Panel

Merlin Fulcher (chair), competitions editor, Architects’ Journal
Phil Sanders, senior consultant, CMap
Martha Tsigkari, senior partner and head of Applied R&D, Foster + Partners
Rada Daleva, architectural designer and founder, Daleva Design
Kira Ariskina, architect, founder and chief executive, ViableSite

Artificial intelligence is already being applied across architectural practice, from the retrieval of office knowledge and the comparison of building regulations to parametric design, project management, feasibility studies and the operation of completed buildings.

The AJ webinar, chaired by competitions editor Merlin Fulcher and supported by CMap, considered what these developments mean in practical terms. Rather than treating AI as a single technology or a replacement for professional expertise, the panellists described a range of approaches shaped by the needs, scale and culture of individual practices.

The discussion brought together Martha Tsigkari, senior partner and head of the Applied R&D group at Foster + Partners; Rada Daleva, architectural designer and founder of Daleva Design; Kira Ariskina, architect, founder and chief executive of ViableSite; and Phil Sanders, senior consultant at CMap.

From tools to practice

Martha Tsigkari opened with an overview of the work undertaken by Foster + Partners, where AI research has been part of the practice’s wider interest in technology, performance and innovation for several years. ‘We have been doing AI at Foster + Partners for many, many years. We started in 2018,’ she said.

For Tsigkari, the important point is that this work extends well beyond generative imagery. She said Foster + Partners has been exploring ways of making the practice’s accumulated knowledge more accessible, including applications that allow staff to search documentation, images, materials and two-dimensional drawings.

‘We started by sharing our 60 years’ worth of knowledge through applications like Ask Foster + Partners,’ she explained. The aim is not only to retrieve information but to connect different datasets and make them available through interfaces that allow architects to ask questions in a more direct way.

‘Just retrieving data is half of the problem,’ Tsigkari said. ‘We are building applications where our different data sets are connected across each other, and they’re facilitated through AI agents, where you can ask questions and retrieve interesting information.’

This approach is being applied to a wide range of activities. Foster + Partners has developed AI-assisted applications for comparing building codes and regulations, translating material, creating documents, searching files and writing reports. Other applications support project management by collecting comments and questions from across a project, then directing them to the appropriate people for response.

The practice is also applying AI to operational data. Digital twins and other systems can continue to generate information after a building has opened, providing a basis for suggestions and analysis. Tsigkari described the use of predictive models for resourcing, cost profiles and profitability, alongside virtual experiences and tools that support performance-driven design.

The underlying models are not necessarily built from scratch. ‘A lot of these are applications that were built on top of models that already exist, but are fine-tuned with our data,’ she said. The value lies in the way the practice structures its own information and integrates it into workflows.

Supporting creative judgement

Tsigkari also described experiments in AI-generated layouts, parametric modelling and design optioneering. Designers can bring sketches, images and three-dimensional models into the practice’s AI portal and use them to explore alternatives across different formats.

The objective, however, is not to hand over design authorship. When Fulcher asked whether the technology risked making Foster + Partners’ service offer less unique, Tsigkari emphasised the importance of culture and intent. ‘The identity of Foster Partners is not driven by any single tool, it’s driven by the culture at Foster Partners and the design excellence that we are working towards and have always been working towards,’ she said.

AI, machine learning and performance-driven design are tools that support that culture, rather than define it. ‘None of the AI tools that I showed are meant to replace what our designers are doing, quite the opposite,’ Tsigkari said. ‘They’re meant to help them be more productive, boost and augment their creative juices, and allow them to do more with less.’

That distinction also shaped her response to questions about governance and professional responsibility. AI-generated outputs are not delivered without review. ‘None of the things that we do, or things that we deliver as is,’ she said. The practice continues to apply the same checks and balances that it would use for a model created in Revit or another BIM platform.

‘In the same way that you would create a model in Revit, or in a BIM software, and somebody would have to check everything from the outputs, the drawings, the 3D model, everything before it’s delivered, we’re doing exactly the same thing now,’ Tsigkari said.

