The Restrepo AI Field Guide

AI in our work · practical, not hype.

AI is the most over-promised feature in our industry. It is also one of the most useful when it is set up right. This is the field guide we wish every client had before they signed with anyone, including us. Models, costs, local versus cloud, open source, control systems, and the privacy stack that keeps a luxury home or a regulated facility clean. Written from twenty years in the field.

Best AI right now · by job, not by hype.

There is no single best. There are tiers, each best at different jobs.

General-purpose chat & reasoning

• GPT-5 (OpenAI). best all-around. Strong reasoning, tool use, voice. Default for most people.
• Claude 4.7 Opus / Sonnet (Anthropic). best for long documents, code review, careful writing, proposals.
• Gemini 3 Pro (Google). best inside Google Workspace, best video understanding, biggest context window.
• Grok 4 (xAI). best real-time web and X knowledge.
• Llama 4 / DeepSeek V3 (open weight). best when you need it on-prem or air-gapped. The pick for government and mission-critical facility rooms.

Voice & assistants in the home

• Apple Intelligence on HomePod / iOS 19. best privacy posture (most processing on-device).
• Josh.ai. best dedicated luxury home voice. Plays clean with Crestron.
• Alexa+ (Amazon). best ecosystem breadth, worst privacy posture.
• Google Assistant w/ Gemini. best for Nest and Google ecosystems.

Image & video generation

• Sora 2 / GPT-5 image. best photoreal video.
• Midjourney v7. best art direction and luxury moodboards.
• Google Veo 3. best long-form video continuity.
• Runway Gen-4. best for editors who want fine control.

Our honest take: most clients use GPT-5 for daily life, Claude for serious writing, and Gemini inside Workspace. That is the stack.

The cost of tokens · where AI bills hide.

A token is roughly 3/4 of a word. Vendors price per million tokens. Two prices for every model: input (what you send) and output (what it writes back). Output is usually 3 to 5x more expensive than input.

What a real conversation actually costs

• Quick chat, 10 turns on GPT-5: ~$0.05
• Read a 50-page PDF and summarize: ~$0.30
• Write a 1,500 word blog on Claude Sonnet: ~$0.10
• One Sora 2 clip, 10 seconds, 1080p: ~$0.50 to $2
• One Midjourney image: ~$0.04
• AI voice assistant in a guest room, 50 commands a day: ~$0.20 a day, ~$72 a year per room

Where the hidden costs hide

• Long context kills you. Sending a million-token codebase 100 times a day is real money.
• Output dwarfs input. A short prompt asking for a long article costs 5x what you expected.
• Reasoning modes silently 10x the bill. Thinking tokens are billed even though you never see them.
• Per-seat AI in business apps. Microsoft 365 Copilot is $30/user/month, Workspace AI is $20. Fifty people = $12K to $18K a year on top of existing licenses.
• Voice and video. Real-time voice runs ~$0.06 a minute. Video generation $0.50 to $5 a clip. Cheap until you scale.
• AI camera fees. Cloud VMS like Verkada and Eagle Eye charge $20 to $40 a camera a month. A 30-camera property is $7K to $14K a year.

How to keep AI cost sane

• Tier the work. Flash and mini for 80%, premium for the 20% that needs it.
• Cache prompts. Every major vendor supports it now (50 to 90% off repeated context).
• Local-first where you can. Apple Intelligence and Llama 4 on a Mac mini cost zero per token.
• Set hard caps on every API key. Every vendor supports it. Most clients get burned because they did not.
• For personal use, flat-rate seats ($20/month tiers) cover most workflows.
• For homes, plan ~$200 to $600 a year in AI subscriptions. That is the realistic recurring fee surface today.
• For commercial, budget AI as OpEx, not a free feature. A real corporate AI bill at a 200 person company is $50K to $200K a year all-in.

Datacenter vs local AI · two cost curves, one decision.

The honest split: 80% local, 20% cloud. Daily voice, transcription, camera analytics, summarization, smart-home brain. local. Bleeding-edge research and the occasional GPT-5 reasoning task. cloud. With proper segmentation, the user never knows which one answered.

Open source AI · the Linux moment for models.

Open weight vs true open source

• Open weight (Llama 4, DeepSeek V3, Mistral, Qwen, Gemma): trained model is downloadable and runnable. Training data is usually NOT public. You can run, fine-tune, and deploy it.
• True open source (OLMo, Pythia, Falcon, BLOOM): weights AND training data AND code are all public. Auditable end-to-end. Smaller and usually less capable, but you know everything about them.

