Use one general assistant, ChatGPT or Claude, at every stage, and add a specialist tool only where a stage has a clear job for it: visualisation at concept, a PDF review tool at tender, a note-taker on site, nothing extra at handover. At list prices a five-person practice pays about $117 to $434 a month.
The pages that rank for this topic are mostly design and visualisation tools, which cover a slice of one stage. A practice's week is also briefs, specifications, tender queries, site minutes and handover paperwork, where the useful tools are less glamorous. Prices below come from each vendor's own pages as of October 2026; where a page would not load, I give no figure.
Five stages, one assistant, and a short list of add-ons
| Stage | Job for AI | Tools worth paying for | Human check |
|---|---|---|---|
| Concept | Test the brief, produce options and mood images | ChatGPT or Claude; one visualisation tool | Brief contradictions resolved with the client; images labelled as concept |
| Design | Specifications, consultant coordination, option visuals, drawing checks | The assistant; in-model visualiser if you already use one | Every clause and every drawing flag read by an architect |
| Tender | Query log, tender comparison, arithmetic and completeness checks | The assistant; PDF markup software | Assessment and advice on who should win stay with the architect |
| Site | Minutes, action lists, site reports, task tracking | Note-taker; task app; the assistant | Minutes read and issued by the person who attended |
| Handover | Document index, user guide, snag list | The assistant; the task app from site | Each product fact checked against the manufacturer's document |
Concept: from a loose brief to options you can show
The first job is not drawing. It is finding out what the client has asked for and where it contradicts itself. Put the brief, the site notes and the client's emails in one assistant project and ask for a room data sheet plus a list of conflicts:
From the brief and emails below, build a table of spaces: name, required area, headcount or use, adjacencies, services needs. Then list every contradiction or gap: where a stated area cannot hold the stated use, where two documents disagree, and where something is mentioned once and never specified. Quote the source for each. Do not propose design solutions.A typical reply catches that the client wants a studio for 40 people in 38 square metres, that the budget in the email differs from the one in the brief, and that "kitchenette" appears in the plans but never in the room list. Each of those is a conversation worth having in week one. The detail is in how to use AI for concept options in a small practice.
For images, the tools worth checking are these. Prices are as the vendors list them.
| Tool | What it does at concept | List price | Check before buying |
|---|---|---|---|
| Veras (EvolveLAB) | Renders from inside Revit, SketchUp, Rhino, Archicad and Vectorworks, or on the web; a Geometry Override slider sets how far the image may depart from your model | Named licence $59 a month, or $29 a month billed yearly ($348); floating licence $51 a month billed yearly; 30 free renders and a 15-day trial | Software versions supported; the page shows a dollar sign without naming the currency |
| Snaptrude | Browser-based early modelling with an AI design agent | Free ($0, 75 credits a week, one project); Pro $60 a month or $50 billed yearly ($600); a team tier at $100 a seat a month billed yearly (one layout of the page drops the per-seat wording) | The pricing page shows two layouts that disagree on which plans export to BIM and Revit; test the export before you rely on it |
| MyArchitectAI | Renders from a SketchUp or Archicad design, style transfer, render editing, video | "Plans from $29/mo"; 10 free renders | Per-plan limits were not shown on the pages I could read |
| Adobe Firefly | Editing and extending images, including Generative Fill | Standard $9.99 a month (2,000 credits); Pro $19.99 (4,000); Pro Plus $49.99 (10,000) | Adobe says outputs from its own Firefly models are safe for commercial use; partner-model outputs are your responsibility to assess |
| Archicad AI Visualizer | Generates images and textures from sketches, model views and prompts inside Archicad | Graphisoft lists it with Archicad 30 on the Studio, Collaborate and Max plans | No render quota or separate price is stated; confirm your plan includes it |
Midjourney is the other name you will hear for concept imagery. Its pricing page did not load when I checked, so I quote no plan or rule for it; confirm current plans and its terms for company use before relying on it. The Midjourney tutorial listed under further reading goes deeper.
