# Rachel Kimura, Principal at Kimura Westin Architects — read of Architect AI, May 11 2026

> 17 years licensed in California, co-run an 8-person firm in San Diego that does mostly civic and mixed-use. I write probably 40% of our specs myself because I don't trust anyone else to do it yet.

## How I got here

My business partner Colin forwarded me this link in Slack at 11pm with the message "thoughts?" No context. He does this. I opened it this morning on my phone while waiting for my daughter's dance class to let out. I was not in a buying mindset. I was in a "convince me this isn't vaporware" mindset, which is the same mindset I have been in since I tried three different AI writing tools in 2024 and deleted all of them within two weeks.

## What I clicked first

The hero line: "Architects, stop translating designs into specs." That one landed. That is the exact friction point. Every time we issue a set and a design change comes through, I end up either rewriting the spec section myself or explaining to a project architect what changed and watching them write it wrong. So I kept reading.

The secondary line "Every design change requires someone to rewrite specs. Architect AI lets your design changes flow directly into specifications" is where I hit pause. That is a more specific and interesting claim than "write faster." If that is literally true -- if it can ingest a changed drawing or a redline note and update the affected spec sections -- that is actually novel. I could not tell from the rest of the page whether they actually do this or whether that was just a metaphor for "describe the change and we will draft the words."

## Where I paused

The Tom Eriksen quote: "It sounds like someone who actually understands what a geotechnical report is and what belongs in a Division 03 spec section."

I stopped here for maybe two minutes. Division 03 is Concrete. If this thing actually knows Division 03 from Division 07 from Division 09 and treats them differently -- knows what belongs in a section 03 30 00 versus a 09 21 16 -- then it is not a generic AI wrapper with a nice UI. That would mean it was built by someone who has used MasterFormat. That matters. Because the failure mode of every AI writing tool I have tried in AEC is: it produces confident, plausible-sounding garbage that a GC would immediately recognize as written by someone who has never been in a building under construction.

I do not know yet if Eriksen is a real person or a persona built into the marketing page. That uncertainty is doing a lot of work right now.

## What I distrusted

Four things.

One: the statistics. "28% of principal time spent on documentation." "4.5 hrs average time to draft one complete specification section." "11 days typical lag from project kickoff to first client proposal." These are presented as if they are research findings, but there is no citation anywhere. No "according to AIA's 2024 Firm Survey" or "based on our user research across 60 firms." These numbers may be completely accurate and sourced from real research. They may be made up. I cannot tell, and that matters because they are the load-bearing evidence for the entire value proposition.

Two: "The client never knows AI touched it first." I understand why they wrote this. But it made me uncomfortable. We have clients who ask us directly if we use AI on their projects. Some of them care. This line felt like it was winking at me to be sneaky, and I am not sure I want that as my vendor's default posture.

Three: the testimonial names. Marcus Delacroix, AIA. Priya Nair, AIA LEED AP. Tom Eriksen. These could be real. Delacroix Architecture Studio, Portland is searchable in about 10 seconds. I have not done it yet. But the names and the quotes have the cadence of testimonials written by the same person who wrote the rest of the page. I have been fooled before.

Four: "Most firms report the output requires only minor editing to match their standard voice." After two to three documents. That qualifier "most" is doing a lot of work and "minor editing" is completely subjective. In my experience, "minor editing" means anything from fixing a comma to rewriting three paragraphs because the AI got the spec standard wrong.

## What would convince me

One real spec section, published. Not a screenshot. A downloadable PDF of an actual Division 07 or Division 09 section that a real firm submitted on a real project. Let me read it. Let me see if the performance requirements are cited to the right ASTM standard, if the submittals section is structured the way I expect, if the execution articles are in the right order. I can tell within two minutes whether someone who understands construction documents wrote the training data or whether someone who read about construction documents wrote it.

Alternatively: a 10-minute recorded demo where someone on their team inputs a real-world scenario -- say, a change from 2x6 exterior framing to structural insulated panels mid-SD -- and I watch what the tool does with the spec. Not a polished sales demo with a pre-loaded project. A live start-from-scratch that shows the follow-up questions the AI asks and the section it produces.

The quarterly code update claim also needs specificity. Which edition of the IBC are they on right now? California has a state amendment cycle on top of the IBC. Does it know Title 24? That is a deal-breaker for any CA-based firm if the answer is no.

## What I'd ask in an email reply

1. The hero says design changes "flow directly into specifications" -- does this mean the tool ingests drawing files or revision clouds and proposes spec edits, or does it mean I describe the change in plain language and it drafts the updated section? Those are very different products.

2. What is under the hood? I am not asking for trade secrets. I just want to know if this is a fine-tuned model, a RAG system over a code and spec corpus, or a prompt-wrapped GPT-4o. The answer changes how I think about output reliability and hallucination risk.

3. Can I see a live demo with a California project -- specifically, does it reference Title 24-2022 and CBC 2022 alongside IBC, or does it only know the model codes?

## Verdict: on-the-fence

If the Eriksen quote is real and the Division 03 claim holds up in an actual document, I would book a demo. The problem space is precisely right and the terminology signals someone built this who has been in a job captain's seat. But I need 15 minutes of honest demo before I give them a credit card number, and I need the hero claim about design changes flowing into specs to be an actual feature and not a metaphor.

---
*Memo by skeptic persona, generated 2026-05-11. Studio breaks own self-grading loop.*