---
title: "What AI projects in Spanish local councils teach us"
description: "Customer service assistants, technical reports, loneliness detection, chatbots, request classification and phone support: what several local authorities did and the lessons learned."
url: https://bidatia.com/en/ideas/proyectos-de-ia-en-ayuntamientos-lecciones
lang: en
date: 2026-07-28
updated: 2026-11-04
author: "José Luis García"
keywords: ["ai in local government", "ai local councils", "public sector ai use cases", "council chatbot", "femp ai guide", "ai local authorities", "ai in municipalities"]
publisher: Bidatia
---

# What AI projects in Spanish local councils teach us

> Customer service assistants, technical reports, loneliness detection, chatbots, request classification and phone support: what several local authorities did and the lessons learned.

## Executive summary

- Brings together six AI cases in Spanish local authorities published by the FEMP (Avilés, Madrid, Las Rozas, Valencia and Alcobendas) and the lesson from each.
- Recurring patterns: assistants over in-house documentation, support for technical reports, classification and routing, and proactive or voice contact.
- Common lessons: data quality determines the outcome, human oversight must be measured, explainability makes justification possible and logs make it possible to respond to complaints.
- The real cost is not the cost of the pilot, and there must always be a human alternative.
- Includes a six-month roadmap to get started, questions to ask before the first project and common mistakes.

The Spanish Federation of Municipalities and Provinces (FEMP) brings together real experiences from city councils and provincial councils in its *Guía práctica y políticas de uso de la IA en las entidades locales* (practical guide and AI use policies for local authorities, 2025). Taken together, they reveal a very clear pattern: the projects that work start from a specific problem, take care of the data and keep people at the centre of decision-making.

## Six cases and their lesson

| Case | What was done | Lesson |
|---|---|---|
| **Avilés: citizen service assistant** | An AI assistant for registry and front-desk staff, based on a knowledge base with the council's documentation, to give consistent answers regardless of who is attending | A single documentary source improves consistency across channels and makes it easier to audit answers. Repeated questions reveal poorly designed procedures |
| **Avilés: minor works permit reports** | Support for drafting technical reports for around 1,200 permits a year, using a cloud model with no specific training. It estimates savings of 20 % to 45 % in the technical process | The savings come from the machine validating and the officer reviewing, not from replacing the officer. Functions such as correcting applications come close to acts with legal effects and must be assessed as such |
| **Madrid: detecting unwanted loneliness** | Conversational AI that called older people with a questionnaire. Of 2,071 completed conversations, 646 asked to speak to council staff and around 240 joined a befriending programme | AI detects and refers; people decide and provide support. With vulnerable groups, impact assessment and transparency are essential |
| **Las Rozas: generative AI chatbot** | A multichannel, multilingual assistant based on the council website and documents, for around €7,000 a year within the telephone helpline contract. It estimates 50 % less handling time | Improve the sources, test extensively and accept some failures. The quality of the documentation and the test bench are decisive |
| **Valencia: routing general requests** | Automatic classification of uncategorised requests to send them to the right unit, with an acceptable margin of error | Validate by sampling, keep human review and measure the error |
| **Alcobendas: automated phone support** | A voice bot to handle out-of-hours calls, with call-backs by staff, arising from an innovation call and paid per use | Always offer the human alternative. Public procurement of innovation makes it possible to test before committing to a large contract, but the cost at real scale must be estimated |

## The lessons that recur

**Data determines the outcome.** The FEMP notes, citing expert opinion, that data quality lies behind around 80 % of AI project failures. The warning signs are well known: nobody knows which system a piece of data comes from, the same data has different values in two applications or the document base contains outdated versions.

**Human oversight degrades if it is not measured.** With high volumes and a system that is usually right, review becomes a formality. Discrepancies between the AI's proposal and the final decision must be measured: if there are never any, that is a sign of dependence, not accuracy.

**What cannot be explained cannot be justified.** Better to use solutions that cite their source (RAG) and documentation in plain language of what the system contributes in each case.