Two rules for concept images. Label them "concept image, not a design" in the file name and on the slide, because a render shows materials you have not chosen and details nobody has drawn. And set the Veras slider low when you want the image to respect your massing, and high only when you want loose inspiration. A client who sees a realistic picture remembers it as a promise.
Where it goes wrong: the client falls for the render. Suppose a Veras image shows a full-height glazed corner that the structure cannot support, and the client asks for it at the second meeting. The fix is procedural. Show a render only beside the massing view it came from, caption it, and keep the unrealistic ones off the slide.
Design: specifications, option visuals and drawing checks
At design stage the volume of text goes up, so the assistant earns most of its fee here. Three jobs repay the effort.
- Specification drafting and consistency. Give the assistant the product data you have chosen and ask it to draft clauses in your house format, then run a second pass for contradictions between clauses and schedules. A sample finding: "Clause 12.3 specifies oak veneer; the finishes schedule says ash for the same door." The tutorial on specs and planning reports with Claude has prompts.
- Option visuals from the model. If you work in Archicad 30 on a Studio, Collaborate or Max plan, the AI Visualizer is built in. Graphisoft describes it as turning sketches, models and ideas into visual output directly from Archicad. Treat that output as concept imagery: it is no substitute for detailed rendering with the materials you have actually specified. In Revit or SketchUp, a plugin such as Veras does the same job for roughly $29 to $59 a month per named user.
- Checking the drawing set. A general assistant reads extracted text and exported schedules to flag index errors, contradictory notes and schedule mismatches. The method and a seeded test are in AI checks for notes, schedules and sheets. Bluebeam's Max plan lists AI drawing review at $590 per user per year, billed annually, shown as an introductory price; its Smart Review looks for issues such as missing sheets and missing door or plumbing fixture tags, and Bluebeam says it is in preview. The Basics, Core and Complete plans are $260, $330 and $440 and carry no AI review.
Whatever the tool, the risk at this stage is mismatch: a visual that shows a detail the drawings do not. Keep a rule that no AI image leaves the practice without a sentence stating what it represents.
Where it goes wrong: a drafted clause cites a standard that has been replaced, or a product that has been discontinued. The assistant does not know which standards apply to your project unless you tell it. Ask it to list every standard and product it cites, in a table, and have someone confirm each against the current source before the specification goes out.
Tender: documents out, tenders back
Tender stage is where an error becomes someone else's pricing. The assistant is useful on both sides of it.
Going out, ask it to read the specification and the drawing schedules together and list items mentioned in one but not the other, because those become tender queries. Keep a query log in a spreadsheet with a number, date, question, answer and whether it needs an addendum, and use the assistant to draft clear answers for your review. For issuing the set and collecting markups, PDF software such as Bluebeam is the usual home. Bluebeam lists "Overlay and compare drawings" from its Core plan ($330 a user a year) upwards, which helps when an addendum changes the drawings; the AI comparison across disciplines is a Max feature.
Coming back, give it the returned pricing summaries and the contractors' qualifications and ask for arithmetic and completeness checks. In this made-up round, four tenders arrive and the assistant is asked to re-add each priced summary:
| Tender | Stated total | Re-added from elements | Qualifications flagged |
|---|---|---|---|
| A | $388,000 | $388,000 | Excludes temporary power |
| B | $402,500 | $402,500 | None |
| C | $412,300 | $421,300 | Assumes site access from 7am |
| D | $395,900 | $395,900 | Excludes scaffolding; provisional sum for glazing |
Tender C's stated total is $9,000 below its own elements, which is a transposed digit in a summary. That is a conversation to have with the tenderer before assessing anything. Tender D looks cheapest on paper only until the scaffolding exclusion is priced, and the assistant cannot price it for you. It can, however, list every exclusion so a person can compare like with like. For the contractor-side view of the same task, see what AI catches and misses in bid levelling. Never let an assistant recommend a winner. The architect assesses the tenders and advises the client.
