A public research programme on where AI is genuinely useful inside architecture practice. Each report should be source-backed, written for business owners and project leaders, and paired with a small proof of concept when a prototype can test the idea better than an essay.
The reports should follow real architecture work: winning projects, briefing, feasibility, design, documentation, coordination, consent, delivery, and learning from finished buildings.
Every useful claim should point back to sources, interviews, product tests, public data, or a working prototype. If the evidence is weak, say so.
Some topics need a small app, checker, dashboard, or retrieval tool. Some need interviews and synthesis. The POC should answer the research question, not become the product too early.
Start with the six areas that can become practical proof fastest: site feasibility, AI governance, documentation QA, specification intelligence, practice productivity, and market pipeline intelligence. They are close enough to real pain that architecture firms can react to them, and they give me small things to build and share as I learn.
Open the prototype lab for the concept demos built from these areas.
Which demand signals should a firm watch before hiring, pitching, or choosing a sector focus?
Building consent trends, construction pipeline data, tender notices, developer activity, council plan changes, and local market signals.
Firms need early warning on where work is moving, not a generic economic commentary after the fact.
Pipeline watch dashboard that turns public data into region, typology, and client-type signals.
How do clients choose architecture firms, and what changes when capital, consent, or speed becomes the constraint?
Client interviews, RFP language, developer websites, listings, project announcements, and post-win/lost-job reviews.
The most useful AI may be in reading intent, risk, and fit before a practice spends senior time chasing the wrong job.
Client-brief classifier that turns a listing, email, or RFP into fit, risk, missing-info, and next-question notes.
Can a site be turned into a reliable evidence pack before concept design starts?
Parcel data, LiDAR, planning rules, hazards, services, listings, survey inputs, and known missing-data states.
Early design improves when the site talks back with constraints, uncertainty, and source links.
Site-pack app: address to terrain, planning, constraints, evidence, and concept-readiness notes.
Where is AI actually being used inside architecture practices, and where is it still theatre?
Practice interviews, task diaries, tool testing, staff surveys, policy documents, and before/after workflow examples.
The useful distinction is not "uses AI" versus "does not use AI"; it is where AI changes a task enough to matter.
Workflow audit template that maps tasks by risk, repetition, review need, and automation potential.
What should an architecture firm allow AI to do, suggest, check, or never touch without review?
Professional conduct duties, insurer positions, firm policies, model-risk examples, and review-gate case studies.
Governance has to be practical enough for daily work, not a policy PDF nobody opens.
AI use register and review-gate checklist for project teams.
What interface lets a designer shape with AI instead of just prompting for outputs?
HCI research, mixed-initiative design tools, studio testing, sketch workflows, and observed friction in prompt-only tools.
The interface matters as much as the model. Spatial work needs a loop of shape, feedback, and revision.
Sketch-house and house-to-3D experiments: draw, sculpt, generate, inspect, and revise.
Can AI help with construction details without pretending to sign off engineering design?
Detail libraries, office standards, connection examples, reviewer interviews, common RFIs, and failed-check examples.
The likely value is retrieval, comparison, packaging, and review support, not autonomous detail invention.
Detail precedent finder that returns similar past details with source, assumptions, and reviewer notes.
What can AI read from a model, and what still belongs in Revit, IFC, schedules, or a coordinator's judgement?
Revit/IFC exports, model element data, BIM manager interviews, clash workflows, schedules, and Autodesk handoff paths.
AI is most useful when it interrogates and packages model information rather than trying to replace the model environment.
Model question layer that answers bounded questions from IFC/Revit exports with element references.
Which drawing errors can a lightweight AI checker catch before issue?
Issue sheets, revision history, common coordination misses, QA checklists, and redline examples from real projects.
The first win is not perfect automated checking; it is a second-pass reviewer that finds omissions and inconsistencies.
PDF drawing QA pass that checks titles, sheet references, detail callouts, revision notes, and missing legends.
