First, what “AI” actually means right now
The term “AI” is doing a lot of work. When people say it today, they almost always mean one of three things.
Large language models (LLMs)
ChatGPT, Claude, Gemini, Llama. These are statistical systems trained on enormous amounts of text. They do not “know” anything in the way a person does. they predict the most likely next token (word fragment) given everything that came before. That sounds limiting, and in some ways it is, but at sufficient scale this prediction starts to look a lot like reasoning. It is also why they hallucinate: a confident wrong answer is just as statistically plausible to the model as a correct one.
Machine learning (ML)
The older, broader category. Pattern recognition trained on data. This is what runs your camera’s face detection, your bank’s fraud system, Spotify’s recommendations, and the noise reduction on your phone calls. ML has been quietly running the world for fifteen years.
Generative AI
A subset of the above that produces new content. text, images, audio, video, code. Midjourney, DALL-E, Sora, ElevenLabs, Suno. This is the part that feels most like science fiction and also the part with the messiest ethical and legal questions.
Knowing which kind of AI is doing what matters, because their failure modes are different. An LLM hallucinating a citation is a different problem than a facial recognition model misidentifying someone, and both are different from a generative image tool reproducing a copyrighted style. If you remember nothing else from this page, remember that the word “AI” on a product page is doing more marketing work than technical work, and your first question should always be: which kind, and trained on what.
AI in control systems
This is the area we work in directly, so we will go deeper than the others.
What was already there
Home and commercial automation has used basic intelligence for decades. Crestron, Lutron, and Control4 systems can run conditional logic. if motion is detected after sunset and the security system is in “home” mode, fade the hallway lights to 20%. That is automation, not AI. It is deterministic. It does exactly what you told it to do.
The limit of that approach is that you have to anticipate every scenario in advance, and any change in the household’s routine requires a programmer to update the logic.
What AI is changing
Three shifts are happening in parallel.
1. Natural-language control. The clearest example in our world is voice control that interprets intent instead of executing pre-programmed commands. “It is too bright” is not a rule we have to write. it is an intent the system has to translate into the right action based on which room you are in, what time it is, and what is currently happening. Apple, Google, and Amazon are all racing toward this with their respective assistants. In the integrator world, a number of platforms promise it. The honest reality from years of installing them is that most depend on a cloud connection, push firmware updates outside the integrator’s control, and degrade in reliability over time. We have written about that pattern at length in why we stopped recommending Josh.ai and in voice control done right in luxury homes. Today we run voice through Crestron’s native voice integration because the processor stays under the integrator’s control and the update cadence is something we manage, not something a vendor pushes to a client’s home on a Friday night.
2. Learned behavior instead of programmed behavior. Newer systems observe what you actually do. which scenes you trigger, what time, in what order. and start suggesting or executing them automatically. This is genuinely useful, but it is also where most systems fail in practice, because households are messier than the training assumptions. Guests, kids, schedule changes, and seasonal shifts all confuse pattern-matching systems. The honest current-generation is that learned automation works well for stable, predictable behaviors (wake-up routines, bedtime) and poorly for everything else.
3. Predictive infrastructure. Platforms like Domotz and OvrC use ML to watch network and device behavior across thousands of installations. They can flag a failing access point, a misbehaving streamer, or an unusual spike in traffic before a human notices. This is the least glamorous use of AI in our industry and probably the most valuable.
Where it still falls short
- Edge cases break it. Voice control with an LLM works great until someone has an accent the model was not well-trained on, or there is background noise, or two people speak at once.
- It is only as good as the underlying system. AI on top of bad programming is just bad programming with extra steps. The fundamentals. clean device naming, properly structured zones, reliable network. matter more than the intelligence layer.
- Privacy trade-offs are real. Cloud-based voice processing means your audio is leaving the building. Some platforms handle this better than others. Anyone selling you “AI in the home” without a clear answer about where the processing happens is glossing over something important.