Where it goes wrong: image-only returns. If a tenderer sends a scanned pricing summary, an assistant may read nothing from it or read it badly, and the arithmetic check looks clean because nothing was checked. Say in the tender instructions that pricing must come back as a spreadsheet or a text PDF, and ask the assistant to list what it could not read.
Site: minutes, tasks and photos
On site the work is recording and chasing. A note-taker and a task app earn their place here.
| Tool | Job | List price | Check before buying |
|---|---|---|---|
| Otter Pro | Transcribes meetings and drafts a summary | $8.33 a user a month billed annually, $16.99 monthly; 1,200 minutes a month; 90 minutes per meeting | A two-hour site meeting will exceed the 90-minute cap |
| Otter Business | The same for longer meetings | $19.99 a user a month annually, $30 monthly; 4 hours per meeting | Whether you need it for every user or only the site lead |
| Fieldwire Basic | Task management, plan viewing, photos, checklists | Free for up to 5 users, 3 projects and 100 sheets | Limits bite on a busy practice |
| Fieldwire Pro | Adds reports, exports and sheet compare | $39 a user a month billed annually | Its Business Plus tier at $89 adds RFIs, submittals and change orders |
Tell everyone before you record a meeting and check the rules on recording consent where you work. Then turn the transcript into minutes with a prompt that keeps the assistant honest:
From this transcript, write minutes for a site meeting. Sections: attendees; decisions; actions (owner, action, due date); open issues; next meeting. Use only what was said. If an action has no owner or no date, write "unassigned" or "no date". Do not summarise opinions as decisions. Flag anything you could not hear clearly.The person who attended reads the output, corrects it and issues it the same day. The tutorial on same-day architect meeting minutes goes through that routine, and the site visit report tutorial listed under further reading turns photos and notes into a report.
Where it goes wrong: a transcript attributes a decision to the wrong person, or records an action with no owner, and the 90-minute cap on a Pro-tier plan cuts the recording off before the last agenda items. Read before issuing, and compare the tool's limit with your longest meeting before you choose the tier.
Handover: O&M files, user guides and snag lists
Handover is mostly collection and translation, which suits an assistant. Give it the contractor's operation and maintenance files and ask it to build an index by system, listing the certificate or manual each one needs, and then flag what is missing. The check is simple: every item the assistant lists must exist in the folder, and every document in the folder must appear in the index.
The second job is a plain-English guide for the client. Before, in typical manual wording: "Self-closing devices must be maintained in working order and fire doors must not be wedged open." After, in the user guide: "Fire doors close by themselves. Please never wedge them open, and tell us if one stops closing properly." Ask the assistant for the second version and then check each fact against the manufacturer's document, because a friendly sentence can still be wrong.
The third job is the snag list, where photos and notes become a prioritised list. Can AI write a snagging list from site photos? covers what works and what to verify. None of these needs a new subscription if you already hold the assistant and the task app from site.
Where it goes wrong: the friendly sentence is wrong. A guide that says the warranty lasts ten years when the manufacturer's cover is two is worse than no guide. Every number in a user guide needs a source page, which you note in the margin of your draft and delete before issue.
One project through all five stages: what the middle stack did
The figures in this walk-through are invented. A five-person practice takes on a refurbishment of the upper floors of a small shop into two flats, and runs it on the middle stack costed in the next section. Site meetings run to about 75 minutes, so Otter Pro's 90-minute cap is enough.
| Stage | What happened | Tools | Human check |
|---|---|---|---|
| Concept | Brief and emails tabulated into a room data sheet; 6 gaps and 2 contradictions found; two massing options rendered from a simple SketchUp model with the Veras slider set low | Claude project; Veras | Gaps taken to the first client meeting; renders captioned "concept" |
| Design | Specification clauses drafted from chosen product data; a consistency pass found 4 conflicts with the finishes schedule; the drawing-set check produced 11 flags, 8 of them real | Claude; spreadsheet | Architect resolved all 4 conflicts and opened every drawing flag |
| Tender | 31 tender queries logged and answered; four returns re-added; one summary total $9,000 out; exclusions listed per tenderer | Claude; spreadsheet | Architect assessed the returns and advised the client |
| Site | 14 site meetings recorded and minuted; 62 actions tracked to closure | Otter Pro; Fieldwire Pro; Claude | Minutes read and issued the same day by the person who attended |
| Handover | O&M index built by system, 9 missing items chased; a six-page user guide drafted; 48 snags listed and prioritised | Claude; Fieldwire | Every product fact in the guide checked; snags confirmed on site |
Notice what the tools did not do. None produced a drawing, assessed a tender or certified work. They read, compared, drafted and recorded, and a person made every decision. That is the line to hold at every stage:
- Concept: choosing the design direction and what to promise the client stays with the architect.
