How AI Is Reshaping Architectural Planning and Production Drawings

From the first sketch to the drawing set, what has already changed, what is changing, and what has not.

This guide follows AI through the three stages of architectural work, from concept to schematic design to production drawings. It shows what has changed, what is changing and what has not, with dated adoption figures. It ends with a stage map and five steps to adopt AI without losing control of the drawing set.

AI in architecture has sped up the first two stages of work far more than the last. Concept images and layout options now arrive in minutes. Production drawings are where AI is newest, least proven and most important. The gain is real only when a practice rebuilds the work around checking. Responsibility stays where it has always been, with whoever issues the set.

The drawing set is still the contract

Production information is the set a project is built from. It holds plans, sections, elevations, details, schedules and the specification. It carries legal and commercial weight, because prices, schedules and disputes are all settled against it.

This guide follows the work through three stages. They map to the RIBA Plan of Work 2020 and to the AIA phases. The edges overlap, but the order does not.

• Concept. RIBA Plan of Work 2020: Stage 2, Concept Design. AIA phase: Schematic design.
• Schematic. RIBA Plan of Work 2020: Stage 2 into Stage 3, Spatial Coordination. AIA phase: Schematic design into design development.
• Production. RIBA Plan of Work 2020: Stage 4, Technical Design. AIA phase: Construction documents.

Four readers depend on the set. The contractor prices and sequences the work from it. The fabricator makes parts from it, usually through shop drawings. Building control, or the code official in the US, checks it against the regulations. The client signs it off and pays for what it shows.

What has already changed at the concept stage

The visible change is at concept. Image generators such as Midjourney and Adobe Firefly produce mood from text. Sketch-to-render tools turn a designer's own model into images. Examples are Veras by Chaos, SketchUp AI Render and the Archicad and Vectorworks AI Visualizers. Autodesk Forma Site Design tests massing on a real site. Massing means the rough 3D volumes of a building, before any detail.

How many architects use AI depends on who is asked and what counts as use.

• RIBA AI Report 2025. Date: 2025. What it measured: UK practices using AI. Figure: 59%, up from 41% in 2024.
• AIA with Deltek and ConstructConnect. Date: March 2025. What it measured: US architects who regularly use AI. Figure: 6%; 8% of firms have implemented AI.
• Architects' Journal survey. Date: October 2025. What it measured: Respondents using each tool. Figure: ChatGPT 77%; Midjourney 22%, down from 42% in 2024.

The figures measure different things. A practice counts as using AI if one person does. Regular use by individuals is far rarer, and firm-wide use is rarer still. The AJ survey also shows where everyday use sits. It is chatbots for bids, admin and research, not image generation.

Concept AI widens the field. It commits to nothing.

What is changing in schematic design

Schematic tools work from rules. The designer sets site boundaries, area targets, adjacencies and limits. An adjacency is a rule about which rooms must sit next to each other. The tool returns options that respect the rules, and the designer chooses.

TestFit runs site and parking solvers for feasibility studies. Finch3D, Snaptrude and ArkDesign generate and test plans and massing against a program. Hypar has moved to space planning, and qbiq makes office fit-out layouts from an uploaded plan.

Autodesk Forma announced Building Layout Explorer on 2 June 2026. It generates and scores floor plan options from a massing model. It is experimental. It is open only to commercial Forma Site Design users with data stored in the US.

Rules go in and options come out. The speed is real, and so is the dependency. An option is only as good as the rules typed in.

What is beginning in production drawings

BIM has automated part of production for two decades. BIM, or building information modeling, keeps the whole design in one model. In Revit or Archicad, plans, sections and schedules are views of that model. So a moved wall moves on every sheet.

BIM still left annotation and sheet layout to people. That means the tags, dimensions and views placed on each sheet. This is where the new layer of AI tools works.

• Swapp. What it does, in the vendor's terms: Construction document sets, with dimensions, tags, views and sheets, inside Revit and Archicad. Status: Commercial product.
• Glyph by Chaos. What it does, in the vendor's terms: Views, sheets, tags and dimensions in Revit. Status: Commercial product.
• Autodesk Assistant in Revit 2027. What it does, in the vendor's terms: Acts on the model from prompts; AI-driven changes are labeled in the Undo menu. Status: Tech Preview since April 2026.
• AutoCAD 2026 Smart Blocks. What it does, in the vendor's terms: Detects repeated objects and converts them to blocks. Status: Detect and Convert is a technology preview.
• Graphisoft AI Assistant. What it does, in the vendor's terms: Assistant built into Archicad. Status: Beta.