FROM THE FIELD
The right question to ask any vendor selling AI in a luxury home is not “what can it do” but “where does the audio go, where does the video go, where does the data go, who owns the model, and what happens to the system if your company stops existing.” If the salesperson cannot answer all five in plain English, it is not a system we deploy.
AI in everyday life
The interesting question is not “what can AI do”. it is “what has it quietly become part of, and what does that change?”
Search and information
Google search results have been AI-influenced for years (RankBrain rolled out in 2015), but the shift to AI Overviews and tools like Perplexity changes the experience fundamentally. Instead of ten links, you get a synthesized answer with citations. This is faster and often better, but it has a serious side effect: it kills traffic to the sites the answer was synthesized from. Publishers are already feeling it. The long-term consequence. fewer original sources because nobody can fund the work. has not fully played out yet.
For an individual, the practical advice is: use AI search for fast answers, but for anything important, click through to the sources. Trust the citation, not the summary. A Columbia Journalism Review audit found that 37% of Perplexity answers contained errors despite the citation-first interface. The citations are evidence the model looked at sources. They are not evidence the model summarized them correctly. We say this as daily users of Perplexity, because it is true.
Writing and communication
LLMs are genuinely good at three writing tasks: drafting from scratch when you do not know where to start, editing for clarity, and translating tone. They are bad at: anything requiring lived experience, anything requiring genuine voice without heavy guidance, and anything where the facts have to be right without verification.
The skill people are slowly developing is knowing when to use it. A condolence note written by ChatGPT reads like a condolence note written by ChatGPT. A complex technical email drafted by ChatGPT and edited by a human is often better than what the human would have written alone. The difference between those two outcomes is entirely about who is in the driver’s seat.
Photography, audio, and media
This is where AI has changed the most for ordinary people without them noticing. Every modern smartphone runs ML on every photo. multiple exposures fused, noise removed, faces sharpened, sometimes entire background elements replaced. The “photo” you took is increasingly a composite the device decided would look best.
The same is true in audio. AirPods Pro use ML for noise cancellation and conversation isolation. Modern hearing aids use it to separate voices in noisy rooms. Video calling platforms remove background noise and even synthesize eye contact.
This raises a quiet question that does not have a clean answer: when does enhancement become fabrication? Most people do not care, and probably should not. But it is worth knowing that “the camera does not lie” stopped being true around 2020.
Health
Apple Watch, Oura, Whoop, and Fitbit all use ML to interpret raw sensor data. heart rate, HRV, skin temperature, motion. into things like “sleep stages,” “recovery scores,” and “stress.” It is important to understand that these scores are interpretations, not measurements. The watch is not measuring your sleep stages; it is guessing them based on patterns. Often the guesses are useful. Sometimes they are wrong in ways that can make a person anxious about a problem they do not have.
The genuine medical applications. ML reading retinal scans for diabetic retinopathy, flagging atrial fibrillation, screening pathology slides. are far more rigorous and increasingly trusted by clinicians. The wearable on your wrist and the imaging tool at your eye doctor are both called “AI.” They are not the same kind of evidence.
Driving
Adaptive cruise, lane keeping, automatic emergency braking, and parking assistance are all forms of AI. They have measurably reduced certain kinds of accidents. They have also introduced new failure modes (drivers becoming inattentive, edge cases the system was not trained on).
Two big bets are running in parallel on what comes after driver-assist. Waymo reports 170 million rider-only miles through December 2025, with serious-injury crashes 92% lower than the human benchmark in its operating areas. The trade-off is that Waymo only operates inside tightly geofenced cities and conditions. Tesla’s own safety report claims FSD users travel roughly 5 million miles between major collisions versus a national average around 700,000 miles. Tesla’s system runs anywhere in the country, on supervised drivers, which means it sees far more edge cases. and NHTSA opened a probe into 2.88 million Tesla vehicles in October 2025 after reports of FSD running red lights and traveling against traffic. Both companies are publishing safer-than-human numbers on the metrics they choose to publish. The honest reading is that geofenced robotaxi and supervised consumer FSD are two different products doing two different jobs, and the comparison is messier than either marketing department would like.