- Design: signing drawings and specification, and deciding which standards apply, is never delegated.
- Tender: assessing returns and recommending a contractor is a professional judgement.
- Site: issuing instructions, assessing work and certifying anything is for the person who is answerable for it.
- Handover: certifying completion and approving what the client is told about warranties and maintenance stays with you.
How these duties interact with your contracts and insurance is a question for your own policy and adviser, not for a tutorial or a tool.
Three costed stacks for a five-person practice
These are list prices as of the pages above, using annual billing where a vendor offers a discount. A practice of five: two architects, two technologists and an administrator.
| Stack | What is in it | Per month | Per year |
|---|---|---|---|
| Lean | 5 assistant seats at $20; Otter Pro for 2 at $8.33; Fieldwire Basic free | $117 | $1,400 |
| Middle | Lean, plus Veras for 2 at $29; one Bluebeam Max seat ($590 a year); Fieldwire Pro for 2 at $39 | $302 | $3,622 |
| Fuller | 5 assistant seats; Otter Business for 2 at $19.99; Veras for 2; two Bluebeam Max seats; Fieldwire Pro for 2; Snaptrude Pro $50; Firefly Standard $9.99 | $434 | $5,212 |
The assistant seats assume a business plan with a two-seat minimum. ChatGPT Business and Claude Team both list at $20 a seat billed annually or $25 monthly, so five seats on monthly billing is $125 rather than $100. Claude's own page lists Pro at $20 monthly or $17 a month billed annually ($200 up front), Team Standard for two to 150 seats, and Enterprise at $20 a seat billed annually with a 20-seat minimum plus usage at API rates.
Monthly versus annual billing changes the answer for project-based tools. Veras costs $59 a month on monthly billing or $348 a year on annual. Five months of monthly billing is $295, six is $354, so if you render in fewer than six months of the year, monthly billing is cheaper. Otter Pro has the same shape: annual billing is about $100 a year, and monthly billing overtakes it in the sixth month ($101.94). Snaptrude's break-even is ten months ($60 a month against $600 a year).
Before buying, put your own hourly rate against the stack. At a charge-out rate of $100 an hour (put your own rate in its place), the lean stack's $117 a month is repaid by about 1.2 hours of saved time across the whole practice each month. The middle stack's $302 needs about 3.0 hours and the fuller stack's $434 about 4.3. Those are break-even points, not promises. Log the hours you actually save for two months and drop anything that does not clear its own line.
Checking a vendor's plan before you pay
Plans change monthly, and pricing pages disagree with themselves. Use a five-minute routine for each tool:
- Open the pricing page in a private window and note the currency and billing period. Several pages show a dollar sign without naming the currency.
- Read the feature rows for the specific thing you need, such as export to Revit or meeting length, not the headline.
- Look for contradictions on the page itself. Snaptrude's page shows two layouts that disagree on which plans can export, which is exactly the kind of detail that costs a week.
- Note whether a price is introductory. Bluebeam labels the Max price that way and says it holds through a renewal in 2027, with the standard price to be announced later, so ask what it becomes.
- Read the terms on data and output rights. Business plans from OpenAI and Anthropic do not train on your content by default, while consumer plans have a privacy setting you must review. Read the copyright and client-data tutorial listed under further reading before anything confidential goes in.
- Screenshot the page with the date, and put it in your records.
