AI for Software Architecture: Why 2026- The architect's copilot still doesn't work. (Category Architecture)
A systematic review of the literatureArtificial Intelligence Support for Software Architecture PracticeIt is a practical question - what architectural tasks AI can already support and why individual successes do not yet add up to a holistic practice. The first version appeared in 2025 year, but in 2026 It was updated in the year and 2 July 2026 year the article came out ACM Transactions on Software Engineering and Methodology. Updated literature review brought to August 2025-Go, added a slice of mass instruments for March 2026-go.
The authors began their analysis with 874 publications from four scientific bases, separately checked by ICSA and ECSA and left 51 peer-reviewed research. Then compare the results with the problems from previous interviews with 32 practices. And we found out the following.
1In limited tasks, AI is already useful The model for choosing one of the three patterns according to the requirements showed accuracy 70%; extraction of architectural areas of responsibility from the text - about 75fullness (recall) on four projects. On the hull 95 ADR GPT-4 He gave coherent decisions, but inferior to man in fullness. This is not a ready-made “architect”, but a first pass accelerator with mandatory verification. 2The most convincing are the tasks with a measurable contour On a small Kubernetes booth, a comprehensive load forecasting framework reduced container deployment time by up to 34RL agents better than random chaos monkey found critical failures, however, so far in the simulation. Industrial confirmation is available in automotive, aerospace and cyber-physical systems where the task is narrow and quality attributes can be calculated. 3Early design and strategic decisions remain largely academic prototypes Models are able to suggest boundaries, patterns, or ADRs in the current context, but are poor at holding organizational constraints, regulations, trade-off history, and the consequences of change. 4е In a separate section 21 The authors saw the same asymmetry. Most AI functions are already in observability, governance, conformance and local generation. The weakest are system-wide reasoning, two-way intent/ADR/code/runtime communication, and the accumulation of architectural erosion signals over time.
In general, at the level of the current AI slice, tools are already helping with architectural solutions, but at the level of helicopter view and the evolution of architecture over time, they do not drag on.
If you look at the authors’ approach to the analysis, they (attached pictures)
- Software Architecture challenges
- AI-specific challenges And a list of topics that have AI tools And they did a mapping between them and then they got insights higher.
If you analyze the relevance of the work in 2026 During the year, you can see that with 2025 year ArchBench, R2ABench, CAKE and SAKE - that is, one of the points of the roadmap of the authors of this review (architecture)They're already closing. The results are more likely to confirm the authors’ diagnosis. In R2ABench, models are good at building syntactically correct diagrams and extracting entities, but weakly linking them with relationships, which results in fragmented architecture; agent processes (agentic workflows) added instability, not sustainable gains SAKE warns that knowledge of architectural terms is a necessary filter, but not a proof of the ability to design a particular system with its trade-offs.
Roadmap of AI development in architecture from the authors of the article generally looks like this Living architectural knowledge base: requirements, solutions, code, telemetry with end-to-end traceability Living Arch Metrics and Testable Benches On top of this, it is already possible to build architectural intelligence that develops with the system: AI plays the role of a strong analyst, and the person remains a strategist and is responsible for context, priorities and compromises.
You can actually use these. 2 This is like a litmus test: If the requirement changes, can we understand the affected ADRs, components, code, quality attributes and timing signals? If there was an incident, can we trace back to the architectural odor, and then to the solution that gave rise to it? If not, the LLM will only make the next static shot more beautiful – useful architectural AI with the coherence of engineering data and the ability to check the recommendation on the history of the system.
In general, AI raises the cost of architectural discipline – the cheaper it is to generate local solutions, the more important it is to retain intent, boundaries, tradeoffs, and consequences over time.
#Architecture #AI #AI4SDLC #Engineering #Research #Software #SystemDesign