How to Make AI Agents Predictable for Your AEC Firm

Every firm on the call had already tried AI in some form. What they hadn’t figured out was how to make it behave the same way twice.

That was the real subject of our live Q&A on August 13, 2026 — a session open to anyone working in architecture, engineering, or construction, with attendees dialing in from Toronto, Ohio, and Abu Dhabi, among other places. No pitch, just questions about what it actually takes to move AI from a demo into something a firm can run on.

The conversation covered a lot of ground: the origin of the platform, how a QA/QC agent for Revit models actually works under the hood, the architecture that keeps an AI agent’s behavior predictable, and a live case study from an attendee’s own construction firm. Here’s what came out of it.

Where the platform came from

The founder’s background with AI goes back further than most people expect. It started with a graduate thesis on speech recognition in 2005. Two years later, in 2007, he was one of two people who built a technology that was later sold to Dell Computers for around $130 million — a deal he’s quick to point out didn’t make him a millionaire on the spot. “Deal making is an art you learn throughout your life,” as he put it on the call.

By 2017, the team had earned top AI company status on Upwork, and the companies they’d worked with had collectively raised somewhere between $65 and $70 million. About a year and a half ago, that experience turned toward architecture, engineering, and construction specifically — after outreach to a handful of firms that turned out to be, in the founder’s words, “pretty adamant about their pains and how AI could fix that.” That became the starting point for Vitru AI.

Out of everything a firm does, QA/QC turned out to be the sharpest entry point. It’s one of the most complex use cases in the design process, and manual QA doesn’t scale. An engineer submits a model, a QA team reviews it, sends comments back, and the cycle repeats — with every engineer developing their own particular style of mistakes that rarely gets written down anywhere a machine, or the next reviewer, could learn from.

The problem with checking every element one at a time

The first version of the QA/QC agent used AI to check every element in a model manually, live, one inference call at a time. It worked, but it turned out to be token-consuming and expensive — the kind of cost that grows in a straight line with the size of the model, which is exactly the wrong direction for a tool meant to save a firm money.

The fix was to change what the AI is actually doing at check time. Instead of asking a model to judge every element fresh, an agent looks at a rule once, figures out how to test it properly, and writes a script for it. From then on, that script runs the check — deterministic, fast, and cheap to repeat across hundreds of elements. The AI still does the hard part, converting a plain-language rule or a line from a building code into something testable. It just doesn’t have to re-derive that judgment every single time.

That shift also determines where AI belongs in a QA/QC process at all. Anything genuinely subjective — a design call, an aesthetic judgment — is where AI-assisted human review makes sense. Anything that’s actually pass or fail — does this match code, does this match the firm’s standard — is better served by a script than by a fresh opinion.

Once a model check is running, the output is a report — a high-level view of what passed and what failed, broken down rule by rule, with the option to jump straight into the flagged element inside the model and fix it.

The constitution: how you make an agent predictable

The single idea that did the most work on the call wasn’t a feature. It was a naming choice: every firm on the platform gets what the team calls a constitution — a foundational document, auto-injected into every agent that runs, describing the firm itself and exactly how its information is organized.

The constitution doesn’t just describe the firm in general terms. It maps the folder system down to the specifics — where standards live, where project files live, where a firm’s own golden-standard templates sit — so an agent isn’t guessing where to look for context. A firm can extend it any time: add a subfolder under templates with the firm’s preferred Revit file organization, for instance, and tell the constitution that the QA/QC agent should check every model against that standard going forward.

Projects get the same treatment one level down. Every individual project has its own constitution, specifying which jurisdiction it falls under and which rules apply to that specific project — auto-injected into any agent working inside it. Inside each project folder sits a QA/QC subfolder holding the reports, the rules, and any proposed additions to the rule set.

The principle behind all of it: if you want AI to behave predictably, you have to feed it context as predictably as the output you expect from it. A system prompt without that context is a job description written for “someone who is universally smart but has no clue how exactly to work within your company” — capable, but directionless. The constitution is what closes that gap.

The workspace itself is versioned, too, which means every update becomes the next revision in a running record. That’s what makes it possible to track how a project — and its rule set — has actually developed over time, instead of losing that history every time a model gets resubmitted.

The three layers underneath any working AI setup

Zoom out from any single agent and a firm’s AI infrastructure tends to split into three layers.

At the bottom sits the tools a firm already runs on — Revit, and whatever else fills out the stack. Above that is an AI and automation backbone: the layer where agents, automations, and schedules actually live and connect to those tools. Above both of those sits a meta-level library — the firm’s design language, standards, checklists, and compiled scripts, the same material that gets referenced inside a constitution.

That top layer is what makes the whole thing format-independent. A rule set built well doesn’t have to live inside Revit alone — it can travel between Revit and ArchiCAD, or Rhino and Revit, which is what makes cross-platform interoperability possible in the first place instead of rebuilding the same standards twice.

An agent itself needs four things to function inside that stack: triggers that wake it up (a new email with a model attached, a scheduled Monday check), a connection to a model provider, memory to retain what it’s learned, and access to the tools it needs to act — sending a message, updating a project management record, submitting a revised model.

What’s live, and what’s next

Today, the QA/QC agent runs read-only by default — it flags issues, it doesn’t touch the model. An edit mode is in active development: point the agent at a firm’s golden standard, and it finds and fixes deviations before a model ever reaches a human reviewer.

Beyond that, the roadmap includes ArchiCAD and IFC support, an AutoCAD plugin, and a project coordinator agent that maintains the project plan, tracks outstanding items, sends reminders, reads incoming email, and flags risk. Document ingestion for project submissions is also in progress. Further out, the team is working toward AI-assisted model creation itself — generating a Revit model from images or a specification document, not just checking one that already exists. As the founder put it, QA/QC might be the most complex use case in the industry today, “but probably the more complex one than that is the Revit model creation itself — and we are also moving in that direction.”

