top of page

What We Learned Building Triton, Our AI Drawing Review Tool

Oct 20
4 min read

Updated: Oct 25

AI adoption in architecture is moving fast, and most firms are still figuring out where to start. We decided the best way to learn was to build our own tool.

Triton Dashboard

We call it Triton, an AI reviewer built on Claude, Anthropic's AI model. We built it thoughtfully to ensure client information remains private while helping our team automate drawing reviews.


Building Triton took hundreds of hours, but it taught us a few things worth sharing.

1 | Start with a real bottleneck

We didn't adopt AI because everyone else was doing it. We had a specific problem: our small but mighty team was spending a lot of hours on small things instead of developing creative, innovative, and functional designs. Now Triton identifies sheet index mismatches, missing code items, and unincorporated consultant markups in real time.  We fed Triton checklists, references, and previous drawing sets to automate plan reviews instantly instead of doing them manually.


Before adopting any AI tool, identify the specific bottlenecks it needs to solve. A tool without a clear job to do is just another subscription.


2 | Weigh build versus buy

Larger firms already pay for AI platforms built by teams that specialize in creating enterprise software. We wanted to build our own platform, specifically to meet our needs. While we don't have a clean answer for when it stops making sense to maintain our own tool or pay for a platform, we do know that a tool built around our own drawings, checklists, and project scale does things an off-the-shelf platform built for bigger firms won't replicate as closely.


If your firm is considering the same build-versus-buy question, weigh your team's capacity to maintain a tool against what a vendor platform solves for your size and workflow.


3 | Watch your costs

Triton runs on a set monthly token allotment tied to our enterprise account. Part of our ongoing work is reducing how much of our server Triton needs to search to answer a single question, since searching an entire project archive to resolve one question spends tokens that a narrower search wouldn't need.


We've also found that it's easy to get lured in by "freemium" offers. Be wary: once your needs become more advanced, the price goes up...and sometimes astronomically! Budget for AI the way you'd budget for any other tool. Track usage costs and how they scale with how a tool searches or generates, not just how often individuals use it.


Examples of AI-supported renderings


4 | Test where the technology holds up

We use AI for early-stage renderings too, sending a rough model or sketch through a prompt to visualize a design direction before committing to a full render. It works well for an overall look, but targeted edits like changing a single object's color without altering the rest of the image remain difficult. For renderings where we need more precise control, we still rely on our own software.


It helps to know where AI is a genuine shortcut and where it still needs your team's usual tools to finish the job. We've discovered (the hard way) that using AI at the wrong time can create unnecessary back-and-forth that you could avoid by spending a few more minutes up front specifying the ask.


5 | Solve the policy problem before it solves you

Video meeting note-takers are the clearest example. A single meeting can have two or three different AI bots joining at once through individual accounts, each with a different premium tier waiting behind a paywall. We've standardized certain AI agents for emails, note-taking, proposal work, and other tasks based on their capabilities, cost, and legal framework.  Issuing clear guidance on which tools are approved for corporate use is critical to avoid the confusion of competing AI models.


Treat AI adoption as a policy decision, not just a technology one. We’ve learned it’s much easier to start with a corporate policy that identifies approved tools, versus backtracking or asking individuals to remove AI agents from their toolkits.


6 | Train your team to always verify

We think of AI outputs similarly to Wikipedia in its heyday in the early 2000s: a good starting point, drawn from sources of uneven reliability, and never a substitute for checking the work yourself. Triton addresses that by citing sources with live links back to the code and drawings it references, so a team member can check the claim instead of taking it on faith. A Triton drawing review example


We've noticed younger staff are more likely to take AI output at face value than experienced architects, so this is as much a training issue as a tooling one. We’re approaching AI through the lens of our core values: become leaders in setting AI policy and best practices for our internal team, use AI to enhance our pursuit of unrelenting service, and always check our AI work against the human component, using our relationships with one another as the governing factor. 

 

Whatever tool your firm adopts, build in a habit of verification from day one, and let your team’s judgment and relationships have the final say. If your firm is weighing whether to build its own AI tools or buy off the shelf, we’re happy to share what we’ve learned along the way.

Comments


@2026 Framework Architects

  • LinkedIn
  • FRAMEWORK ARCHITECTS
bottom of page