BIM, AI and the Estimate: What We Actually Do With Revit, ArchiCAD and AutoCAD

Every job we build starts as a drawing, and most of the slow, expensive mistakes start there too. The permit set that comes back with corrections. The renovation where nobody has an accurate plan of what is already standing. The estimate that missed a wall because the takeoff was done from a PDF at the wrong scale.

For two decades the fix was more people: more drafters, more estimators, more hours. This year we started doing it differently. We connect an AI agent (Anthropic’s Claude) directly to our design software, let it do the repetitive model work, and keep the judgment, the review and the signature with people. Here is what that looks like on real jobs, with the numbers we have and the ones we don’t have yet.

The bottleneck is the drawing

A general contractor lives downstream of drawings. Permits need them. Estimates come from them. Subs price from them. And on renovation work, the drawing of the existing building usually does not exist, or exists as a scanned PDF from another decade.

Three programs run most of this world:

• AutoCAD is still the universal language of 2D drawings. Almost every architect, engineer and sub can open a DWG.
• Revit is the dominant BIM platform on larger commercial work: a single model where a wall is a wall, with height, material and cost, not just two lines.
• ArchiCAD is the other major BIM platform, strong with architects and smaller firms. It is the one we run in-house.

The difference that matters for an owner: in 2D, a change means redrawing every sheet that shows it. In BIM, the change is made once and the plans, schedules and quantities follow. That is also why BIM is where AI becomes useful. An AI can read and edit a structured model. It is much worse at guessing what a pile of lines means.

Where Revit, ArchiCAD and AutoCAD stand on AI

As of November 2026, all three vendors are building the same idea: an assistant that lives inside the model.

AutoCAD 2026 added AI-assisted Smart Blocks, which find repeated objects in a drawing and convert them into blocks, and Markup Assist, which reads markup on an imported PDF or image and can carry out simple instructions such as move, copy or delete. Autodesk Assistant answers questions without leaving the program.[1][2]

Revit 2027 brought Autodesk Assistant to Revit. It can color and tag elements and apply schedule filters on request. Autodesk added two controls that matter to anyone who has had software change a model behind their back: AI-made changes are labeled in the Undo menu, and there is a Stop button.[3][4] At Autodesk University in September, Autodesk previewed more: drawing compliance checks that flag issues like missing wall dimensions, analysis of what changed between drawing revisions, automated documentation, and a daily log agent that turns a superintendent’s spoken update into a structured log. Availability for most of it has not been announced.[5]

ArchiCAD moved fastest toward letting outside AI in. Archicad 29 shipped with an AI Assistant in beta.[6] Archicad 30, released this fall, adds an Archicad MCP server in beta, a standard connector that lets other AI tools, including local ones, query the model, edit properties and classifications, and control views. Graphisoft is careful about one limit: its agent “cannot create, move or delete model elements.”[7][12]

The pattern is clear. The vendors are putting an AI between you and the model. We did not wait for the polished version. We built the connection ourselves.

Job one: a demolition plan, drawn by two hands

The job was an interior demolition permit for a condo unit: kitchen out, two tubs out, a bar wall removed, a game room turning into a bedroom. The building department needed a clean demolition plan.

Our first attempt was the wrong one, and it taught us the most. We had the AI generate the drawing directly as lines in a file, with no real CAD behind it. It looked amateur. No plan reviewer would take it seriously, and we threw it out.

The second attempt split the work by what each side does best. The model lived in ArchiCAD 29. ArchiCAD has a built-in programming interface, and a free add-on extends it. Through that connection, the AI handled the bulk operations: creating layers, moving walls onto them, tagging the demolition zones, placing labels, and filling the title block. A person traced the walls, because that interface cannot create walls, and built the final sheet, because it cannot set up a layout or plot a PDF.

By our count, roughly 70% of the model operations ran through the AI. The manual remainder was about half an hour of layout and output work. The package around it, cover letter, demolition notes, legend and checklists, was prepared overnight while nobody was in the office.

