Architects should put AI to work on the submission set

25 September 2026

By Al Vigier

An architectural submission can run to hundreds of pages and still leave a reviewer unsure which version of the building they are being asked to approve. A revised drawing may disagree with a specification, or the fire strategy may refer to an old layout.

This is where I think AI could do some of its most useful work for architects. Practices have plenty of reasons to experiment with visualisation. They also have a less photogenic opportunity: finding mistakes in the submission set before those mistakes cost another round of review.

The RIBA’s 2026 survey found that 74% of respondents said their practice used AI on at least some projects. Its research identifies regulatory compliance checking among the areas where AI is being used. Those findings show that practices are willing to try it. They do not tell us how much time a practice will save on a regulatory submission.

England’s Gateway 2 process shows why that question matters. In the 12 weeks to 31 August 2026, the Building Safety Regulator reported a median approval time of 22 weeks for new higher-risk buildings and conversions, with complex cases reported separately at 33 weeks. The standard assessment period for new higher-risk buildings is 12 weeks, although extensions can be agreed.

The regulator has improved considerably since 2025. Its September update points to specialist staffing and its Innovation Unit as reasons for faster approvals. Practices cannot solve a shortage of reviewers by buying software, and the figures do not establish that AI caused the improvement.

But the quality of submissions is a problem too. In its analysis of a sample of rejected applications, published alongside data to March 2025, the BSR found insufficient detail in 73% of the rejected new-build applications it examined. This was more than a matter of presentation. The reviewer needed evidence that the proposed work would comply.

That distinction matters when choosing software. Checking that a fire strategy has been uploaded is a useful administrative task. Checking whether it addresses the project’s actual design requires a much more demanding review.

Honolulu offers an early test of what automated checking can achieve. Since 1 September 2026, its Department of Planning and Permitting has required CivCheck, an AI-assisted pre-check, for eligible residential projects. Applicants must complete the process before submitting. The department is explicit that this does not replace its official review or guarantee approval.

The department cites a comparison of 19 residential permits processed with CivCheck and 17 without it. Average time to a permit decision was 32.5 days for the CivCheck group, against 73 days for the other group. Average review cycles were 1.4 and 3.4 respectively. That is encouraging, but it is a small comparison, not proof that AI alone caused the difference or that another authority would achieve the same result.

Singapore is addressing coordination through CORENET X. Its default process brings more than 20 approval touchpoints into three main gateways, with coordinated submissions reviewed collectively by the agencies. A simpler direct submission route applies to some projects. The relevance for practices is the demand for consistent information across disciplines. A digital submission system should not, by itself, be treated as evidence of AI’s effectiveness.

The timetable has also changed. Under the July 2026 update, the mandate expands on 1 October 2026 to new projects meeting the 5,000 m² gross floor area threshold. Projects below that threshold can continue using CORENET 2.0 for now. Practices need to check the current requirements before building them into a workflow.

For a practice deciding where to begin, I would choose one submission and give the software a narrow job. Build a document register from the files actually being issued. Check it against the authority’s requirements, and link each requirement to the document and page that addresses it. Where evidence is missing or unclear, the system should say so.

Then test for contradictions across the set. Do the door ratings agree? Does the specification describe the same wall build-up as the drawings? A useful tool should take the reviewer straight to both sides of a suspected mismatch. The project team can then decide whether there is a real problem and how to resolve it.

Some of this is ordinary document management. File naming and revision control may be better handled by existing systems or simple rules. AI is worth testing on work those tools struggle with, such as comparing differently worded descriptions across documents. It should earn its place by reducing the total review effort.

Approval does not end the job. In England, controlled changes must be assessed and recorded, and major changes need BSR approval before the related work proceeds. Comparing revised documents against the approved set could help a team identify what needs attention. Deciding how a change affects compliance remains a professional responsibility.

I would test any proposed tool on a completed submission with known errors before relying on it. Count the problems it catches and the ones it misses. Record how long staff spend checking its findings and clearing false alarms. A system that produces a long list of dubious warnings may create more work than it saves.

Client information needs the same care. Before uploading drawings or specifications, establish where the files will be stored, who can access them and whether they can be used to train a model. The people responsible for the project still need to review the work and exercise their own judgement.

The return should show up in fewer avoidable corrections and less time spent chasing inconsistencies. A complete submission cannot guarantee approval, but a practice should not have to wait for the regulator to point out that two of its own documents disagree.

Al Vigier is CEO of Caseway, an AI company that develops products with enterprise partners and shares in the revenue.

Comments on this guide to Architects should put AI to work on the submission set article are welcome.