THESIS
The model is not the engineering. The chain of artefacts around it is.
Software engineering was built on the assumption that a human writes the code. That assumption is gone, and the industry response has been to insert a generator into the old lifecycle and measure how much it produces. That answers the wrong question. Generation is already cheap and improving without our help. What is neither cheap nor solved is everything around it: capturing intent in a form that survives regeneration, deciding mechanically whether what came back is acceptable, and preserving the trace from a sentence someone said in a meeting to the line of code that now runs in production.
ANSE treats requirements, architectural decisions, acceptance criteria, test plans and generated code as artefacts in a single governed chain, where the output of each stage becomes the controlled input to the next and a human approves at the points where approval changes the outcome. We are not building a better coding assistant. We are building the discipline that makes a coding assistant accountable.
A NÄYTÖNPAIKKA PROGRAMME
ANSE is funded under Näytönpaikka (Rise to Challenge), Business Finland's instrument for radical research openings and its largest single commitment to research organisations. The call is open in theme and narrow in ambition. It funds a small number of programmes selected from an open field, and it asks each of them for something specific: build capability that multiple industries will need, combine disciplines without regard for their conventional borders, assemble the strongest people working in the field in Finland, and take on a mission whose commercial consequences are real but not yet precisely estimable.
The instrument also imposes a boundary that defines what ANSE is. A Näytönpaikka project must address a theme that does not sit on an existing industrial roadmap and that could not be built as a company collaboration project instead. Work that industry can already specify and fund does not qualify. ANSE was admitted on the basis that its question is not yet in that state.
Funding is staged and conditional. The first phase runs on measured objectives agreed with the funder, and continuation into the second phase is decided on whether those objectives were met, not on a new application. The measures cover the growth of the research team, progress along the published roadmap, and evidence that the people who will use the results are committed to the next phase. A programme of this type is expected to show, at a fixed date, that it did what it said it would do.
THE PROBLEM
The industry response to generative AI has been additive: keep the process, insert the tool, measure acceptance rate. This produces genuine local speedups and no change in the discipline. It also produces a specific pathology. Fluency at the keyboard rises while the artefacts that carry meaning, the specification and the validation criteria, stay informal and undocumented. The work gets faster and less legible at the same time.
The harder problem is that the AI-native question is not a tooling question. It is a question about where authority sits. If a human no longer writes the implementation, then the specification and the validation rules are the only places where human intent is recorded, and they were never engineered to carry that weight.
METHOD
The chain has four stages and one thing that holds them together.
For brownfield systems, existing code and documentation are read back into a reconstructed current architecture, and the requested feature is analysed against it. The proposed changes are presented as engineering decisions to be approved, rejected, or dropped, not as edits to be applied. This is where the human approval point sits, because this is the stage where a wrong decision is cheapest to catch and most expensive to miss.
WHERE ANSE SITS
ANSE is not the first Finnish programme to work on AI and software engineering, and it is deliberately not doing what the others do. MAISA established AI-assisted development at industrial scale. GENIUS, running under ITEA, works on extracting the best available support from generative AI across the conventional software engineering lifecycle. ELFMo addresses the integration of large language and foundation models into software-intensive systems, and Excellent-AI extends that to how organisations operationalise AI at the strategic and product-management level. Together they have produced validated tools, working practices and developer-level productivity evidence that ANSE does not need to recreate and builds on directly.
What they share is a premise. Each takes the shape of the software engineering lifecycle as given and asks how AI can be applied within it: to a task, to a phase, to a product, to a management layer. That premise is reasonable and it is where most of the value has been found so far. ANSE exists to test what happens when it is withdrawn. If generation rather than authorship is the default mode of production, the lifecycle is not a container for AI to be placed into; it is the thing under revision.
The boundary is not an editorial preference, it is a condition of the funding. Näytönpaikka finances only themes that sit outside existing industrial roadmaps and that could not be built as company collaboration projects instead. Every neighbouring programme named above is such a project. ANSE was admitted on the opposite basis: the question is not yet specified well enough for industry to fund it directly, and the work required to specify it is the research. Shared people and shared partners across these programmes are a feature of the Finnish ecosystem and are coordinated through the SW4E network. Shared questions would be a defect, and the instruments are designed to prevent them.
DEMONSTRATORS
In June 2026, ANSE and Harvia took a real case from a sauna control system and asked how much of the software engineering lifecycle could be carried by an AI-native approach. An existing web application monitored sauna temperature. The requested feature was a timer, letting a user pause, resume and reset while holding temperature. Ordinary, small, and complete, which is what makes it a usable demonstration: a reader can judge for themselves whether the feature that came out is the feature that was discussed.
What makes the case a genuine test is that the system is brownfield. There is existing code, architectural history, dependencies, and decisions nobody documented. Greenfield is where these tools look best and where almost every published demonstration takes place.
The run is recorded end to end: the conversation, the requirements it produced, the architectural decisions taken and rejected, the engineering contract handed to the coding agent, the failing tests, the resulting changeset, and the execution telemetry underneath all of it.
Harvia sauna timer (brownfield)
A timer feature — pause, resume, reset while holding temperature — added to an existing Harvia sauna temperature-monitoring web application, carried through the full ANSE chain from recorded conversation to reviewable changeset. June 2026.
A recorded working session between a Harvia product owner and a developer discussing a timer for the existing temperature-monitoring application: what pause should do to the heating element, whether reset returns to the configured target, and what the display shows while paused. The transcript is the primary artefact, not a summary of it.
RESEARCH OUTPUT
ANSE is a research programme, so the peer-reviewed record is the primary evidence, not a by-product. The list below is generated from public bibliographic data and updated automatically; preprints and work under review are marked as such.
