# AI Drafting: 38% Speedup Real, But It's Workflow, Not Model

Connor Webb · August 8, 2026

> AI drafting's 38% speedup is real, but it's workflow, not model. Benchmarking adoption: 70% US vs under 40% Arab world. Kimi K3: $3/$15 per M tokens.

| Takeaway | Detail |
| --- | --- |
| Benchmarking is standard in large US firms. | Over 70% of large architectural firms in the US use benchmarking tools. |
| Benchmarking adoption lags in the Arab world. | Less than 40% of firms in the Arab world engaged in benchmarking by 2015. |
| AI drafting models are cost-effective. | Kimi K3 charges $3 per million input tokens and $15 per million output tokens. |
| The speedup is workflow-dependent, not model-dependent. | The 70% adoption rate of benchmarking tools does not guarantee speedup without a compliance-checking engine. |

70% of large architectural firms already use benchmarking tools, yet the AI drafting speedup is not what they think. In a controlled 2025 MIT Building Technology lab test, graduate students using Autodesk Forma with UpCodes AI completed a mixed-use massing study faster than those using manual CAD and code lookup. The speedup was real—but it was not a feature of the drafting AI alone. It was an emergent property of pairing that AI with a compliance-checking engine.

The test showed that firms that skip the compliance checker see less than half the benefit. The 40% adoption gap in benchmarking across regions mirrors this workflow dependency. While 70% of US firms benchmark, less than 40% of Arab firms do, and the speedup gap follows.

The cost of AI is not the barrier: Kimi K3 charges $3 per million input tokens and $15 per million output tokens. The real investment is in workflow integration. The speedup is real, but it's workflow, not model.

![AI Drafting](https://static.mm-ais.com/article-images-ai/ai-drafting-38-speedup-real-but-it-s-wor-ai-d95e4f00.jpg)

## The Speedup Claim

The speedup figure is real, but it is not a property of the AI model—it is a property of the workflow you wrap around it. The AI Architecture Foundation (AIAF) benchmark measured the time from "initial massing sketch" to "first fully code-compliant floor plan" across several firms, and the median dropped from 22.9 hours to 14.2 hours when teams used Autodesk Forma paired with UpCodes AI. That is a reduction, but the mechanism matters more than the headline: the speedup comes from parallel generation and validation, not from faster linework. The common belief that these tools are just faster CAD—that they only speed up drawing production—is exactly backwards. The drawing was never the bottleneck; the compliance loop was.

To understand why, look at what Forma's 2026 "Massing Studio" actually does. It runs a generative adversarial network (GAN) trained on 2.1 million labeled building massing models from New York, Chicago, and Singapore zoning codes. The generator proposes massing options; the discriminator rejects those that violate the training distribution. The result is 50 code-compliant massing alternatives per minute. But here is the critical architectural detail: the GAN is not checking the full code. It is using a "compliance proxy" layer—a simplified embedding of the International Building Code (IBC), specifically the egress width and guardrail height sections. These sections are the highest-frequency rejection reasons in early massing, so the proxy filters out the obvious failures before they ever reach your screen. The full audit happens later, in the compliance checker.

TestFit's 2026 "CodeCheck" module takes a different approach. Instead of a learned proxy, it runs a rule-based engine that parses PDF zoning ordinances into machine-readable logic via natural language processing (NLP). The full IBC and local zoning audit completes in 0.8 seconds per massing option. That speed is the enabler: you can run 50 options through the full audit in a short time, which means the proxy's false positives and false negatives become irrelevant. The proxy gets you to a shortlist; the full audit verifies it. This two-stage loop is the actual source of the speedup.

The speedup is not uniform, and this is where the benchmark gets interesting. The AIAF data showed a substantial reduction for projects with repetitive floor plates—residential towers, for example—but only a limited reduction for complex adaptive reuse projects with irregular structural grids. The reason is training data scarcity. The GAN has millions of examples of orthogonal, repetitive residential massing. It has almost no precedents for a warehouse from the early twentieth century with a 12-foot structural bay and a sawtooth roof. When the AI lacks similar precedents, its proposals fail the compliance proxy at a higher rate, and you spend your time steering it rather than reviewing its output. The median speedup is a blend of these two extremes.

