AI in public authorities and administrations: getting started

Generative artificial intelligence is an ideal tool for public authorities and administrations given the large volume of documents they deal with. But the challenges are different from those faced by many companies. Approaches to getting started.

Dusty offices? Simply set foot in any public administration building today and you’ll quickly find that this cliché does not hold water. Modern PC workstations predominate. If you’re looking for stacks of files or cabinets brimming with folders, you’ll have to go down to the archives. Still, the fact remains that administrations work with a huge number of documents: forms, ordinances, minutes, decisions, permits, tax returns – the list goes on and on.

Generative artificial intelligence, also known as generative AI or GenAI, is therefore an ideal tool for public administrations. Its large language models (LLMs) such as OpenAI’s GPT (on which Microsoft’s Copilot is also based), Anthropic’s Claude and Meta’s Llama are distinguished by their ability to ‘understand’ and process texts written by humans. ‘As public administrations work particularly intensively with unstructured texts in the form of e-mails, objections and meeting minutes, they can particularly benefit from the support of GenAI applications,’ says Christof Zogg, Head of Business Transformation at Swisscom.

Such support can have a major impact on office work and increase efficiency by eliminating tedious routine tasks. ‘Summarise all unread e-mails and make me a to-do list’ is much quicker and more satisfying than digging through countless unread e-mails after some time off. ‘Create a template for an approval based on this document’ or ‘translate these minutes into French’ are other examples of how generative AI can help save time.

In its paper Governing with Artificial Intelligence: Are Governments Ready?, the OECD has also identified increased productivity and, as a consequence, improved service delivery for ‘customers’ such as individuals and companies as a major benefit of generative AI in public administration.

However, the starting point is different from that of many private companies.

AI: the legal and ethical context

Administrations have access to a large amount of private information and personal data, from tax returns, injunctions and permits through to court decisions and psychiatric assessments. This also includes ‘sensitive data’ as defined in the Swiss Federal Act on Data Protection (FADP). Privacy must be protected under all circumstances when using generative AI, and erroneous or even discriminatory results in decision-making processes must be prevented. As Christof Zogg explains, this is a special starting point: ‘The requirements in respect of data protection and, in the future, AI governance are particularly high for the public sector. That’s why it’s important to identify suitable use cases for early AI innovation projects – and these do exist.’

A study commissioned by the canton of Zurich addressed ethical and legal issues surrounding the use of AI in public administrations. And for the Swiss federal administration, the Federal Council has adopted guidelines on the use of AI that emphasise transparency and security.

Examples of use of (generative) AI

Even if legal and ethical requirements are taken into account, AI and generative AI can save time and boost efficiency in the day-to-day handling of documents and information. Here are possible public administration use cases with increasing levels of technical requirements:

Use case 1: Meeting summarisation and transcription

Generative AI simplifies the creation of documents such as meeting minutes, resolutions, etc. and can translate these into numerous languages. Recordings of (online) meetings can be transcribed with AI and saved as text.

The results should always be treated as a draft and checked for correctness and completeness in order to identify and correct ‘hallucinations’ (incorrect outputs by generative AI systems).

Use case 2: Creation of drafts for directives, decrees, etc.

Generative AI can create templates based on existing texts for various official documents, which then do not need to be written from scratch.

Use case 3: Document search bot

AI makes it possible to search documents and retrieve information from them. Such search bots are suitable for internal purposes (‘enterprise search’) as well as for publicly accessible documents such as building information for architects, resolutions, proposals, etc. One example (implemented by a private company) is ZüriCityGPT.

Use case 4: Trend forecasting, e.g. for urban planning

AI can use existing data to forecast population, traffic and other trends. This helps in urban planning when deciding on school buildings, public transport connections and so on. The information can also be used as a basis for political decisions, for instance when it comes to budgets for public construction projects. One example is the Smart City LuzernNord.

Such use cases are complex and require machine learning models to be trained with internal data.

‘The requirements in respect of data protection and AI governance are particularly high for the public sector. That’s why it’s important to identify suitable use cases for early AI innovation projects – and these do exist.’

Christof Zogg, Head of Business Transformation, Swisscom

First steps towards implementation: approaches to AI

Broadly speaking, there are three areas of application for (generative) AI that can be equated with rising levels of difficulty or increasing complexity according to the use cases:

Approach 1a: Evaluation and introduction of standard AI services such as Microsoft Copilot for M365. Off-the-shelf services like these can be implemented quickly, the provider takes care of the operation and governance is clarified.

Approach 1b: Licensing/activation of AI functionality in existing software. This involves using the AI functions of standard software, for example for automated invoice processing in an ERP such as SAP or Abacus. The effort is similar to Approach 1a.

Approach 2: Configuration or development of an in-house (generative) AI process application. Existing processes are automated using pre-trained multimodal foundation models such as GPT or Claude. One example is the pre-processing of standardised applications, for example in the construction and asylum sectors. The effort required for this is medium, because the models have to be re-trained with in-house data for a specific task, for example by means of retrieval augmented generation (RAG).

Approach 3: Training and operation of an entirely in-house AI model. Here, a business unit develops and trains its own model for a specific application, which is precisely tailored to the respective needs. Examples include forecasting demographic trends or detecting fraud based on specific patterns. These approaches, which often use ‘traditional’ AI such as machine learning, require significant development and operational effort.

Getting started with generative AI

The use of generative AI needs to be well planned. The technical and legal requirements must be clarified, and consideration must be given to data protection guidelines and employee training.

It often makes sense to start with projects that are easy to implement (Approach 1). ‘The technical maturity of the administration will determine the recommended approach and selection of initial useful AI deployments,’ says Christof Zogg. ‘But it’s already worth it if a few smart power users start creating illustrations for the internal newsletter using GenAI image generation tools, and thereby getting the ball rolling on the long journey into the age of artificial intelligence. This is definitely more of a marathon than a sprint.’ Alternatively, employees who already have experience with ChatGPT could form a pilot group to use Copilot.

The question of data storage and processing plays a role for in-house AI applications: depending on the sensitivity of the data and legal/regulatory requirements, the AI services of a hyperscaler or a Swiss offering such as the Swiss AI Platform might be considered. In any case, it’s important to start with an overview: how would AI benefit us most? Where do we want to start? This will open up potential applications in which the use of AI is sure to be advantageous – and blow the last speck of dust from the office.