AI for cities – the BRISE Example

Currently, artificial intelligence approaches are largely based on the processing of very large amounts of data, artificial neural networks, or self-learning algorithms. Especially "deep learning" and "artificial neural networks" are currently attracting a lot of attention in the public perception. Both refer to pattern recognition, which analyses data to identify regularities, repetitions or similarities. IT systems are often capable of analysing much larger amounts of data in less time than would ever be possible for a human being. In addition, IT systems do not suffer from fatigue or careless errors. One advantage of pattern recognition systems is the identification of correlations that were previously neither known nor noticed by humans. This means that data cannot only be checked for previously defined correlations (what statisticians used to do), but also evaluated in an open-ended way.

In the meantime, the distinction between weak AI, strong AI and superintelligence has become established in science. Weak AI is usually developed and used for specific applications, for example expert systems, speech recognition, navigation, or translation services. Applications based on weak AI are already in widespread use today and can be found in everyday life in the form of intelligent search suggestions or optimised route calculations, for example. With reference to Howard Gardner's theory of multiple intelligences,[1] weak AI primarily describes the emulation of linguistic and logical mathematical intelligence. In contrast, strong AI furthermore describes systems that can think logically, plan, learn and make decisions under uncertainty on their own.[2]

AI is a technology which is increasingly capable of taking own decisions and replacing humans in work processes and in decision making. This calls for a strong governance of AI to make sure we stay on top of its impact and remain the decision makers when it comes to issues affecting us directly. Municipalities are the carriers of the public good. They have authority to take decisions with direct and real impact on the lives of individuals. Thus, it is imperative to deal with this question before putting AI to work in a city context. Many scientific publications have been written about this issue. Here let us only state that the use of AI in conformity with fundamental rights means that we only delegate decisions to machines that do not require human judgement and do not entail any democratic and ethical problems.

AI and Machine Learning (ML) offer enormous opportunities and possibilities for municipal organisations where boring and error-prone routine tasks can be delegated to machines, leaving people more time for creative and meaningful tasks.

In addition to the potentials of the introduction and use of cognitive technologies and processes, the focus must be on the training of employees for understanding the potentials and risks of AI.

Already today AI can be used for a range of different types of tasks in differing municipal contexts. Basic AI technologies and applications have shown in the past, which human skills can be replaced or enhanced by AI to deliver better services. A study by Fraunhofer IAO[1] has identified five areas of application for AI in the municipal administration: a) Front office management, b) Background process management, c) Decision Support, d) Decision Automation and e) Real-time decision making.

Let us take a brief look at each of them:

Numerous digitisation projects in public administration are currently focusing on direct contact with citizens. Often there are too many touchpoints for citizens or businesses to interact with the city administration and it takes too long until the right officer has been identified. Often citizens even need multiple accounts, numbers, and identities to interact with the different municipal service providers (waste, energy, mobility, housing etc.). Once the location of the case within the organisation has been achieved, lengthy and complicated forms and untransparent processes take too much of the valuable time from citizens and businesses. In all these points AI can help. Already today there are a range of AI-based services that improve citizen interaction with the municipality:

• Chatbots and personal voice assistants
• Service robots as digital assistants on site
• Single Sign-on and Once Only Identities
• Application support tools

As the size and complexity of an organisation increases, so does the need for support processes in addition to the actual production of services. These processes do not generate any added value themselves, but they make the actual desired processes possible. In a thematically broad and complex organisation such as a public administration, support processes take up a not inconsiderable share of resources. To use the available resources as far as possible for the relevant processes, there is a great interest in simplifying or automating the support processes. AI can help to smoothen this:

• Support for ongoing transaction processing: workflow management
• Automatic performance of standard procedures and processes 
• Personnel management: Recruitment

In public administration, many different decisions are made with legal binding force. The actual decision always consists of the weighing of possible alternatives and the final decision on one alternative. A decision, despite all the use of technical aids and procedures, has always been an internal human process, regardless of whether decisions were made authoritatively, patriarchally, collegially or collectively.

While some decisions can be made based on single, defined and measurable criteria, others require a large amount of data and information to be evaluated and weighed against various criteria. In the context of evidence-based government, ideally almost in real time, decisions based on data and facts are currently gaining strongly in importance, even if other actors want to discredit decision-making bases by deliberately spreading false facts. However, this does not mean that administrative decisions are purely a matter of analysing and evaluating facts and figures. Especially in complex cases, where several criteria must be weighed individually and discretion must be exercised, human decision-makers have so far been superior to technical systems.

Acknowledging this, the interaction of humans and machines can bring the respective strengths together and support the human processors with technical systems. In this context AI can be used for the following purposes:

• Use of existing “data treasures” and new types of smart data bases
• Intelligent planning and predictive maintenance
• Decision Control Radar
• Recommendations and forecasts

In addition to supporting the decision-maker, artificial intelligence can also be used to automate decisions. This means that the human is removed from the decision-making process and the binding decisions are made autonomously and thus exclusively by a technical system.

If an administrative process is automated, IT systems thus make partial decisions or the full decision independently. This means that the administrative staff, depending on the extent of partial or full automation, are no longer involved in the processing to the same extent as before. This primarily has an impact on the level of the decision-maker. While the (partially) automated tasks are thereby eliminated, new tasks for monitoring and controlling the systems are also added at the same time, as well as possibly additional work due to the automatically generated results.