A Context-Aware Multi-agent Approach to Enhancing User Story Management in Agile Software Development
A Two-Way Survey of Intrinsically Interpretable Deep Learning for Tabular Data: Bridging Methods and Systems
Agentic Frameworks for Reasoning Tasks: An Empirical Study
Anomaly detection in smart power grids with graph-regularized MS-SVDD: a multimodal subspace learning approach
Bridging Humans and LLMs: Investigating Human-AI Collaboration in Multi-agent Requirements Analysis for Organizational AI Adoption
CodePori: Large-Scale System for Autonomous Software Development Using Multi-Agent Technology
Editorial: Benefits and detriments with respect to the ethics and sustainability of gamification
ERC4AI: A BERT‐Based Model for Ethical Requirements Classification in AI Systems
From Business Meetings to Requirement Artifacts: An Agentic AI Approach with MARARE
From Meetings to Product Vision: An Agentic Approach with MARARE
LLM-Based Multi-Agent Systems for Code Generation: A Multi-Vocal Literature Review ⋆
Operationalizing AI Ethics in the Public Sector: A Cross-Context Replication in Brazil
PLACEHOLDER: A Validation Layer for Agent-Generated Software
PLACEHOLDER: Measuring Requirement Traceability in Agentic Pipelines
PLACEHOLDER: Specification-First Construction with Autonomous Coding Agents
Specifying Fairness and Transparency Requirements for Public Benefit Allocation
Towards AI Transformation in Software Engineering - Selecting Generative AI Use Cases
Vibe Coding in Software Development: A Multivocal Literature Review
Vibe Coding in Software Development: A Multivocal Literature Review
Vibe Coding in Software Development: A Multivocal Literature Review
Vibe Coding in Software Development: A Multivocal Literature Review
A Multi-agent LLM System for Automated Requirements Analysis: A Study on User Story Generation and Prioritization
AI and Agile Software Development: A Research Roadmap from the XP2025 Workshop
AI and Agile Software Development: From Frustration to Success -- XP2025 Workshop Summary
AI and Agile Software Development: From Frustration to Success XP2025 Workshop Summary
Assessing Small Language Models for Code Generation: An Empirical Study with Benchmarks
Assessing Small Language Models for Code Generation: An Empirical Study with Benchmarks
Autonomous Agents in Software Development: A Vision Paper
Autonomous Legacy Web Application Upgrades Using a Multi-Agent System
Autonomous Legacy Web Application Upgrades Using a Multi-Agent System
Can Large Language Models Serve as Data Analysts? A Multi-Agent Assisted Approach for Qualitative Data Analysis
Distributed Approach to Haskell Based Applications Refactoring with LLMs Based Multi-Agent Systems
Engineering RAG Systems for Real-World Applications: Design, Development, and Evaluation
Enhancing Regulation-Adherent Requirement Engineering with Contextual AI: An Industrial Study
ERC4AI: A BERT-Based Model for Ethical Requirements Classification in AI Systems
ERC4AI: A BERT-Based Model for Ethical Requirements Classification in AI Systems
ERC4AI: A BERT-Based Model for Ethical Requirements Classification in AI Systems
FaireduPlus: Enhancing Intersectional Fairness in Education-Focused Machine Learning Using Synthetic Data
From Specification to Service: Accelerating API-First Development Using Multi-Agent Systems
Generative Artificial Intelligence for Software Engineering—A Research Agenda
Large Language Models for Code Generation: The Practitioners Perspective
LLM-Based Multi-agent System for Intelligent Refactoring of Haskell Code
LLM-Generated Microservice Implementations from RESTful API Definitions
LLM-Generated Microservice Implementations from RESTful API Definitions
PLACEHOLDER: AI-Native Software Engineering: A Research Agenda
PLACEHOLDER: What Remains of Code Review When Humans Stop Writing Code
Still just personal assistants? – A multiple case study of generative AI adoption in software organizations
The EU AI Act is a Good Start But Falls Short
Toward AI Evaluation of Student Essays
VAPU: System for Autonomous Legacy Code Modernization
Vibe Coding in Practice: Flow, Technical Debt, and Guidelines for Sustainable Use
INDUSTRIAL STEERING GROUP
ANSE is a joint programme of the University of Jyväskylä and Tampere University, funded by Business Finland. Jyväskylä coordinates the programme through its Empirical Software Engineering Research group, whose research line is built on constructing technical artefacts and then measuring their effect on the development method rather than surveying opinions about tools.
Tampere University contributes through GPT-Lab, with work on the specification-driven control surface, requirements capture, orchestration, and developer experience. Both groups work with industry partners on their own production systems.
| COMPANY | FI | STEERING GROUP — Kalle Mäkelä | |
| COMPANY | FI | STEERING GROUP — Jari Partanen | |
| COMPANY | FI | STEERING GROUP — Tuuli Lindroos | |
| COMPANY | FI | STEERING GROUP — Timo Harvia | |
| COMPANY | FI | STEERING GROUP — Henri Wenlin | |
| COMPANY | FI | STEERING GROUP — Antero Taivalsaari | |
| COMPANY | FI | STEERING GROUP — Petri Liimatta | |
| COMPANY | FI | STEERING GROUP — Iiro Sundberg | |
| COMPANY | FI | STEERING GROUP — Manu Setälä | |
| NETWORK | FI | STEERING GROUP — Markku Oivo | |
| COMPANY | FI | STEERING GROUP — Mika Karaila |
PARTICIPATE
ANSE works with companies that are past the tool-adoption stage and have started to notice that their process no longer describes what their teams actually do. Participation means bringing a real system and a real constraint, not a pilot sandbox, and getting method work done against it alongside researchers who publish what they find. The programme's next phase is being scoped now.
CONTACT
Funded by Business Finland under the Näytönpaikka (Rise to Challenge) call for radical research ideas.