The mindset shift: above the loop, not in it

The most direct line from the call describes what actually changes once agents go autonomous:

“There is no human in the loop. There is a human above the loop, because the human has designed this process.”

That’s a real shift in how a person’s day looks, not just a slogan. The founder described running his own company this way — not writing code by hand, but steering a workflow-based development process where AI writes code against defined rules, other agents check that output for quality and architecture, and the whole loop runs with minimal human intervention. The job moves from producing the work to designing and supervising the system that produces it.

That shift is also where AI adoption gets uncomfortable, independent of any technology question. People build their sense of usefulness at work around how busy they are and how much they personally touch. Moving above the loop means becoming more of a director and less of a producer — and no amount of ROI math resolves that discomfort on its own. Firms that get through it are the ones where leadership treats AI-directed output as real work, not a shortcut.

A firm already mid-pivot

One attendee on the call — a design, installation, and construction firm with roughly 500 to 600 employees — offered a candid look at where most firms actually are.

“We are just getting started. We don’t have an AI policy in place yet. We don’t have any kind of a roadmap.”

They described an earlier evaluation of a closed-loop, on-premises AI system that didn’t have Revit integration. Security was the sticking point: “Everybody’s concerned about where this information is going. Is it going out into space, into the cloud?”

Their first real step has already paid off, though. The firm implemented Trunk Tools, integrated with Autodesk Construction Cloud, giving their teams the ability to query project documents directly — cutting the time it takes to find information on large projects. That capability has since trickled down to field teams, who now search specs, submittals, and contracts through a pre-built AI agent rather than digging through folders by hand. It’s an early pilot, not a transformation yet, but it’s a real one — and given the firm’s size, running that setup on their own private infrastructure makes sense rather than depending entirely on the public cloud.

That distinction matters more broadly, too. Enterprise accounts with major AI providers already guarantee contractually that your data isn’t used to train their models — a risk profile similar to the one most firms already accept using Microsoft or Google for email. Firms that need tighter control can run an open-source or local model entirely inside their own private cloud, with access limited to their own network.

How to actually prioritize what to automate first

1. Map the process as it really runs. Sit down with your leads and document a specific workflow the way it actually happens — not the org-chart version. This becomes the raw material both for automation and for the constitution an agent eventually needs.
2. Overlay real data to find where time and value leak. Turn that process map into a heat map using your own operational numbers — where work backs up, where delays happen, where value is created or lost. That heat map tells you where to point AI first, instead of guessing based on what’s trending.
3. Score every candidate use case against ROI and effort. Not every proof of concept is worth building first. An information-retrieval tool can be genuinely useful and still be hard to put a dollar figure on. A process automation you can measure — faster submissions, fewer review hours per model — is what lets you tell leadership exactly how much time was saved this week. Rank candidates by ROI against difficulty, and start with whatever sits in the highest-ROI, lowest-effort quadrant.

Quick tips for making AI agents predictable

• Write a constitution before you write a system prompt. Describe your firm and your folder structure once, in a document every agent references, instead of re-explaining context in every conversation.
• Compile the rule, don’t re-judge it every time. For anything pass/fail, have AI generate the check once, have a person review it, then run it as a script — not a fresh opinion on every model.
• Give agents a versioned workspace. If every update becomes the next revision, you get a real record of how a project and its rule set evolved — not just the current snapshot.
• Design the process before you automate it. Map the workflow with your leads first. The automation plan and the AI’s eventual instructions both come out of that map.
• Match your infrastructure to your data sensitivity. A firm running hundreds of employees with strict security requirements is a reasonable candidate for a private, on-premises setup — not every firm needs to default to the public cloud.

Book a demo → Join our next Q&A

FAQ

What is a “constitution” in an AI agent system?

It’s a foundational document auto-injected into every agent a firm runs, describing the firm itself and exactly how its information is organized — which folders hold which standards, templates, and rules. Individual projects can have their own constitution too, specifying jurisdiction and project-specific rules. It’s what lets an agent find the right context without being told from scratch every time.

Is our project data safe if we use a commercial AI provider?

Enterprise accounts with major AI providers contractually guarantee your data isn’t used to train their models. The realistic risk profile is similar to the one you already accept using Microsoft or Google for email and file storage. Firms with stricter requirements can run an open-source or local model on their own private cloud, keeping everything inside their own network.

What does “human above the loop” actually mean day to day?

It means the person’s role shifts from doing each step of a process to designing the process an AI agent executes on its own — setting the rules, defining what triggers action, and reviewing outcomes rather than approving every individual step. It’s a real change in how the workday feels, not just a description of the technology.

We haven’t touched AI yet. Where do we actually start?

Start by mapping a real workflow with your leads, not by shopping for tools. That map becomes a heat map once you overlay your own operational data, which tells you exactly which process to automate first — and it becomes the foundation for the constitution an AI agent eventually needs.

Will AI eventually create Revit models, not just check them?

That’s the direction the team is moving in. QA/QC is one of the most complex use cases in the industry today, but model creation from images or specification documents is considered even more complex — and it’s already on the roadmap.

The takeaway

The technology under a predictable AI agent isn’t a smarter model — it’s a constitution that tells it exactly where to look, and a rule set that got compiled into a script instead of re-judged from scratch every time. That’s what turns a chat window into infrastructure a firm can actually run on.

Want to see what that would look like on your own project structure? Book a demo and we’ll map it out with you — or join our next live Q&A.