For her, AI changes the speed of certain processes but not the responsibility attached to professional decisions. ‘Technology is just a facilitator, it’s not a… it’s a means to an end, it’s not the end itself,’ she said.

Designing with intention

Rada Daleva offered a perspective from a smaller, design-led studio. Daleva, founder of Daleva Design and formerly project lead at Studio Tim Fu, described how AI can help communicate design ideas quickly, particularly when clients are working to tight programmes or need to understand a proposal through images and iterations.

She began with a request from a recent client: ‘Fix this with AI, please.’ Daleva said that this captures a growing misconception that AI functions as a magical solution capable of resolving design problems in minutes. In practice, the design decisions still belong to the architect.

‘I have fixed it, but through design,’ she said, describing decisions about materials, depth, furniture and the relationships between elements. AI helped her communicate those decisions ‘faster and better to the clients’.

For Daleva, the technology has altered the economics of exploring options, but not the responsibility involved in selecting one. ‘AI removed the cost of making a lot of options, but not the cost of choosing the right options,’ she said. The architect’s value lies in deciding which proposal should be developed, built and ultimately presented as the answer to a brief.

She illustrated this through an ongoing project in which she was asked to improve parts of a design while construction was already under way. The challenge was to bring together elements that were close to one another but lacked a coherent relationship. Daleva developed a common language across the interior, drawing on the clients’ interest in the peacock and the peacock feather and interpreting that idea through different materials, forms, doors and pieces of furniture.

The work was produced using a mixture of AI-assisted exploration and conventional modelling. Daleva said that the design elements were modelled in Rhino for production, as current image-to-three-dimensional tools did not yet provide the level of detail required for manufacture.

Her broader conclusion was that AI makes it possible for smaller teams to undertake more ambitious work. ‘Architecture is being decentralised,’ she said, arguing that one architect with the right tools can produce considerably more than was previously possible.

That does not remove the need for judgement. Instead, it makes the question of judgement more important. ‘The question is not, is it faster, but is it better?’ Daleva said.

A new role for expertise

Kira Ariskina brought a data and planning perspective to the discussion. As founder and chief executive of ViableSite, she is developing an AI-enabled platform intended to speed up and automate early-stage feasibility work on small urban sites.

Ariskina’s central argument was that generic AI systems lack the detailed knowledge required to understand how architectural schemes work in the real world. As models improve, routine tasks such as drafting clauses or summarising reports become more widely available. The scarce resource, she suggested, will be domain-specific knowledge and professional judgement.

‘The question kind of moves from, can I use AI, to what do I know that LLM doesn’t know?’ she said. That expertise is not only located in the heads of individual architects. It is also embedded in drawings, specifications, cost plans, planning outcomes and project records.

Ariskina challenged practices to consider whether they record the reasoning behind decisions, rather than merely retaining the final document. ‘You need to start thinking about building a system that not just records the documents or stores the documents, you need to think about your internal system, how that records your decision-making process,’ she said.

The reason is straightforward: the combination of historic project data and the reasoning behind it can become a powerful resource. Without that link, a system may know what a practice did but not why it did it.

Ariskina advised practices to begin with a goal rather than with the technology itself. That might be speeding up fee proposals, using previous projects to understand the likelihood of planning approval or examining how a particular borough has responded to comparable schemes. Once the goal has been defined, the data can be structured accordingly.

That work is not glamorous. It may involve cleaning files, removing duplicates, converting long PDFs into usable text, dividing information into manageable sections and adding metadata to images. Ariskina compared the process to organising information for a new member of staff. ‘Think about AI pretty much like about your new hire,’ she said. ‘If part one or part two will find it, or newly hired architects will find it, AI probably will be able to make sense out of it.’

Rubbish in, rubbish out

Ariskina was also clear about the risks. ‘Please remember, rubbish in, rubbish out,’ she said. Poor-quality or inconsistent data does not simply disappear when processed by an AI system. It can be turned into an answer that appears authoritative while being fundamentally wrong.

