# The Engineering Test: Is Your Building's AI Engineered, or Just Vibe-Coded?
A buyer's discipline for the age of agentic PropTech.
Every week another vendor demos an "AI agent" for your building — one that reads your BMS, drafts your work orders, forecasts your energy, answers your tenants. The demo is smooth and the slide deck is confident. But the question that actually protects your capital is not what did it do on stage. It is how was it built?
The lens worth borrowing from software engineering
At AI Engineer Europe 2026, the engineer Matt Pocock made an argument that should land hard for anyone buying building intelligence: now that AI models are good enough to write complex software, the oldest software-engineering disciplines matter more, not less. An AI agent, in his framing, is only as good as the engineering constraints, workflows, and quality standards you give it. Feed it vague requirements, skip the tests, ignore the architecture, and treat its output as disposable — and it can amplify those bad habits at machine speed.
The industry already has a name for the opposite of discipline: vibe coding — accepting whatever the model produces because it sounds confident. Vibe-coded software is the hidden liability inside a glossy demo. Vibe-coded building AI is a hidden liability inside your operations.
And the asymmetry is the whole point. A spreadsheet that is wrong fails quietly, on one tab. An autonomous agent that is wrong fails at machine speed, across your portfolio — a bad setpoint propagated to forty floors, a fabricated maintenance forecast that reshapes a capital plan, a tenant message that should never have sent. The downside is not symmetric with the upside, which is exactly why evaluation cannot stop at the demo.
Five questions that separate engineered from vibe-coded
Here is the test we apply at AI Smart Buildings. None of these questions is about features. All of them are about how the thing was built.
1. Does it show its work, or only its answer? Engineered systems carry a verification loop — they check their own output against data before they present it (the building-AI equivalent of tests and type checks in software). Ask the vendor: when your agent flags this chiller as a near-term failure risk, what did it verify that against, and can I see the chain? If the answer is merely "the model is very capable," you are looking at vibe coding.
2. Is it narrow and deep, or broad and shallow? The best-engineered tools are what the computer scientist John Ousterhout calls deep modules — a small, simple interface over rich, genuinely hard internals. A single agent promising energy and leasing and tenant experience and maintenance is usually the opposite: a shallow wrapper with a wide surface and little real depth. Prefer the tool that does one consequential thing exceptionally and exposes a clean seam over the one that demos ten things shallowly.
3. Can it say "I don't know"? Confidence is not competence. An engineered agent surfaces its own uncertainty and flags low-confidence outputs for a human. A vibe-coded one is uniformly, fluently certain — which is far more dangerous in a live building than an honest "I'm not sure, escalate this."
4. Where is the human gate? Ask precisely which actions the agent takes autonomously and which it routes for approval. An engineered system has an explicit, defensible line: consequential or irreversible actions — a setpoint change, a tenant communication, a capital recommendation — sit behind a review gate. "It just handles everything end-to-end" is not a feature to admire; it is an unmanaged risk to price in.
**5. Does the vendor talk about how it's built, or only about what it outputs?** This is the tell. Engineering-grade vendors tend to discuss their constraints, their measurement-and-verification discipline, their guardrails, and their failure modes — often unprompted. Vibe-coding vendors talk only about outputs and demos. Ask the simplest hard question — "what does your system do when it's wrong?" — and listen for whether there is a real answer or a change of subject.
The counter-intuitive part
Here is what surprises buyers: the smarter the underlying model gets, the more this discipline matters, not less. A more capable model fails less obviously and more confidently — its mistakes are better disguised, which makes them harder to catch and more expensive when missed. The fundamentals — narrow scope, verification, an honest "I don't know," a human gate — are not training wheels you remove as the AI improves. They are the load-bearing structure that makes a capable model safe to deploy in a real asset.
None of this is new. It is the oldest engineering discipline, pointed at the newest tools. Our job is to keep pointing it — to judge building-AI by how it is built, not by how it demos. The vendors who can answer these five questions are the ones worth your pilot. The vendor who cannot answer question five is the one to walk away from.
重點（繁體中文）
每週都有新廠商向你展示「建築 AI 代理」——讀你的 BMS、起草工單、預測能耗、回覆租戶。展示很流暢，但真正保護你資本的問題不是「它在台上做了什麼」，而是「它是怎麼被打造的」。
工程師 Matt Pocock 在 AI Engineer Europe 2026 的論點值得每位建築智慧買家借鏡：正因為 AI 已聰明到能寫複雜軟體，最老派的軟體工程紀律反而更重要。AI 代理的好壞，取決於你給它的工程約束與品質標準;餵它模糊需求、跳過驗證、忽略架構,它就可能以機器的速度放大壞習慣。業界稱紀律的反面為 vibe coding——因為它聽起來很自信就照單全收。一份算錯的試算表只在一個分頁安靜出錯;一個出錯的自主代理,卻會以機器速度、跨整個資產組合出錯。
分辨「工程化」與「vibe-coded」建築 AI 的五個盡職調查問題:(1) 它展示推理依據還是只給答案?(2) 它是窄而深(Ousterhout 的「深模組」:小介面、真本事),還是廣而淺的萬能 demo?(3) 它會說「我不確定」嗎?(自信≠能力) (4) 人為審核關卡在哪?哪些動作自主、哪些需核准? (5) 廠商談的是怎麼建的(約束、M&V、護欄、失效模式),還是只談輸出?
最反直覺的一點:模型越聰明,這套紀律越重要,而非越不重要——更強的模型出錯更隱蔽、更自信。基本功不是隨 AI 進步就拆掉的輔助輪,而是讓強模型能安全部署在真實建築裡的承重結構。能回答這五個問題的廠商,值得你的 pilot;答不出第五題的,該走開。
Editorial note: AI Smart Buildings is an AI-assisted intelligence desk. This piece synthesizes publicly presented work by Matt Pocock ("Workflow for AI Coding," AI Engineer Europe 2026) and John Ousterhout ("A Philosophy of Software Design"). It is general commentary for evaluating building-AI procurement, not legal, financial, or engineering advice, and names no specific vendor. Any forward-looking statements about how tools or vendors may behave are illustrative only and are not guarantees of any particular outcome; results may vary by system and deployment.
Research compiled by the AISB agent fleet from primary sources; every claim verified against the public record. Cost figures are labeled industry estimates. Full source list available on request — hello@ai-smart-buildings.com.