AI in the Built Environment

Why we wrote this

The luxury AV industry is currently full of people selling AI who could not explain the difference between a transformer and a transistor. We are tired of it. So we wrote a small library on the subject, from the point of view of the people who will actually have to install it, integrate it with your Crestron processor, document it for your insurer, and answer the phone when something goes wrong at 11pm.

This pillar page is the front door. It is the shortest honest answer we can give to the question, "Should I put AI in my house, my hotel, my office, or my facility, and if so, how?" The longer answers live in the four sister pages linked below: the model field guide, the local-first build guide, the reality check, and our working review of Perplexity, which is the tool we use to run our own business.

We are Restrepo Innovations. We are an Elite Pro Crestron Dealer, HTA Luxury Certified, OSHA Certified, and a Ubiquiti UISP and Pro partner. We deploy Lutron, Ketra, and Colorbeam on a regular basis. We work in New Jersey, Connecticut, and the New York metro. We have done this for twenty years.

The thesis: AI as a guest in your house, not a tenant in someone else’s building

Most of what is sold as "smart home AI" today is a thin client to someone else’s data center. Your microphone is in your kitchen. The model that interprets it is in Virginia, owned by Amazon, Google, or Apple, governed by a terms-of-service document that can change unilaterally and has. Your house is the tenant. The provider is the landlord. You will be evicted on the day they sunset the product line, and you have no equity in the building.

We do not run our company that way and we do not recommend you run your home that way. The thesis of this pillar is simple: AI should live as a guest in your house, not as a tenant in someone else’s building. That means the compute, the models, and the data should default to local, on hardware you own, and reach out to the cloud only when there is a specific reason to do so. The cloud is a tool, not a host.

FIELD NOTE

"Local-first" does not mean "no cloud." It means the cloud is a service you call when you need it, not a substrate your house cannot operate without. There is a difference between a guest at the table and a landlord with the deed.

The four divisions where AI matters now

Residential

Concierge automations, scenes that learn your routine, pre-arrival climate and lighting based on calendar and traffic, voice control that does not pipe your kitchen audio to Seattle. Camera analytics that recognize your housekeeper without uploading her face. A handoff system that makes "I am leaving Manhattan" trigger the right cascade of HVAC, lighting, music, and security states without you having to touch a panel. None of this requires cloud. Most of it works better local.

Hospitality

Pre-arrival communication with guests, F&B forecasting, energy optimization on rooms that are sold but unoccupied, predictive maintenance on AV gear so the next guest does not get a black television. We deploy these for boutique hotels and private estates with rental programs. AI here saves real labor hours; it does not replace the front desk and we do not pretend it can.

Commercial

Conference rooms that recognize speakers, transcribe accurately, and post the right summary to the right Teams or Slack channel. Crestron Sightline AI, Microsoft Teams Premium, and Copilot for SIMPL+ are real products that real boards of directors notice. The bar for commercial AI is reliability, not novelty. A demo that fails in a board meeting costs more than the whole system.

Government and military

This is its own pillar, coming in Wave 2. The short version: air-gapped local LLM deployment, IL5/IL6 considerations, NDAA-compliant hardware, no Chinese chips. We do not push frontier cloud models into spaces that have to assume the network is hostile. If you are reading this from a classified-area scope, the rest of this page is for your residence, not your facility.

What AI is genuinely good at

It is good at converting unstructured input into structured output. Spoken request to lighting scene, photograph to scene description, twenty PDFs to a single page summary, calendar plus traffic plus weather into "leave at 3:42 to make your reservation." It is good at first drafts. It is good at search when paired with citations. It is very good at code if you let it focus.

It is good at boring things you would otherwise pay people to do: triaging a 200-message inbox, rewriting a vendor quote into your spec format, watching a camera feed for the FedEx truck and unlocking the gate. We use AI internally for all of this. We are credible on the topic because we run our shop on it.

What AI is bad at, period

It hallucinates. Every model. The frontier models in 2026 still confidently produce false statements at rates between 1% and 80% depending on benchmark and task, and the failure mode is not random noise but plausible-sounding fabrication. Vectara’s public leaderboard tracks this in detail.

It cannot do safety-critical real-time control without a deterministic system underneath it. We will not let an LLM decide when to unlock your front door, change your gas valve state, or override a fire panel. That is the job of a Crestron processor with documented logic, not a probabilistic model with a 4% miss rate.