Job two: rebuilding a house inside the model

The second job was harder: an impact window and door replacement on a two-story house. To permit it properly we needed an accurate plan of a house that only existed as the original architect’s PDF sheets.

The AI read the original floor plans, detected the double lines of the interior walls at a calibrated scale, and created 81 interior walls in the model. We traced the main walls of both floors by hand, and we placed and reviewed every window and door mark ourselves. Then we laid the original sheets under the model as a calibrated image and checked wall by wall. Typical deviation: 4 to 8 inches.

Two results mattered more than the speed.

First, the window and door schedules now come from the model itself. Change a window in the model and the schedule on the sheet changes with it. No retyping, no version where the plan says one thing and the table says another.

Second, a tag-by-tag audit against the original sheets caught a real error. We had drawn a glass block run on the wrong wall. The original showed four panels along the south side, the zero-lot-line side. It was fixed in the model and on every sheet before anything was priced or ordered. That kind of mistake normally shows up when the material arrives.

BIM and AI for estimates

An estimate is only as good as its quantities, and quantities are only as good as the drawing they came from. That is the real reason BIM matters to a contractor: when the model is right, the counts of walls, openings, areas and fixtures come out of it instead of out of a highlighter and a scale ruler.

The takeoff market is moving fast. AI-native platforms read 2D plans and count elements for you. Togal.AI, one of the most reviewed, starts at $299 a month per user.[8] Even the industry’s own buyer’s guide says to test any AI takeoff on your own plans, not a scripted demo, and to ask the vendor one question first: what happens when the AI is wrong?[8]

That question is the heart of how we use AI on estimates. Speed is easy. Accuracy comes from three habits:

1. Every line says where it came from. Each line of our estimates is tagged as verified (live supplier price or a cited cost source), assumed, or pending. Assumed and pending numbers never get mixed into a total without a flag.
2. The AI lists what it does not know. On an eight-unit, four-building project, the estimate went out with four open items stated up front: no soil report yet, an expired roofing product approval, a pending utility load letter, and a missing fee invoice on one building. Each one could move the price. Each one was in front of the owner before the number was.
3. People sign, AI does not. Engineering is sealed by a licensed engineer. Prices are confirmed by suppliers. The AI prepares, checks and flags.

The owner’s math

Here is the honest part. We have not timed these jobs with a stopwatch. What we can document is the 70% of model operations, the half hour of manual remainder, the 4-to-8-inch overlay, and the glass block error caught before ordering. Starting this month we log every drawing and every estimate: hours, revisions, and errors caught before they reached the field. We will publish those numbers here.

The cost side of our setup is small. ArchiCAD was a license we already owned. Its programming interface is built in, and the add-on we use is free. The rest is an AI subscription. If you want to run the math for your own company, it is one line:

Payback in months = monthly cost of the tools ÷ (hours saved per month × your loaded hourly cost)

Count the errors too. One window ordered for the wrong wall, one rejected permit set, one missed wall in a bid, usually costs more than a year of software.

What it can’t do

• It does not replace CAD skill. The AI-only drawing was rejected. Good results came from a real model with a person driving it.
• The interfaces have hard limits. In our setup the AI could not create walls, set pens or plot sheets. Graphisoft’s own agent cannot create, move or delete elements.[7]
• Scanned drawings lie. Reading old PDFs needs scale calibration and a human overlay check. Four to eight inches is fine for a permit plan and not fine for ordering glass.
• Vendor AI is still early. Several features are previews, tied to specific subscription plans or limited by monthly AI credits; Archicad Collaborate includes 50 per license per month.[7]
• Nobody seals with AI. Licensed professionals still sign and carry responsibility for the drawings.
• Check where your data goes. Read each vendor’s terms. Graphisoft states that it does not use project content to train its models.[7]

What comes next

The next step is checking before submitting. Drawing compliance review and change analysis between revisions, both previewed by Autodesk this fall, point at the same place we are heading on our own: an AI that reviews the permit set against the code and the plan reviewer’s last comments before it leaves the office.[5] The firms that win the next five years will not be the ones with the most drafters. They will be the ones whose model, estimate and paperwork agree with each other.