AI for your business
For someone running a small or mid-sized business, AI is genuinely a use tool. But the framing that matters is this: AI is good at reducing the cost of producing first drafts of almost anything. That is the real shift. The cost of starting a piece of work dropped close to zero. The cost of finishing it well did not.
Where it pays off
Anything with a template. Proposals, scopes of work, follow-up emails, case studies, service agreements, FAQ pages. Feed the model your past examples and the new specifics, and you get a first draft in 30 seconds that would have taken an hour. You still have to read it carefully, but you are editing instead of staring at a blank page.
Summarization. Meeting transcripts, long email threads, customer complaints, research documents. LLMs are very good at this and getting better.
Research. Background on a prospect, a competitor, a vendor, a technology. Perplexity and ChatGPT with browsing are good at the first pass. Anything that will be cited in a client-facing document needs verification.
Content production. Blog posts, social captions, email campaigns. The honest version: AI content without a clear brand voice and human editing reads exactly like AI content, and Google is increasingly good at recognizing it. AI-assisted content with strong human direction is different and works well.
Internal tools. This is underrated. Tools like Zapier, Make, and n8n now have AI nodes that can read incoming emails, classify them by intent (quote request, service issue, vendor invoice, recruiter, junk), pull the relevant data out, and route them to the right person or system. Five years ago this was a multi-week engineering project. Today it is a one-evening Zapier flow with an LLM step in the middle. The use is real, but the failure mode is also real: if the classification is wrong and nobody catches it, a quote request can sit unread in a folder. The rule we use internally is that AI can sort and draft, but a human still owns the final outbound communication.
Where it does not pay off, or actively hurts
- Anything where being wrong has a cost. Legal language, tax language, medical claims, regulatory commitments, anything that ends up in a signed document. The Air Canada precedent is now law in Canada and persuasive in the US: a chatbot is treated as an agent of the company that deployed it, and the company owes what the chatbot promised. Do not put an LLM in front of customers without bounded scope, logging, and a human escalation path.
- Anything requiring genuine institutional knowledge. The thing AI replaces is sometimes the thing a junior person needed to do in order to learn. If your associate generates a memo with an LLM and your senior reviews it without context, three years of that pattern produces a team that can produce output but cannot defend it.
- Brand voice in creative work. Marketing copy, design assets, and brand voice that have been generated end-to-end by AI read like AI even when nobody can articulate why. Use AI as a polish step, not as the source of original creative work. Document the human contribution. We do this for our own marketing, including the page you are reading.
The practical framing for an owner
Sit down once with a clear head and write three lists. First, the work in your business that is mostly template. same shape every time, just different specifics. Second, the work that needs your actual judgment and voice. Third, the work that is regulated, contractual, or otherwise expensive to get wrong. AI belongs in the first list immediately, in the second list as a draft tool only, and in the third list only with documented guardrails and human review. That is the entire strategy. Everything else is implementation detail.
The honest summary
AI in 2026 is a real shift, but not the one most of the marketing suggests. It is not a thinking partner with judgment. It is not a replacement for skilled people. It is a substantial drop in the cost of producing a first draft of almost any kind of work, and an even more substantial drop in the cost of pattern-recognition across data sets that used to require an analyst. Used well, it gives a small team the use of a much larger one. Used poorly, it produces a flood of plausible-sounding output that nobody can defend and that exposes the deployer to liability they did not anticipate.
The deciding variable is who is in the driver’s seat. We deploy AI in homes, hotels, commercial spaces, and government facilities across NJ, NY, and CT. We default to local processing when the use case allows it, we document what data leaves the property, and we keep a human in the loop on every safety-relevant or financially-consequential action. None of that is anti-AI. It is what taking AI seriously looks like.
If you want to go deeper on the specific risks, read the Reality Check. If you want to know which model we use for which task, read the Model Field Guide. If you want to run AI locally on hardware you own, read Build It Yourself. If you want us to design and deploy it for you, the number is at the bottom of this page.