Workflow coordination
Single AI tasks become coordinated work: context, model checks, options and review prepared in parallel, so the same team carries more projects and margin holds as fees tighten.
Research report
Architecture firms
Grounded in built prototypes
The state of AI in architecture practice.
This source-backed review examines where AI is changing architecture work, the controls firms are adding, and the workflows that can reduce labour while keeping judgement, approval and liability inside the practice.
AI creates practical value when it works inside a controlled project workflow.
The sections below examine market behaviour, useful applications, the project context AI requires, and the common control loop identified across three prototypes.
01MarketWhere the work is going
Architecture firms are adding controls around project AI use.
Autodesk's 2026 AI Pulse reports that 98% of design and make leaders already use at least one AI tool, and 59% are using or plan to use agentic AI within a year. Firms now need to define where AI can reduce labour and speed decisions, and where accountability remains with their people.
NoFromTo · where the market is going
01AI features
→Workflow products
The strongest products are built around whole work packages: early planning, site checks, fit-out, model data, drawing review and trade coordination, not isolated add-ons.
02One assistant
→Specialist product layers
Forma, TestFit and Ark focus on feasibility; Snaptrude on browser BIM; qbiq on planning packages; Speckle on model data; Augmenta on coordinated building systems.
03Design output
→Decision package
The output is no longer just a drawing, render or model. The useful package carries geometry, assumptions, source status, schedules, exports and risk signals.
04Hands-off automation
→Accelerating review loops
The strongest pattern is human-controlled automation: AI prepares, compares and generates options while professionals keep judgement, approvals and liability.
02BehaviourWhere adoption is heading
Staff behaviour is changing ahead of formal firm systems.
Specialist products continue to improve, but firms are already adapting their own workflows around gaps in those platforms. The resulting process, including context, handoffs and review, is becoming an important internal capability.
User uptake signal
Production AI is already normal next door, in software
Architects get the most from AI in early, exploratory design.
2026 State of AI in Architecture: of 1,200+ architects, 43% said conceptual and pre-design is where AI adds the most value, ahead of documentation and delivery.
~27%
of production code across 4.2 million measured developers is AI-generated.
DX · 2025>30%
of Google's new code is AI-written, up from 25% six months earlier.
Google · 202580%+
of code merged into Anthropic's production codebase is authored by Claude.
Anthropic · 2026The expertise that wins work is already specific: how a firm tests a site, prices a job, checks a model and runs a review. AI creates leverage when it works inside that context. The advantage comes from turning the business a firm already has into repeatable workflows.
03Business impactMargin · capacity · risk
AI changes practice economics through coordination, data and repeatable workflows.
The five shifts below affect margin, capacity, differentiation and risk across the practice.
Data quality
Naming, model standards, source records and project history become structured assets. Better records allow each future project to start with more usable context and less rework.
Differentiation
Generic AI makes everyone faster but more alike. Scaling AI around the firm's own judgement, data and standards keeps a defensible position: work that carries your expertise is what commodity tools cannot reproduce.
Browser-first work
Design, model viewing, data exchange, review and AI orchestration converge on shared browser environments, lowering IT and coordination cost as clients and consultants work in one place with fewer handoff errors.
Trust and liability
Winning systems carry source links, confidence levels, missing-data flags, change history and approval paths, turning liability from an adoption blocker into a controlled, insurable process the firm can sign off on.
Practice requirement
Firms need people who can direct and review AI, supported by work that is readable, reviewable and reusable. Together these disciplines allow project knowledge to carry forward.
04The operating layerWhat AI is onboarded into
AI needs project context, controlled commands and review.
A project workflow begins by giving AI the site context: address, parcel, LiDAR terrain, planning rules, hazards, client intent, material preferences, model state and source status. This gives later outputs a defined evidence base.
What the AI has to be onboarded into
01Site reality
Address, parcel, boundary, LiDAR terrain and source links are resolved first.
02Planning frame
Zones, overlays, hazards, constraints and consent pathways are made explicit.
03Client intent
The brief, priorities, assumptions, budget and open questions sit beside the site data.
04Model context
Massing, rooms, surfaces, edits and assumptions connect back to the evidence.