For an independent view, read Martyn Day's AEC Magazine review of Swapp. He found it suited to predictable building types, with output that still needs review.

What has not changed

Responsibility has not changed. The person who issues the set answers for it, whatever drew it. Rule 4.102 of the AIA Code of Ethics puts it in one line. "Members shall not sign or seal drawings, specifications, reports, or other professional work for which they do not have responsible control."

Coordination with structure and services is still human work. So are site conditions, the reading of codes and the judgment about what the client really needs. Machine errors make checking harder, because they are confident errors. A wrong dimension looks exactly like a right one.

The sober counterweight comes from Deltek's Clarity A&E study of May 2026. AI adoption among architecture and engineering firms rose from 53% to 70%. Yet only 38% reported a measurable positive business impact.

Temporary architecture as an early test

Exhibitions and events squeeze the whole cycle into weeks. They work on fixed venue plans, under rules that repeat from show to show. Aisle widths, exit clearances and height limits come up every time. That is why AI reached drawings there early. Some event tools already build a to-scale layout on an uploaded venue plan. They draw plans, elevations and renders from that one layout, and re-run the layout checks after every change.

How to adopt AI in architecture without losing the plot

Start from a map of what each stage asks.

• Concept. Task: Explore directions. Tool category: Text-to-image, sketch-to-render, massing. Example tools: Midjourney, Veras, Forma Site Design. Typical output: Images, massing. What a person must still check: Scale, site truth, what the image commits to.
• Schematic. Task: Fit the program. Tool category: Test-fit and layout solvers. Example tools: TestFit, Finch3D, Snaptrude. Typical output: Plans, area schedules, options. What a person must still check: The rules entered, adjacencies, code assumptions.
• Production. Task: Document the design. Tool category: Annotation and sheet automation. Example tools: Swapp, Glyph, Autodesk Assistant. Typical output: Sheets, tags, dimensions, schedules. What a person must still check: Every dimension, tag, clearance and revision.

Then adopt in five steps.

1. Choose one stage. Start where the hours are, not where the demos are.
2. Baseline the hours. Take the current hours for that stage from timesheets, so the gain is measured rather than felt.
3. Write the check before the tool arrives. Decide what a person verifies on every output. Our eight-point protocol for machine-drawn sets is a starting point.
4. Settle data and IP with each vendor. Get in writing who owns uploaded drawings and whether the vendor trains models on them.
5. Retrain juniors as checkers. Teach them to mark up and explain a detail, not only to draw it.

Tools at each stage are mapped in AI tools by workflow stage.

The hours move to judgment

The gain from AI is the hours moved from drafting to judgment. The responsibility does not move at all.

Frequently asked questions

How is AI used in architecture today?

Mostly for text. In the Architects' Journal survey of October 2025, 77% of respondents used ChatGPT, mainly for bids, admin and research. Midjourney use fell to 22%. Concept images, layout solvers and drawing tools are growing, but they are still less common than chatbots.

Can AI produce construction documents?

Partly. Tools such as Swapp and Glyph place views, sheets, tags and dimensions inside Revit or Archicad, and Autodesk Assistant in Revit is a Tech Preview. These tools draft annotation. They do not coordinate structure and services or read codes, so every sheet still needs a qualified checker.

Who is responsible for drawings made with AI?

The person who issues them. The AIA Code of Ethics bars members from signing or sealing work they do not have responsible control over. The ARB Architects Code expects skill and diligence. The software that drafted a line carries no professional duty.

Which stages of architectural work has AI changed most?

Concept and schematic design. Image, massing and layout tools now produce many options quickly. Production drawings are where AI is newest and least proven. Several drawing features are still previews or betas, and an error costs the most at that stage.

Events squeeze this whole cycle into weeks, which is where Monet works. It builds a to-scale layout on your venue plan and draws production-ready plans and elevations from it. Its exit and walkway checks re-run after every change, so checking runs through the job instead of waiting for the end.