What to buy first, and what to leave alone
- Buy the assistant first. It touches every stage and costs the same whether you use it twice a week or daily. Run the same specification and minutes task through ChatGPT and Claude on your own documents, then pick one.
- Write the policy second. Decide which projects and documents can go in, who approves a new tool and how outputs are marked, before more tools arrive.
- Add the note-taker when site meetings become your biggest admin load. Mind the 90-minute cap on Otter Pro.
- Add visualisation if you already show clients images often. Start with the free trial or free renders, and use the one built into your modelling software if it has one.
- Add PDF review last. Bluebeam's AI review is in preview, so run a seeded test before paying for it.
Leave alone anything that promises to produce drawings, certificates or compliance sign-off. Those stay with the architect. If you want a staged introduction, the first-quarter pilot plan listed under further reading shows how to sequence it.
Questions about an AI toolkit for architects
Do I need a different AI tool for every project stage?
No. One general assistant, ChatGPT or Claude, covers briefs, specifications, minutes, tender queries and handover guides. Add a specialist only where a stage has a job the assistant cannot do well, such as in-model visualisation at concept, PDF review at tender or a task app on site. Start with the assistant and add one tool at a time.
Are AI renders safe to show clients?
Show them as concept images, never as drawings. An AI image is no substitute for detailed rendering with the materials you have specified, and it can show finishes you have not chosen. Adobe says outputs from its own Firefly models are safe for commercial use but that you judge partner-model content yourself.
How much should a five-person practice budget?
In the worked example, a lean stack costs about $117 a month, a middle stack about $302 and a fuller stack about $434, all at list prices with annual billing where offered. Prices change, and several vendors list different figures for monthly and annual billing, so check each page before you commit.
Can AI check my drawings or the tenders I receive?
It can act as a second reader, not a final check. For drawings, extracted text and exported schedules work better than screenshots. For tenders, an assistant can recompute arithmetic and list qualifications, but the architect assesses and advises on who should win. Verify every flag before you act on it.
Is ChatGPT or Claude better for an architecture practice?
Both list at $20 a month for an individual plan, and the business tiers are $20 to $25 a seat with a two-seat minimum. Choose by running the same brief, specification and minutes task through each on your own documents, then standardise on one so prompts, projects and policies are shared.
Further reads
- How Much Does AI Cost a Small Architecture Practice per Month? — What AI costs a practice each month, tool by tool
- Claude vs ChatGPT vs Gemini for Architects: Which to Pay For — Choosing the general assistant for the whole practice
- Copyright and Client Data Risks of AI in Architecture Practices — Image rights and client data before you upload anything
- AI Policy for Architects: A Ready-to-Use Checklist and Clauses — A policy that tells staff which tools they can use
- Architect Site Visit Report: Notes and Photos to a Clear Report With AI — Turning site visit notes and photos into a clear report
- AI Floor Plan Generation for Architects: Where It Fails a Real Brief — A test protocol for architects: write one real brief as a ledger, score a generated plan against it, and set the sign-off rules before clients see anything.
- AI Marketing for Architects: Case Studies, Website Copy and LinkedIn — Turn a site-visit voice note into a case study, service-page copy and LinkedIn posts that sound like your practice, with permission and claims checks.
- AI Programs for Architects: Room Data Sheets From a Client Brief — A worked example of turning a messy client brief into room data sheets with AI, and the review pass that catches what the draft invents or drops.
- AI for Architects: Draft Design Schematics From a Client Brief — A brief-to-schematic workflow for small practices: stress-test the brief, draft an area schedule and options, then run the checks that stay with the architect.
- Architect Website: Project Descriptions and Service Pages Written With AI — A copy workflow for architects: project descriptions from verified facts, service pages by project type, honest area pages and an FAQ on fees and process.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: vendor pricing, product and help pages read in October 2026 for ChatGPT, Claude, Bluebeam, Veras, Snaptrude, Adobe Firefly, Otter, Fieldwire, MyArchitectAI and Graphisoft (AI solutions page). Pages that would not load are not quoted.