**Without logs there is no way to respond to a complaint.** Keep the input, output, model version and who validated it, and secure access to those logs by contract.

**The free pilot is not the real cost.** Recurring costs (licences, cloud consumption) must be estimated before the pilot, and the tender specifications must include exit from the provider without data loss.

**Citizens may take an answer as the official position.** Answers for guidance only with a link to the source, reminders of deadlines and procedures, and a human channel that is always visible.

**Start small.** Well-defined, low-risk cases, a mixed team of technical, legal and service staff, and training from the outset.

## What type of project each one is

The cases follow four patterns that recur in almost any organisation:

| Pattern | Cases | Usual risk | Key to success |
|---|---|---|---|
| Assistant over in-house documentation (RAG) | Citizen service, web chatbot | Answers taken as the official position | Quality and currency of the documentation, citing the source |
| Support for technical reports | Building permits | The officer stops reviewing | Real oversight and measuring corrections |
| Classification and routing | General requests, registry | Errors that delay procedures | Validation with samples and human review |
| Proactive or voice contact | Unwanted loneliness, phone support | Vulnerable groups and the digital divide | Human alternative and transparency |

Identifying the pattern helps you know in advance which controls are needed.

## A roadmap to get started

1. **Months 1–2: put your house in order.** An inventory of the AI already in use (including AI embedded in applications), basic usage rules, a list of authorised tools and initial training for all staff.
2. **Months 2–3: choose a first case.** Low risk, with a clear, measurable problem: an internal assistant over public or internal documentation, with no personal data, is often a good choice.
3. **Months 3–6: pilot.** With a department involved from the start, a test bench and acceptance criteria set before starting.
4. **Month 6: decide.** Continue, adjust or stop, based on data. If you continue, formal authorisation and publication of the system.
5. **Afterwards: scale up.** The second case is easier: the governance model, the record sheets and the contracts already exist.

## Questions worth asking before the first project

- What specific problem are we solving and how will we know if it has worked?
- Is the source documentation or data up to date, and does it have an owner?
- Who oversees the results and with how much time?
- What happens to people who cannot or do not want to use the digital channel?
- How much will it cost per year at real scale, not in the pilot?
- Can we change provider without losing data or configuration?

## Mistakes that recur

- **Choosing the first case for how impressive it is** rather than how feasible it is.
- **Launching a chatbot on an outdated website**: it will confidently repeat whatever the website gets wrong.
- **Not measuring beforehand**: without knowing how long things used to take, it is impossible to demonstrate savings.
- **Leaving out the staff who will use the system** until the end.
- **Forgetting transparency**: citizens trust administrations more when they explain which systems they use and why.

## How to apply it in your organisation

Before launching a use case, it is worth going through a few minimum questions: what problem it solves, whether AI is needed, what data it uses and on what legal basis, who oversees it and how success will be measured. We explain this in [the lifecycle of an AI project](/en/ideas/ciclo-de-vida-proyecto-ia) and in [which regulations apply to AI and data](/en/ideas/ia-y-datos-normativa-rgpd-ria-ens).

At Bidatia we work with public administrations on data governance and responsible AI projects, from strategy through to implementation.

## Frequently asked questions

### Where can you find examples of AI in Spanish local councils?
The FEMP brings several together in its Guía práctica y políticas de uso de la IA en las entidades locales (2025), with experiences from city councils and provincial councils.

### How much does a generative AI council chatbot cost?
It depends a great deal on the scope. One of the cases collected by the FEMP costs around 7,000 euros a year within the telephone helpline contract, but the recurring cost at real scale should always be estimated.

### Can a council chatbot resolve procedures?
It can inform and guide, but its answers must be for guidance only, link to the official source and offer a human channel. Resolving without human intervention would be an automated administrative action with specific requirements.

### Why do AI projects fail in the public sector?
Mainly because of data quality, human oversight that becomes a formality, recurring costs that were not estimated and failing to involve the service that will use the system from the start.

### What is public procurement of innovation?
A procedure that allows public administrations to contract solutions that are not yet on the market, or to test them before committing to a large contract. Several municipal AI projects started this way.