In the first part of this series, Vladislav Ivanov explored how new AI technologies – especially language models – are changing architecture practice. In this second part, I look at how language-model technology can be used in architecture work today: what works, and what does not, at least not yet. Have AI agents already replaced enterprise architects? Has the hype been completely unfounded? Or is the truth somewhere in between – in small, practical improvements to everyday work?
Language models have now been tested for a range of purposes for several years. People have been waiting for tangible results, and some have begun to emerge. It is tempting to think that architecture work could simply be outsourced to AI: if language models can produce text, why could they not produce architecture descriptions as well?
And they can – that much has been proven. Carefully instructed, the latest cloud-based LLMs – by which I mean generative services based on language models, such as Anthropic’s Claude and OpenAI’s ChatGPT – can produce fairly extensive descriptions of the whole formed by an organisation’s operations, data and information systems. The description can also be downloaded directly in a machine-readable format for use and further development in a modelling tool. This works particularly well for organisations with a large amount of information available on the open internet.
Generated architecture still requires validation
Architects cannot be dismissed just yet. Creating descriptions is only one part of an architect’s role. Generating them is useful, but they still need to serve a purpose – as current-state or target-state material in development work. That means confirming that they reflect reality. Let us assume a language model produces a description stating that:
- we have 189 processes;
- we have 79 information systems;
- there are 395 integrations between our information systems; and
- our operations are governed by 63 national or EU-level regulations.
How would we verify that the description is true – that each of these things actually exists? We could turn to the intranet, only to discover that it offers no final certainty either. A page created by Pertti in 2019 appears to confirm some of the details, but other information is clearly outdated. Unfortunately, Pertti retired a couple of years later. And wait – where are the technologies? Did the AI find them at all? A PowerPoint deck created by Liisa in 2023 mentions a cloud transformation, but was the target state she described ever implemented as planned? Should we call a meeting to clarify the situation and start compiling yet another PowerPoint?
Human architects excel at collaboration and sense-making
You can ask the language model for more, but it soon starts going in circles. You can urge it to “try harder”, yet the underlying problem is often a lack of sufficiently current, comprehensive and high-quality material.
People know things, but they have not written that knowledge down anywhere. Existing documentation can often be summarised into a surprisingly neat whole, but the available ingredients are soon exhausted. Similar bursts of progress are harder when people need to agree on development goals across organisational boundaries or when language barriers are involved.
A patient and empathetic human architect can achieve things that remain difficult for a language model. As before, it is still valuable to talk with people, collect what they know and structure it into coherent wholes.
Data security shapes AI adoption
This quickly raises another question: do we trust a cloud-based LLM provider enough to give it everything relevant to the organisation’s architecture – from executive team meeting minutes to the contract register and the contents of the CMDB?
Contracts can help manage the risks, but organisations should be prepared for data-flow governance to become highly complex.
Local language models offer greater control
An experimental architect can also install a local language model on their workstation – or in the organisation’s data centre or a cloud service deemed sufficiently secure – and vibe-code an agent alongside it with access to the required data. This provides more control over where organisational data moves and how it is used.
The challenge is that local language models are not yet on the same level as cloud-based LLMs.
The biggest benefits come from everyday productivity gains
Whether the organisation uses a cloud-based LLM service or a locally installed language model, information processing remains a challenge. Language models can produce architecture descriptions, but without guidance they struggle to work with descriptions of organisational structures.
Cloud-based LLM services can do more than generate diagrams. They can also create scripts that extend the capabilities of modelling tools. In other words, an architect can vibe-code new features into their tool.
Architecture scripts are typically compact enough for this approach to work well. I have personally gained a great deal from scripts that arrange boxes in diagrams. It is not yet a disruptive game changer, but it makes a meaningful difference in everyday work.
When does AI deliver the most value?
Modelling-tool vendors are understandably moving as well. At the time of writing, there have been no major breakthroughs, but sooner or later one tool will probably gain features that significantly support architecture work – and competing tools will likely copy them quickly. Because architecture work has long time horizons, the latest features are rarely essential. I would therefore not rush to buy an expensive AI architecture tool, however eagerly a vendor might offer one.
In short, the situation is interesting. AI already affects architecture work, but with caveats. Language-model-based services are not yet tools in the conventional sense: they cannot simply be used in a standardised way and expected to produce consistent results. At their best, they are like sourdough starter – in the right conditions, they can produce something genuinely good.
What are the right conditions? Perhaps something like this:
- A large amount of high-quality, unambiguous, current and consistent information about the organisation, especially its structures. Without it, a language model blends present and outdated views, facts and assumptions.
- A clear “production line” that defines the technologies used and their constraints, together with the information used and the limits governing its processing. Without this, we may become overly cautious – or share information about our organisation that we never intended to disclose.
The architect’s role is changing, not disappearing
New technologies are often expected to be miracle cures, and language models are no exception. Few technologies have fulfilled every promise made about them, and forecasting the future is now harder than usual.
Even so, organisations that have done their homework and retained skilled architects will find new opportunities in these technologies. Those still at the beginning of their architecture journey may gain valuable momentum from them.
Gofore’s architecture experts help organisations build sustainable architecture capabilities, leverage AI responsibly and create a strong foundation for long-term transformation.
