Management of AI-driven digital service development

The management of software development has changed significantly in recent years and is expected to undergo further transformation in the future (Ozkaya 2023; Sauvola et al. 2024). The work of programmers has evolved considerably due to AI-assisted coding, no-code development, AI agents, and the use of AI-powered tools. As AI becomes a more independent actor due to the rapid development of agentic AI, its roles, privileges, and responsibilities should be carefully defined (Paasonen 2026).

Similarly, advances in AI have transformed the software development process, with AI models increasingly being incorporated into traditional software products and services. This creates new competency requirements for developers and designers, as well as for product managers and IT consultants.

This article reviews the development processes through which new AI-based digital services are designed, piloted, and implemented. Digital services consist of software products that enable the provision, delivery, and use of services.

Responses to the skills gap in generative AI development

In recent years, Europe has responded to the growing demand for new skills driven by the rapid development of AI technologies and their applications. For example, Haaga-Helia is a partner in the Erasmus+ Generative AI Skills Academy (GenAISA) project, which develops new learning resources for generative AI development.

The project partners developed tailored curricula for higher education and vocational education and training, which were launched for piloting this summer. The partners have also provided open-access learning opportunities through a MOOC platform where students and professionals can develop competencies in aligning business needs with generative AI opportunities. In this work, experts from Haaga-Helia developed the course Management of Generative AI Transformation for GenAISA, providing knowledge and skills to support the adoption of AI in alignment with business needs.

Since AI is developing rapidly and new AI applications are being created weekly, skill requirements are also evolving at an unprecedented pace. Hence, contributing to the field by offering new insights into the management of AI-driven software development, I have structured various AI development approaches and refined their development process models.

Three ways to integrate AI models into software products

The easiest way to integrate AI models into traditional software products is through the user interface (UI) of a digital service or other IT system. Perhaps the best-known examples are chatbots, which many organizations now provide for their users and customers.

Another way to integrate AI models into traditional software systems is through the application logic of the software. They are sometimes called ‘hybrid solutions’. This requires the use of an Application Programming Interface (API), which enables communication between the AI model and the software system. The software system can then use the integrated AI model for specific purposes, such as analyzing and processing data or monitoring workflows.

There are hundreds of thousands to millions of AI models and components available on the market that organizations can use in their own development projects. Some of these are open-source AI models, while others are commercial products. For example, Hugging Face community includes over two million AI models. Large language models (LLMs) can be both commercial and open-source models.

The third approach to applying AI models in software products is to develop an ‘AI-native’ application, in which the core processes and business logics are driven by AI. Such applications are AI-based systems that incorporate few features of traditional deterministic software. Hence, most functions and value creation are executed by probabilistic AI models rather than by deterministic software code.

Classifying software types is challenging, as the appropriate category depends on the type and complexity of the software. However, most AI systems often include traditional deterministic software modules that should be kept deterministic for reliability, controlling and security reasons, such as access rights management, system controls, security functions and user privileges.

Naturally, AI features based on probabilistic language models and deterministic software features may overlap from the perspectives of user interfaces and end users. Nevertheless, some form of classification helps students and professionals understand the main approaches and components that modern IT systems and applications consist of.

Crucial tasks introduced by AI models in software development management

Together with my colleagues (Alamäki et al. 2026b), we have described the new AI-related tasks introduced into the software development process through the integration of AI models. In the article, we identified and described the following nine design and development phases associated with integrating ready-made AI models into digital services and software products:

1. The software development process begins with business needs analysis. During this phase, designers assess the potential value that AI models could provide for new digital services, while also evaluating the associated risks.
2. The second phase focuses on use case definition, during which the role of AI and the required data are specified. It is essential to clearly define both the capabilities and limitations of AI, and the requirements of data.
3. The third phase is requirements specification, where the functional and non-functional requirements of AI components are defined together with information security requirements, AI regulatory requirements, and data quality criteria. Particular attention should be paid to defining the AI model’s role, privileges and responsibilities.
4. The fourth phase addresses architecture and system design, including roles, authority, responsibilities, cybersecurity, system security and trustworthiness as integral elements of the IT system. This phase requires a simultaneous review of Phases 2-4.
5. The fifth phase covers system development, during which the AI model is integrated into the IT-system.
6. The sixth phase is testing, where the trustworthiness and performance of the AI model are evaluated in a secure environment.
7. The seventh phase is implementation, which places particular emphasis on monitoring the trustworthiness of the AI model in an actual usage environment.
8. Maintenance involves continuous monitoring and evaluation of performance and trustworthiness, as well as the potential renewing or retraining of the AI model.
9. Further development includes replacing the existing AI model with newer versions or introducing significant improvements.

We have also described AI-related software development processes using a five-phase model (Alamäki, Khan & Lagstedt 2025). This model focuses on AI lifecycle management, where the final stage of the AI lifecycle is retirement, in addition to the design, development, deployment, and operation phases. The model distinguishes between actions, actors, and ethical considerations.

In another article, we describe AI adoption from the perspective of AI tools in SME companies, particularly in the restaurant and tourism sectors (Alamäki, Hiekkanen, & Laintila 2026a). The model shows that AI adoption creates new workflows and working practices in which AI tools take over some tasks that have traditionally been performed manually.

The actors involved in AI development include professionals such as software developers, machine learning scientists, data scientists, domain experts, system integrators, end users, and third-party suppliers (Clement et al. 2023; Alamäki et al. 2025).

Designing and developing a new AI-native application

‘AI-native’ software design and development differ from the integration of AI models into traditional software products. In this context, innovation and proof of concept evaluation are more important, as many situations remain uncertain. Therefore, an iterative development process is needed, in which designers learn about product requirements through the experimental development of minimum viable products (Ries 2010). Similarly, involving end users and stakeholders in the development process and testing the application in its actual usage environment are important (Alamäki & Dirin 2014).