05Material context
Material direction, product options and surface intent become part of the design.
06Review state
Professionals can see what has been checked, trusted or left for judgement.
07Project memory
Corrections, approvals and source status stay with the project for the next AI task.
What this unlocksContext becomes an operating layer
A shared project context allows AI to support how the job is started, checked and carried forward.
Faster start
Site, rules and brief arrive together.
Cleaner decisions
Assumptions and risks stay visible.
Parallel checks
Planning, terrain and model health run together.
Reusable knowledge
Approvals and corrections carry forward.
05Controlled toolsThe dead end and the pivot
A design engine gives AI a controlled way to change the model.
An early prototype allowed AI to edit imported 3D models freely. The test exposed weak knowledge of site, rules, intent, object meaning and approval status, even when the visual file looked correct. A design engine addresses this by making those constraints part of the state AI works within.
The dead end
Treat the 3D file as the source of truth.
It feels efficient, and you make real progress at first: upload a model, chat to it, let AI edit geometry. But this is about as good as the approach gets. The system spends too much effort guessing what objects mean and too little making accountable design decisions.
The pivot
Make the design engine the source of truth.
Move the work into an AI-first design state: site context, terrain, planning rules, rooms, levels, roofs, materials, validation, diffs and approval history all live in the same place AI reasons from.
Old pathFree model editing
→
New pathAI-first design engine
What the new engine carries
Site context
Address, parcel, LiDAR terrain, planning rules and source status are brought in before the AI starts work.
Design state
Rooms, wings, levels, roofs, openings, decks and materials exist as project objects, not just pixels or mesh edits.
Controlled commands
AI uses named moves like create, move, fit to terrain, apply material and validate, instead of unrestricted geometry edits.
Review memory
Validation, diffs, approvals and failed checks stay with the project so the next AI task starts from better context.
06The same loopWhat kept recurring
Three prototypes use the same four-stage control loop.
Three systems built for different tasks independently converged on the same sequence: propose, validate, approve and retain the decision for the next run.
01
→
Propose
AI suggests through a few named moves.
02
→
Validate
The system checks it against the rules.
03
→
Approve
A person accepts, rejects or redirects.
04
Accumulate
The decision is kept and reused next time.
↻The decision re-enters as context, then the loop runs again
The hard part
The unresolved part of the loop is the human review step: a qualified, accountable person needs enough evidence to assess what the AI proposed.
01
Liability has a name
The professional who signs carries the risk, not the software.
02
Everything in one place
The change, the reason, the evidence and the rule impact, together and fast.
03
The judgement is specific
To this firm, this site, this client. It does not generalise.
04
Concept-ready is not consent-ready
The gap between plausible and signable is the whole problem.
Current prototypes do not yet resolve this review step cleanly. Further work needs to close the gap between a plausible output and one a professional can sign, while keeping the decision with the accountable person.
07Where this goes nextWhat to do about it
The next decisions concern workflow ownership, data and review.
The market is moving from AI experiments to workflow control. For a business owner, the issue is whether the firm has the workflows, data and review structures to use AI safely.
What matters now
Workflow ownership
Know which decisions AI can speed up, and which must stay with the team.
Data readiness
Project knowledge, past work, standards, costs and site context need to be findable and reviewable.
Review discipline
Approvals, evidence, change history and responsibility need to be built into the workflow.
Practical value
Start where the work is repeated, painful and commercially meaningful.
Where to start
Map real workflows
Look at how work is won, scoped, designed, reviewed, specified and delivered.
Test one workflow
Pick a bounded problem where better speed or consistency would be easy to measure.
Keep people accountable
Use AI to compress work around decisions, not to take ownership of judgement.
Build from evidence
Let the first workflow prove value before committing to broader systems.
Closing thought
Durable advantage comes from workflow discipline and reusable project knowledge.
Firms can move faster by making their work readable, reviewable and reusable by AI while retaining human responsibility for each decision. The value comes from compressing the work around professional judgement and carrying reliable project knowledge into the next task.
Matt Strawbridge · Landform ResearchAotearoa New Zealand · 2026