The Three Dog Theory: Is Your Office a Dog Shelter?
Axoworks Commentary | August 2026
AI maturity in AEC is determined by the handler's skill, not just the model's quality.
Smarter doesn't mean better; smarter means the errors are now even harder to spot. This is where our original 3-dog theory still holds true.
Just over 6 months ago, AEC firms were prompting ChatGPT to summarize a zoning or building code. The chatbot was the know all, cure all.
Then the ground shifted. Openclaw (Clawbot/Moltbot) crawled its way to the front page, an unruly personal agent, mostly hijacked and exploited days of going viral. Frontier moved from prompting to autonomy for the tech industry exploring the edge.
US models evolved to capture the new bot market: Claude Opus 4.6 with agent teams, GPT-5.4 routes tasks across parallel tool calls. China's open-weight models like Deepseek dominated their Lobster craze.
These new models are not only regurgitating their reward-trained data, they are now performing tool calling: checking the web, scheduling your day, responding to emails, or cleaning out your hard drive (for better or worse). They read your files, call your tools, and run for hours without you watching.
Power in the wrong hands is a loaded gun. Even Meta's Director of Alignment at its Superintelligence Labs had to watch in horror as Openclaw stubbornly obliterated her entire Gmail inbox in real time. It's a brave new world.
As capability exploded, failure modes got worse. Hallucination rates on most complex reasoning still run at 5–20%, a 2024 Stanford study found legal hallucinations jumped to 69–88% for state-of-the-art (SOTA) models. Even with corporate RAG guardrails, 17–33% of the time. Agentic systems now hallucinate entire software packages. Exploits through ‘slopsquatting’, registering the fake packages an agent suggests so it installs malware. One USENIX study found 19.7% of packages AI recommended didn't exist at all.
1. The Puppy
Majority of firms start here. Sign up for the latest frontier model subscription, paste a permit application with little to no real directions, and get a beautiful, confident, perfectly grammatical response. We call this Puppy love. The puppy doesn't know what you need, so it makes wags and cuddles just to please you.
This is an unwary human handling a very smart dog, a frontier model. Often too smart for the human handler. The smarter it is, the more convincing the hallucination. The Puppy doesn't understand the zoning code, it just understands that you want a ’yes’ from you. So it gives a confident, beautifully written, completely hallucinated ‘yes’.
Using it for professional work is like hiring a toddler to do your taxes. It's cute, but a disaster waiting to happen.
2. The Confused German Shepherd
These are the firms who dutifully did the upgrade. They did their research and deployed frontier models. MCP servers connect their agent to Revit, their document management system, their schedule. They built a real badass agent.
Then they grab the leash and mumble: "Hey buddy, can you like… look at this project? Maybe make it more professional? Check some of the stuff?"
The dog tilts its head. It looks at you with pity. This dog is not stupid; it's confused by you. You have a Ferrari of intelligence, but you're driving it into a brick wall because you don't know how to give clear commands.
The agent runs for hours without guardrails and structure, confidently doing the wrong thing. It clicks through software and destroys a drawing set. It burns $2,000 in tokens on a $42 task. It writes its own verification and signs off on its own mistakes. Nobody built a second pair of eyes into the loop.
3. The Elite K-9 Unit
The firms that run Elite K-9 units deploy the same models and agents, except that they leave nothing to chance.
It's the same German Shepherd. But it's not tilting its head anymore. Why? Because it's been trained to move with precision.
- Model routing. SOTA thinking models for planning tasks, cheaper models for simple classification, local models for privacy. Pay for intelligence when it matters.
- Grounded workflows. Every agent runs against a verified, versioned knowledge base and spec. Every answer cites its source, if not, flags that it ‘cannot’.
- Verify, verify, verify. The Elite Unit assumes the dog lies. Every output passes a second or third model, a rule engine, a Red Team to flush out the BS, and a human who is accountable.
- Error logging. Every failure gets logged, categorized, and fed back into the playbook.
The Verdict
The gap between the top and bottom of the industry is widening faster than ever. The unsavvy now have more powerful tools than experts had 12 months ago. What is scarier … they are more dangerous to themselves than ever before.
The theory still holds: AI maturity is determined by handler skill, not model quality.
The models got better. The stakes got higher. And the dogs are already in your yard.
The only question left is: Are you running a shelter, or are you part of an Elite Unit?
The full 2026 manifesto, with the companion technical report on how AI lies, lives at axoworks.com/articles/three-dog-theory-v2.