Table 1 illustrates the design and development process in which designers and developers iteratively create a new AI-based application or solution. In this process, not all requirements and features of the final application are clear from the outset; therefore, an innovative and experimental approach to the design process is required.

Table 1. The iterative design and development process for creating a new AI-based application.

The concept planning phase is the starting point in which the overall concept is designed, including business needs, use case analysis, data requirements, and functional and non-functional requirements. Business needs define the actual needs to improve specific business processes, while use case analysis describes how users interact with the application and what features are expected throughout the workflow or end-user journey.

Data requirements specify the types of data needed for the application. Functional requirements provide detailed specifications for the required features of an AI application, whereas non-functional requirements define qualitative and quantitative characteristics, such as excellent usability, security, cost efficiency, accuracy, and other specific attributes of the AI application.

The AI system design phase focuses on the building blocks of the technical application. While the concept planning phase addresses user and business needs, the AI system design covers technical requirements and specifications. First, designers must define the role of AI, including its added value, expected outcomes, tasks, and specific features. Technology selection outlines the technical components of the application, such as the user interface, database type, APIs, and operating system provider.

The selection of an AI model determines which algorithms, and AI components will be used, such as the type of machine vision, the provider of large language models (e.g., Claude, MS Copilot, ChatGPT or Mistral), or the other type of machine learning model, if LLMs are not applied. System architecture and cloud infrastructure refer to the IT environment in which the AI application will be deployed, hosted, and maintained (e.g., Microsoft Azure, Amazon Web Services AWS, Google Cloud Platform or other servers).

The proof-of-concept (POC) phase includes the development of a minimum viable product (MVP). This may take the form of a simple and non-functional sketch or demonstration, or functional prototype developed using a software development platform (e.g., Lovable, Claude Code, or another AI-based tool).

Depending on the technical capabilities of the POC team, the performance and trustworthiness of existing LLMs can be improved either by utilizing the RAG (Retrieval-Augmented Generation) method, or, requiring more resources and expertise, through training or fine-tuning. Once the MVP is ready for experimentation, it should be deployed, implemented, and shared with test users for piloting. Collecting feedback from pilot users is an essential part of this phase.

The final phase of this process is the evaluation of proof-of-concept (POC) experiences. Since the ultimate goal of an application is to address business needs and fulfill use case requirements, the results of the POC project should be evaluated against the requirements defined during the concept planning phase. In addition to user and business needs, the experiences and outcomes of the POC should also be assessed against the technological requirements established during the AI system design phase.

AI design and development are particularly demanding with regard to performance and trustworthiness, as outcomes and outputs are often more difficult to evaluate than in traditional software testing of deterministic software products. Therefore, significant effort should be devoted to the evaluation of POC experiences, as assessing performance and trustworthiness requires domain experts or experienced end users to evaluate the AI application in real-world use cases. This provides more accurate feedback on whether the AI application is capable of solving similar problems or performing tasks comparable to those carried out by human experts. Such evaluation is especially important in POC projects involving agentic AI, where AI agents are expected to assist with or replace tasks that have traditionally been performed by humans.

Deterministic and AI-based software systems

Traditionally, software products in areas such as finance, human resources, enterprise resource planning, and other business processes have been based on deterministic logic. This means that the software operates according to predefined and programmed rules, workflows, and sequences. It does not generate its own interpretations or choose alternative ways of processing data; instead, it processes information systematically according to its predefined logic and workflows.

Unlike traditional software products, large language models (LLMs) are not built on deterministic business logic. They do not necessarily produce exactly the same output when given the same input. This is a significant difference compared to traditional software products, even though LLMs are also classified as software products.

While both traditional applications and LLMs can be considered software products, their approaches to data processing differ. Traditional software relies on predefined rules and deterministic execution, whereas LLMs generate outputs based on probabilistic models.

Differences in data processing and management

Traditional software systems, which still account for the majority of software applications, typically require structured data to function reliably. As a result, they are usually supported by a database with a predefined data model, and their user interfaces contain fields, forms, and drop-down menus that users employ when entering, processing, and managing data. Examples of structured data include names, addresses, status information, specifications, classifications, and customer records.

Multimodal AI models, on the other hand, are capable of processing unstructured data, such as long-form articles, social media posts, speeches, images, videos, and software code. They convert data into numerical representations and vectors with different values and weights and use probabilistic calculations and complex transformer architectures to predict e.g. the next element in a sequence based on the patterns of billions of numbers learned from vast amounts of training data. LLMs perform remarkably well in this task, and their output often appears as though it has been created by a human expert.

Conclusions

As demonstrated in this article, advances in AI models and their integration into digital services have created new competency requirements for software designers, developers, IT consultants, and managers. Although most users still interact with LLMs as conversational chatbots to support various everyday information needs and tasks, digital and AI-driven transformation is also reshaping software development processes.

Companies are purchasing and developing agentic AI solutions for a variety of business processes and operational IT systems. As companies, organizations, and societies become increasingly digital across multiple sectors and domains, IT professionals continuously design, develop, and implement new AI-based digital services, agentic AI applications, and large-scale IT systems.

The world of AI extends far beyond the use of LLMs as chatbots or handy assistants. Multimodal AI models and agentic AI solutions are increasingly being integrated into large IT systems and digital services, while the number of new AI-native applications continues to grow rapidly. Thus, managing the development of digital services and software products is becoming increasingly complex, and it requires a multidisciplinary approach where diverse competencies and areas of expertise are needed.

Picture: Shutterstock