The leaders right now

Why open source matters in our work

• Privacy by default. Run locally, nothing leaves the building.
• No vendor lock-in. Your client’s smart-home AI brain is portable.
• No surprise pricing changes.
• Air-gap compliant for government, military, hospital, financial services.
• Customizable. Fine-tune Llama 4 on a family’s preferences, a hotel’s brand voice, or a boardroom’s standards.
• Auditable. When a family asks "what did you put in our home", you can answer with specifics.

Watch the license

Llama 4 has a 700M-user clause that triggers a paid license at scale. DeepSeek and Qwen are made in China, relevant for some federal work. Mistral and Falcon are permissive Apache 2.0. Read the license. Have counsel sign off for regulated deployments.

Security and privacy with AI · the real talk.

The seven privacy postures every AI device has

1. What it captures. audio, video, occupancy, behavior, calendars, contacts.
2. Where it processes. on-device, on a local hub, or in the vendor cloud.
3. What it transmits. full audio, transcripts, just metadata, or nothing.
4. Where it stores. for how long, in what jurisdiction, who can subpoena it.
5. Who it shares with. third-party analytics, ad networks, training pipelines, law enforcement.
6. Who can access it. vendor employees, contractors, partner cloud platforms.
7. What happens when you cancel. does data delete, or sit in a backup forever.

If your integrator cannot answer all seven for every AI device they put in your home or building, do not let them put it in your home or building.

The eight real attack surfaces

1. Always-on microphones. Echo, Google Home, Josh.ai, smart TV mics, conferencing rooms. Each one is a recording posture.
2. AI cameras. Cloud-based VMS had real breaches in 2023 to 2025. Local-first is the answer.
3. Voice cloning. Three seconds of your voice from a podcast or a Zoom is enough to clone you. Establish a family or company safe word.
4. Prompt injection. Hidden text that tricks an AI agent into doing something the user never asked.
5. Model exfiltration. Cloud AI can be jailbroken to leak training data.
6. API key theft. Leaked keys get used to mine crypto or run scams on your bill. Cap every key.
7. Default-credential AV gear. Devices left at admin/admin are scanned constantly.
8. Network segmentation failure. AI cameras + IoT + control + work laptops on one flat VLAN is a free pivot.

The Restrepo privacy stack

1. Local-first by default. Llama 4 / DeepSeek on a Mac Studio or NVIDIA box in the rack. Cloud is opt-in.
2. Network segmentation. AI cameras, voice assistants, control processors, work devices, guest Wi-Fi. separate VLANs, separate policies, separate firewall posture. Ubiquiti UISP / Pro Partner work.
3. Hardened credentials. Every device gets a unique strong credential, stored in our team password manager, rotated annually.
4. Patched firmware. Every device on a documented update cadence.
5. Privacy posture document. Every client gets a written summary of every AI device, what it captures, where the data goes, and how to revoke it.
6. Voice-policy review. Where do you NOT want a microphone. We honor that. We document it.
7. Recording retention. Cameras and voice get explicit retention windows, not "forever" by default.
8. Family / staff safe-word policy. For high-net-worth and executive clients, a verbal safe word for any AI-related transaction request.
9. Quarterly security review. Network, credentials, firmware, and privacy posture reviewed every 90 days.
10. Documented exit plan. When you leave us or change platforms, we hand you the data, wipe what is wipeable, and document what is not.

The compliance lens

• GDPR / CCPA / CPRA. Cloud AI usually fails this without a DPA.
• HIPAA. No cloud AI without a BAA. Local-only is the safe answer.
• CMMC, FedRAMP, FISMA. Cloud AI is mostly a non-starter. On-prem, approved-product list, sustainment plan.
• NDAA Section 889 / TAA. Affects which AI cameras and AI gear are even buyable for federal work.
• NJ Consumer Privacy Act (active 2025). Your home state. New disclosure requirements.
• PCI-DSS. Hotels and restaurants. AI that touches the POS network is a PCI scope expansion.

Ten questions to ask any integrator before signing

1. Where does the audio from this device go?
2. What VLAN will the AI cameras live on?
3. Are you giving me a written privacy posture document?
4. Can the system run if the internet drops?
5. Who has access to the cloud account behind this?
6. What happens to my data if I cancel?
7. Is the firmware patched and on a schedule?
8. Are there any default passwords still in this system?
9. Have you read the GDPR / CCPA / HIPAA implications for my use case?
10. What is your incident-response plan if an AI device is compromised?

If they fumble three of those, walk away.