It does not know what it does not know. The AA-Omniscience benchmark showed that GPT-5.5 produces a confident wrong answer 86% of the time when it does not know something. That is not an AI you let near a contract. We address this in the Reality Check.

RISK

Any integrator who tells you AI "doesn’t hallucinate anymore" or that "the new models are basically perfect" is either lying or has not read a benchmark in twelve months. Both are disqualifying.

The honest risks

Lawsuits in flight: OpenAI versus Musk goes to trial in Oakland in April 2026. The New York Times versus OpenAI continues to surface evidence of training-data regurgitation. Air Canada was held liable for what its chatbot promised a customer, and that precedent now applies to anyone deploying a customer-facing bot, including us. We tell clients about all of this before we deploy anything.

Data leaks: Samsung banned ChatGPT internally in 2023 after engineers pasted proprietary code into the free tier. JPMorgan and Amazon have similar restrictions. The free tier of every major chatbot may train on your inputs. Energy strain, vendor lock-in, deepfake fraud, and the sunset risk of "smart" subscriptions all live in the Reality Check.

How we deploy it

We use a router pattern. No single model wins every task, so we do not pretend one does. A typical Restrepo workflow sends search to Perplexity Sonar, drafting to Claude, polishing to GPT, deep research to Gemini, and image work to whichever model the client’s brand guidelines allow. The orchestration sits behind one interface so the client never has to think about which model is doing what. Full breakdown in the Model Field Guide.

The local-first option

For clients who want AI without a cloud dependency, we build local stacks. A Mac Studio M3 Ultra, a Framework Desktop, or an NVIDIA RTX 5090 box runs Llama 4 or 5, Whisper for voice, MiniCPM-V for vision, and Home Assistant or OpenClaw for orchestration. It costs between $1,500 and $15,000 depending on capability. The full hardware budget and architecture lives in Build It Yourself.

The Perplexity story

We use Perplexity Computer to run Restrepo Innovations. The blog you are reading was queued by it. The hero image was generated through it. The publishing pipeline runs on a cron it manages. We are not an affiliate of Perplexity. We pay full price like everyone else. We wrote a working review at our Perplexity page because if we are going to recommend a tool, we should be using it ourselves and willing to say where it fails.

The integrator’s perspective on hype

We have lived through every wave of home technology since the late 1990s. We installed the first generation of touch-panel control systems when they had to be programmed in a proprietary language only the dealer understood. We saw streaming kill physical media and then watched physical media come back as a luxury choice. We saw the smart speaker land in 2014 and the entire residential AV industry pivot to assume Alexa would be in every house. We saw cloud-tied locks ship to thousands of homes and then watched the manufacturers go bankrupt and brick the locks remotely.

The lesson from twenty years of installs is that hype cycles in this industry are reliably four to six years long, and the buyers who do well are the ones who picked architectures that survived the cycle. Crestron survived. Lutron survived. The deterministic local processor as a category survived. The cloud-tied gimmick of the moment generally did not. Our position on AI is informed by that history. AI is not a hype cycle; it is real. But the cloud-tied AI gimmick of the moment may very well be one. We pick architectures that work in the year you buy them and in the year you sell the house.

What we will not do

We are happy to be specific about this. We will not let an LLM autonomously control gas, water, alarms, or front doors. We will not deploy a customer-facing chatbot for a hospitality client without bounded scope, logging, and a human escalation path. We will not put a free-tier ChatGPT in any workflow that handles regulated client data. We will not promise a client that AI is going to replace a staff member; we will tell them which tasks AI handles well and which it does not. We are an Elite Pro Crestron Dealer, HTA Luxury Certified, OSHA Certified, and a Ubiquiti UISP and Pro Partner. We deploy Lutron and Ketra products on every project type we run.

A 60-second decision tree

Do you want AI in your home right now?

If you want to dim lights by voice without an Amazon account: yes, and we can do it locally with Whisper plus Home Assistant on a $1,500 box.
If you want a "concierge" that books your reservations: cautiously yes, with manual confirmation on every action that costs money.
If you want AI to control safety systems autonomously: no. We will not build that. No reputable integrator will.
If you want AI to draft your emails and brief you on your day: yes, and Perplexity Computer or a Claude/GPT router does it well.
If you want AI to recognize family members on cameras without sending video to the cloud: yes, Frigate plus a local model handles this.