from content, refused and handed over, per topic
- Accuracy and groundedness on a weekly human reviewed sample
- Repeat contact on the same topic within seven days
- Satisfaction and feedback on answers, compared with search and phone
- Contact centre volume on the topics in scope
Human in the loop
Content owners approve the corpus and the topics in scope; a service team reviews samples of conversations every week, with extra review for refusals and complaints. Advisers take every handover with the transcript, and changes to scope go through the same sign off as a change to published guidance.
Common failure modes
- Confident answers that break the law
- New York City's MyCity chatbot was reported in 2024 to tell businesses they could do things that are illegal. Ground strictly, refuse outside the content and test legal edge cases before launch.
- Stale guidance in the index
- The assistant repeats last year's rules after a policy change. Tie the index to the publishing pipeline and show the source date.
- Personal questions answered generically
- Residents ask about their own case and get a generic answer that sounds personal. Detect case questions and route them to the authenticated service.
- Usage mistaken for success
- High conversation numbers can hide abandonment. Measure unanswered questions and repeat contacts. Microsoft's case study on Montgomery County's Monty 2.0 reports a drop in the share of unanswered queries from between 35% and 45% to between 10% and 15%, next to its conversation count.
What are the risks and rules?
EU AI Act
Limited risk (transparency)
An information assistant must tell people they are interacting with AI (Article 50). It is not high risk as long as it does not evaluate eligibility for public assistance benefits or services (Annex III point 5(a)); an assistant that starts to pre assess eligibility should be reassessed.
Rules that apply
Guidance
- Article 50, transparency obligations for providers and deployers of certain AI systems (European Union, Europe). People must be informed that they are interacting with an AI system unless this is obvious from the context.
- AI Playbook for the UK Government (UK Government, Europe). Guidance for civil servants and people working in government organisations on using AI, including generative AI, safely, effectively and securely.
- Algorithmic Transparency Recording Standard Hub (Government Digital Service, Europe). Mandatory for all UK government departments and for arm's length bodies that deliver public or frontline services; the GOV.UK Chat, DVLA and FCDO records on this page were published under it.
- M-25-21: Accelerating Federal Use of AI through Innovation, Governance, and Public Trust (US Office of Management and Budget, North America). Requires US federal agencies to inventory their AI use cases at least annually and to apply minimum risk management practices to high impact AI.
Controls to put in place
- AI disclosure and a statement that answers may be wrong, with links to the official source
- Entry in the public AI or algorithm register (for example ATRS in the UK or the Dutch algorithm register)
- Data protection impact assessment covering question logs and retention
- Red teaming for jailbreaks and legal edge cases before launch and after model changes
- Retention limits on conversation logs and access control for reviewers
When it went wrong elsewhere
- NYC's AI chatbot tells businesses to break the law. In March 2024 The Markup found New York City's MyCity business chatbot giving answers that contradicted local law on housing, worker rights and consumer protection, for example that bosses could take workers' tips.
Frequently asked questions
- Can a government chatbot use generative AI safely?
- Yes, if it answers only from official content and shows its sources. GOV.UK Chat retrieves from GOV.UK pages, rejects questions that contain common personal data patterns and runs every answer through a second model check; FCDO's consular triage goes further and only picks from approved templates, so users never see generated text.
- What share of questions can it handle?
- Reported figures are modest but real. DVLA automates around 20% of web chat enquiries, and FCDO's triage picked the correct response for 76% to 81% of historic enquiries in testing. Treat any number as topic specific and measure it on your own questions.
- Is a citizen information chatbot high risk under the EU AI Act?
- Not by default. It carries the Article 50 transparency duty. It becomes high risk when it evaluates eligibility for public assistance benefits or services, which is why eligibility questions should route to the official service.
- Should every agency build its own assistant?
- Not necessarily. Estonia's Bürokratt is one assistant, developed under the Information System Authority, that about 20 public bodies and websites use, from the state portal to the Tax and Customs Board. A shared platform means each agency does not have to build, secure and evaluate its own assistant.
How to cite this page
Blits.ai AI Use Case Library, "AI assistant for citizen information and government services", last verified 27 September 2026, https://www.blits.ai/ai-use-cases/citizen-information-assistant. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 27 September 2026: First published