However, this also changes the job profiles of administrative staff. Ideally, the aim is to relieve employees of burdensome, monotonous work. In each individual case, it must be examined which changes occur through partial and full automation, which new activity and requirement profiles arise and in which way it must be reacted to. It may also be necessary to weigh up several factors here instead of implementing what is technically possible. In addition, transparency about the functioning and results of the decisive systems must be ensured to be able to adequately review their activities. AI can help municipalities with the following approaches:

• Simple bound decisions based on existing regulations
• Complex discretionary decisions
• Automatic authorisations in the tax administration

In addition to administrative processes where decision-making and execution can be easily realised within minutes, hours or days, other administrative decisions require near real-time decision-making and implementation. On the one hand, this is related to the need for an immediate reaction, such as a reaction in road traffic in the case of traffic light controls or autonomously driving vehicles. Wrong decisions can endanger human lives. On the other hand, the impact of processes can also be greatly improved. Instead of waiting several days after traffic offences, real-time processing enables law enforcement to react promptly with appropriate warning and financial punishment. Latency times in the range of milliseconds or seconds, combined with high computing power, make it possible to respond to changes in near real time. In many cases, this makes it possible to react immediately to the causes instead of waiting for a long time for the visible consequences, thus protecting property and human life. AI is already applied for use cases such as:

• Traffic management
• Disaster management
• Hazard prevention

The UIA funded project BRISE Vienna uses AI to significantly accelerate the process of issuing building permits. In this project, AI is one of several technologies which are used to inform and support a fully digitized process, linking the planner and the municipal building authority through various digital tools.

In BRISE Vienna, AI is used as a decision support system and as for Background Process Management (Category B and C of the above classification). It helps the staff of the Vienna building authority MA37 to quickly assess, whether a submitted plan for a new building is in line with all existing regulations and requirements. This verification process used to be a manual task and it required review of a range of specific regulatory documents which are specific to the extent of individual sites (e.g. specifying requirements for green roofs, shadowing, terrasses, etc.). These “textual requirements” are part of the formal land-use plan of Vienna and they are attached to it as PDF documents.

The AI-supported verification process in BRISE Vienna automatically analyses all pdf documents linked to a specific site, it then extracts the relevant information in a machine-readable way and provides it as input to the IFC-based digital reference model of the building, against which the submitted plan is being checked.

A textual requirement could read as follows, for example:

“In streets with a breadth of less than 16m, bay windows and protruding loggias may only reach beyond the building line to a maximum of 0,8 m.”

The AI now needs to understand and classify the terms “bay window” and “loggia” and it needs to interpret the additional information “0,8 m” as “MaxDistance”.

To enable the AI to perform these tasks automatically, a manual training of a machine learning program had to be undertaken first. To this end, the staff of MA37, supported by a company specialized on AI, selected and analysed a range of textual requirements and manually defined six categories of requirements with reference to the Vienna building code. The MA37 employees then analysed a sample of 100 PDFs manually and classified the relevant terms into the six categories.

An automated codification of text into numbers was applied to enable statistical learning operations. After this, a range of statistical models were put into place to enable Machine Learning based on the training data provided by MA37.

As a result, the AI is now able to automatically analyse and classify all requirements specified in all legal documents that are linked to the land use plan. From here this information can be transformed into a machine-readable file, which informs a 3D reference model for the site, and it can be made available as decision making support within the verification tool used by the employee of the Vienna building authority. All in all, this saves a significant amount of time throughout the verification process, since all of these steps were undertaken manually before.

A second area of application of AI within BRISE Vienna refers to the automated search of similar law cases.

The authorities of the city of Vienna – like all other municipal authorities – are obliged to take decisions within the boundaries of the current law. In case of a lack of clarity or ambiguiry in the legal provisions, additional documents such as written interpretations must be consulted. These can either stem from past law cases in court or from internal commands of the municipal authority. When analysing and verifying a building application, employees of MA37 often reach the limits of their active knowledge, since the overall number of relevant laws, court decisions and commands is very high. In this situation, in the past, the staff needed to consult the Austrian legal information system RIS (Rechtsinformationssystem), which publishes legal provisions and court decisions. The RIS supports a search based on keywords, but it often provides unsatisfactory results, since the word being searched for was spelled incorrectly or is not contained within the legal text.

This problem was solved by an AI-supported semantic search function. After automatically transferring all documents from PDF to text, a semantic model could be applied which is able to grasp the meaning of a word and relate it to similar or linked terms. With this solution, employees of MA37 can now easily find all relevant past court decisions, law cases, legal provisions, and internal commands in one place. It helps them save a lot of time which they used to spend on researching various data sources with inadequate tools before.

A third application of AI in BRISE Vienna refers to an automated check of submitted documents. In the past, verification staff had to open all submitted documents to check, whether signatures have been provided at the required places. An automated analysis of signatures (based on AI object recognition models) helps prevent this step and shows at one glance, whether all signatures have been provided at the right place in the document. Again, this saves time for the building authority to concentrate on more complex tasks.

Widespread adoption of AI in the public sector is not trivial. Much of the knowledge about administrative processes, concrete procedures and decision-making patterns is only implicitly anchored in the minds of administrative staff. For a sustainable design of AI-based processes, this knowledge must be preserved and new competences in the field of AI must be built up. So far, such competences are often still bundled in those companies that develop and implement AI-based applications for the administration. In the future, care should be taken to jointly build up and bring together the necessary competences. For administrative authorities, this means successively creating basic competences in the field of AI.

On the other hand, it seems obvious that not every municipality and every public authority should deal with all the basic questions about AI in detail. A coordinating body for the use of AI in the public sector would help to ensure a comprehensive transfer of knowledge at all administrative levels. State and local authorities need a reliable legal framework as well as technical and organisational support and training. They can provide an important impetus for the identification and harmonisation of suitable procedures as well as for the derivation of procedural models.

The example of BRISE shows that there is tremendous potential even in smaller AI based applications, since they can take over routine tasks that used to consume a lot of time in the past.