Can AI help choose and document products while keeping manufacturer data and compliance evidence visible?
Product technical literature, EBOSS-style listings, manufacturer manuals, office spec masters, substitutions, and warranty documents.
Specification work fails when claims lose their source. Product intelligence has to preserve evidence and boundaries.
Product evidence card: source, use-case fit, constraints, warranty notes, substitutions, and unanswered questions.
Where can AI reduce consent friction without giving false legal or code certainty?
Council checklists, RFIs, Building Code clauses, planning overlays, past consent packs, and consultant review notes.
The best target is evidence completeness and RFI avoidance, not automated compliance sign-off.
Consent evidence matrix that maps project documents to checklist requirements and missing evidence.
How can firms make climate and carbon decisions earlier, with assumptions visible?
Embodied-carbon data, EPDs, material schedules, climate hazards, passive design assumptions, and rating-tool requirements.
AI can help surface trade-offs early, but only if the assumptions and data quality are shown.
Carbon and climate assumption card attached to early material or site decisions.
Can AI flag design decisions that create avoidable procurement or buildability risk?
Quantity assumptions, supplier lead times, contractor feedback, cost plans, RFIs, substitution history, and build sequence notes.
Good AI support may look like early risk notes, not a fake precise cost estimate.
Constructability memo generator for a concept option: unknowns, long-lead items, risk flags, and questions for QS/contractor.
Which tasks actually consume margin, and where can AI reduce waste without lowering quality?
Timesheets, project stage plans, fee proposals, rework causes, issue logs, and staff interviews.
AI should be judged by whether it improves project economics and quality, not whether it looks impressive in isolation.
Project-margin diagnostic that maps repeated time drains to possible automation, templates, or review changes.
How should a firm explain its value as AI makes generic output cheaper?
Firm websites, project pages, pitch material, client interviews, competitor scans, and win/loss notes.
As generic production gets cheaper, trusted judgement, sector knowledge, and delivery credibility become more important.
Firm positioning review that compares public proof, project evidence, sector language, and client fit.
What should different roles in a practice actually learn about AI?
Role interviews, software usage, common failure cases, policy gaps, training material, and project-stage responsibilities.
Training should be role-specific: directors, designers, graduates, BIM managers, and project architects need different skills.
Capability matrix that maps role, task, risk level, tool pattern, and required review behaviour.
Can a firm turn its archive into a source-backed precedent system?
Past project files, specs, drawings, meeting minutes, lessons learned, correspondence, and office standards.
The boring archive may be one of the strongest AI assets a firm has, if retrieval stays cited and permissioned.
Internal precedent search that answers from a small project corpus with citations and "not enough evidence" states.
How can practices learn from occupied buildings and feed that knowledge back into design?
Occupant feedback, defects, maintenance data, energy use, comfort surveys, photos, and project review notes.
AI can help structure feedback, but the firm needs a repeatable loop from built outcome back to design decisions.
Post-occupancy insight pack that clusters feedback into comfort, maintenance, energy, and design-decision lessons.
Which AI tools matter to architecture firms, and which are demos without practice fit?
Vendor demos, pricing, product docs, user reports, workflow fit tests, integration paths, and export quality.
The key question is not model quality alone. It is whether the tool fits review, source control, liability, and handoff.
Tool scorecard comparing source traceability, review gates, export, data control, and actual practice use-case fit.
What could architecture practice look like if AI becomes normal infrastructure rather than a novelty?
Signals from current tooling, professional regulation, client expectations, delivery models, and adjacent industries.
The strongest future scenarios should be practical enough to guide decisions now, not abstract predictions.
Scenario workshop board for firm leaders: likely shifts, risks, bets, and capabilities to build this year.
Each report can follow the same structure: thesis, why it matters to an architecture firm, evidence base, workflow map, risk boundary, small POC, what the POC showed, and what still needs testing. That makes the series feel like ongoing research rather than marketing copy.