‘Be absolutely ruthless about what you put in, and please always keep a qualified human who will verify what comes out,’ she continued. ‘AI can make mistakes… but at the end of the day, you will be liable for those mistakes.’

Confidentiality and intellectual property must also be considered. Information used internally may still be shared with a third party, depending on the terms of the relevant large language model. ‘Always check the terms and conditions of this particular LLM, who you’re sharing with,’ Ariskina advised.

Where AI-generated material is used externally – in documents, submissions, applications or published work – the level of responsibility increases. Ariskina argued that architects should consider declaring when AI has been used to draft or substantially rewrite text, create summaries or alter images, alongside explaining what checks have been carried out.

Her advice was to approach AI with the same professional caution that architects apply to other parts of their work. ‘We know the duty of care, liability, and record keeping,’ she said. ‘So don’t be scared of it, just approach the AI the same way you approach any architectural project.’

Making information useful

Phil Sanders, senior consultant at CMap, connected these questions of data, practice management and operational efficiency. With a background spanning design, visualisation and practice management, Sanders described the problem facing many firms: they hold large amounts of information but do not always have a practical way to interrogate it.

‘We have all this data, but we’re not using it properly,’ he said, describing the question being asked by practices considering AI. Architecture and construction generate substantial quantities of data, but most architects are not trained as data analysts. The opportunity, therefore, is to make the information already held by a practice easier to use.

CMap is developing AI agents that can surface previous project performance when preparing fee proposals, identify client payment behaviour at the point of invoicing and update invoice schedules, resourcing and project information when a project stalls. Sanders also outlined CMap Chat, a conversational interface that allows users to ask questions of their practice data in plain English.

The wider ambition is to connect information across systems. Sanders explained the idea of a Model Context Protocol, or MCP, as a more flexible form of integration that could allow an AI system to interpret and coordinate information from different sources. A practice might hold brand guidelines in one system, terms and conditions in another and project information in a third. In future, these could potentially be brought together to create a structured proposal or explore business scenarios.

Sanders stressed that AI is not intended to replace expertise. ‘AI, for us, isn’t about replacing expertise, it’s about making it easier to use what we already know,’ he said. The value is in surfacing existing information, joining it together and presenting it when it is needed.

Value, not speed

The final discussion returned to the effect of AI on the profession’s economic model. Fulcher asked how practices should respond if work that once took several weeks can be completed in a matter of hours, and whether clients will expect the same output for less money.

Tsigkari argued that this should prompt a broader discussion about the value of architecture rather than a narrow focus on hours. ‘People have associated the value of what we’re doing based on how many people you need and how many hours it will take,’ she said.

The profession, she argued, needs to reconsider how it defines and communicates that value. ‘That value has to do with the final asset that you’re providing, the quality of the building that you’re creating on the built space, the quality of the drawings,’ she said.

Daleva agreed, arguing that the profession has been weakened by competing primarily on lower fees. ‘If we ourselves can’t define our value as architects, then of course the clients are not gonna trust us,’ she said. She suggested that the emergence of AI makes hourly charging look increasingly vulnerable and should encourage practices to consider different ways of structuring and communicating their services.

Sanders also supported a shift away from a purely time-based model, although he acknowledged that changing a system so deeply embedded in the industry would be difficult. Tsigkari, meanwhile, cautioned against describing the future simply as commoditisation, arguing that the consequences of such a shift need to be understood carefully.

Asked whether AI would reduce construction costs, Ariskina was more circumspect. ‘Building costs are still happening in real life,’ she said, adding that AI could speed up communication within the design team but would not, by itself, alter the cost of building.

The webinar’s overall message was that AI is most useful when it is applied deliberately: to a defined problem, with reliable information, clear governance and a qualified professional responsible for the result. The technology may change how quickly practices search, model, communicate and explore, but the panellists argued that the core of architecture remains the ability to make informed judgements and take responsibility for the spaces that result.