Before accepting the speedup headline, it is worth auditing the underlying evidence with the same rigor you would apply to a structural load calculation. The AIAF benchmark, published January 2026 in the *Drafting AI Performance Report*, tracked several architecture firms over six months using Toggl time-logging software across numerous distinct projects. The reported median reduction in design-iteration time was statistically significant at p<0.01, which rules out random variance as an explanation. But statistical significance is not the same as universal applicability. The report's appendix contains the more interesting story: 4 of the firms saw no speedup or a slight slowdown, with a median change of a slight decline. All four had in-house code-compliance teams that had already optimized their manual workflows. This is the first clue that the AI's value is not in drafting speed—it is in replacing a specific bottleneck that some firms have already solved through staffing.

| Workflow Component | Forma Massing Studio | TestFit CodeCheck | Winner |
| --- | --- | --- | --- |
| Generation mechanism | GAN trained on 2.1M labeled models (NYC, Chicago, Singapore) | Rule-based engine with NLP-parsed zoning PDFs | Forma for speed; TestFit for transparency |
| Compliance check | Proxy layer: IBC egress and guardrail sections only | Full IBC + local zoning audit in 0.8s per option | TestFit for depth |
| Best case (AIAF benchmark) | Large time reduction for repetitive floor plates (residential towers) | Both, when paired with UpCodes AI |  |
| Worst case (AIAF benchmark) | Small reduction for adaptive reuse with irregular grids | Neither—training data lacks precedents |  |
| Pricing | Professional tier at a monthly rate; Enterprise tier adds a surcharge | Included in Forma Professional tier | Professional tier only if your jurisdictions match training data |

The mechanism behind the speedup is independently corroborated by a 2025 study from the National Institute of Building Sciences (NIBS). NIBS found that manual code-checking consumes a large share of total drafting time, and that automating it with UpCodes AI reduces that share to a small share. This reduction in the compliance burden explains the iteration speedup without requiring any assumption about the AI's generative capabilities. The AI is not drawing faster; it is eliminating the manual verification loop that previously forced architects to stop, check, and rework. A 2026 case study from Skidmore, Owings & Merrill (SOM) on a high-rise Chicago tower reported a reduction in design-iteration time, closely matching the benchmark. But SOM noted the savings were concentrated in the first three design iterations, not the final two. This is a critical edge case: the AI's value diminishes as the design converges, because late-stage changes involve coordination across disciplines that no drafting tool can automate.

![misty dawn over modern glass and steel atelier stretching into](https://static.mm-ais.com/article-images-ai/ai-drafting-38-speedup-real-but-it-s-wor-ai-1479c04f.jpg)
misty dawn over modern glass and steel atelier stretching into

## Evidence Check

The iteration-speed gap between Forma and TestFit is not marginal—it is the primary driver of the difference in speedup. Forma's generative adversarial network produces 50 massing options per minute, while TestFit produces fewer, and manual drafting yields one option per hour. That is not a trivial difference in tool efficiency; it is a difference in the shape of the design space you can explore before a client meeting. The GAN's parallel generation lets you test code-compliant permutations that a human drafter would never reach in the same wall-clock time.

Manual drafting retains one decisive advantage: flexibility on non-standard projects. In the AIAF benchmark, adaptive reuse projects saw only a limited speedup with Forma, and 3 of 5 such projects required manual rework of the AI's output, effectively negating the time savings. The GAN is trained on repetitive floor-plate geometries; when you hand it an irregular existing structure with load-bearing walls where you need openings, the generated options fail in ways that are faster to fix from scratch than to correct. This is the edge case that the marketing materials omit.

Decision tree: (1) Most projects are new-construction with repetitive floor plates? → Forma Professional. (2) Jurisdictions with complex local amendments? → TestFit Studio. (3) Fewer than 2 concurrent projects? → Manual CAD. (4) Adaptive reuse or irregular existing structures? → Manual CAD, regardless of project count. (5) Everything else, with 5+ concurrent projects? → Forma Professional, but only if you integrate an automated compliance checker into the loop—otherwise the speed gain is lost to manual rework.

| Evidence Source | Key Finding | Implication for Adoption |
| --- | --- | --- |
| AIAF benchmark (Jan 2026) | Median iteration reduction, p

Canonical: https://archparse.com/blog/ai-drafting-38-speedup-real-but-its-workflow-not-model.php
Markdown: https://archparse.com/blog/ai-drafting-38-speedup-real-but-its-workflow-not-model.php/index.md