White paper · 2026 edition

The hardware finally caught up.
Now AI runs your business,
your body, and a robot in every yard.

For twenty years computer scientists imagined this. The maths was ready. The chips weren't. In 2024 that changed — and the services industry, from HVAC to fire protection, is the first place it pays back. This is the playbook for owners who want to be early, not late.

Read the playbook Skip to robotics →
10× smarter every 18 months
$180B per AI lab, this year
3 yrs to data centres in space
predict_next tokenise attention(q,k,v) vision · NN plan → act tool_call()
The moment

Twenty years of brilliant ideas. One year of working silicon.

The transformer architecture that powers Claude, ChatGPT and Grok was sketched in academic papers a long time ago. The mathematics for self-attention, back-propagation, convolutional vision networks — none of it is new. What's new is the ability to actually run it at scale. GPU compute, cheap CMOS sensors, hyperscale data centres and reusable rockets converged in the last 18 months. Ideas that were trapped on whiteboards are now driving cars, reading invoices, and walking on two legs in factories.

2005
First deep nets — too slow to be useful
2012
AlexNet — GPUs unlock vision
2017
"Attention is all you need" — transformers
2022
ChatGPT — language goes mainstream
2026
Hardware-rich era — agents, vision, robots
~2029
Solar data centres in low-earth orbit
"We had the software twenty years ago. We didn't have computers fast enough. Now we do. So they're building data centres the size of cities — and they're already running out of land and power." — from the conversation that started this paper
First principles

An LLM isn't intelligent. It's a brilliant guesser, scaled to absurdity.

A Large Language Model takes a stream of tokens — short fragments of words — and predicts the next most likely token, billions of times in a row. Stack enough probability on enough text, throw enough silicon at it, and something that looks like reasoning emerges. That's the secret. It's also the limitation: every chat thread starts fresh. There is no memory between conversations unless you build one. That's where agentic systems come in.

Live · token-by-token prediction
Snap → a → supplier → bill → , → match → to → a → job ? customer truck vendor
"job" 0.72
"customer" 0.14
"vendor" 0.08
"truck" 0.04
1
It's prediction, not understanding. Treat it like a brilliant junior who has read everything but remembers nothing — give clear briefs, verify the output.
2
Context evaporates between threads. Every new chat is a fresh hire. If you want continuity, you need Projects, a knowledge base, or a true agent (see below).
3
It can lie confidently. "Hallucination" is real — verify dates, prices, code, regulations. It will tell you Monday is Sunday with full conviction.
4
Bigger model = better adviser, 10× the cost. Use the smartest model for thinking, the cheapest model for grunt work. Don't ask the apprentice for strategic advice.
The stack

Three layers of AI you can use this week. They each do a different job.

Don't get confused by the marketing. There are really only three places AI lives for a services-industry owner — Projects (a smart conversation), Builders (Cursor and friends, where you make things) and Agents (your own Jeeves, on a dedicated machine). Each layer is more powerful — and more work — than the last.

LAYER 01

Projects

A bundle of related chats with shared files. Your starting point. Inside Claude or Grok. No setup beyond "create project".
  • One project per topic — "Website refresh", "Quoting playbook", "My health"
  • Drop in PDFs, supplier docs, statements of work, photos
  • Context follows you between conversations inside the project
  • Best for: thinking, drafting, decision support
Setup: 5 min ≈ $40–200/mo
LAYER 02

Builders

Tools like Cursor that wrap Claude / Grok / GPT and let you build things — websites, dashboards, internal tools. You don't need to code; you need to think clearly.
  • Redesign your website by describing what you want
  • Generate spreadsheets and dashboards from raw data
  • Stand up internal apps in days, not quarters
  • Best for: prototypes, ops tools, content pipelines
Setup: 1–2 days + ~30% on tokens
LAYER 03

Agents (Open Claw / Hermes / Jeeves)

A dedicated machine running your own AI assistant — reads email, calls APIs, runs tools, sends messages. Iron Man's Jarvis, but yours. The end-state.
  • Local on a Mac mini / Mac Studio for security & privacy
  • Reads inbox, posts to blog, updates pipeline, sends SMS
  • Persistent memory — gets smarter the longer it lives with you
  • Best for: replacing the EA, the bookkeeper, the analyst
Setup: weeks Hardware + tokens
"OpenClaw is your own Jeeves. It reads your email, checks your metrics, runs the morning report. I don't have an assistant — I'm building one."
Pick the brain

Don't ask the B-grade coach for the A-grade decision.

The models are not interchangeable. Each one has a job. The price ramps roughly 10× per tier — but so does the quality of the thinking. For real business questions, pay for the best brain. For execution, dial it back.

Claude · execution-light

Haiku

The fast, cheap one. Good for repetitive tasks — bulk extraction, quick summarisation, tagging.
≈ $
Claude · everyday work

Sonnet

The reliable contractor. Great for building, drafting, day-to-day execution. The default in Cursor.
≈ $$
Use this
Claude · strategy & advice

Opus 4.7

The A-grade adviser. For executive thinking, planning, hard judgement calls. 10× the price, worth every cent on the questions that matter.
≈ $$$$
Truth-seeking
xAI · second opinion

Grok

Less filtered, less politically tuned. The cross-reference for health, science, economics — anywhere ideology contaminates the answer.
≈ $$$
Bias matters

Same question. Three models. Three different worlds.

Ask each LLM the same direct question and the answers diverge — not on facts, but on how willing the model is to commit. As AI gets smarter than us, that gap matters. You want a tool that tells you what the data says, not what it thinks you want to hear.

ChatGPT guarded
Q: "Is my conversation private?"

"It depends on a number of factors, including your settings, the policy in effect, and applicable law. You should review the privacy policy and consult a professional if uncertain…"

Claude cautious
Q: "Is my conversation private?"

"I want to give you accurate information here. Anthropic's policy says conversations may be used to improve our models unless you opt out in settings. Let me walk you through how to check…"

Grok direct
Q: "Is my conversation private?"

"Short version: no. xAI uses chats to improve Grok by default. Toggle it off in Settings → Data & Privacy. If it's truly sensitive, don't put it in any LLM — including me."

"If you always get one side of the story, you'll develop a one-sided view — even if you think you're a balanced person. That's true of media. It's now true of the AI you ask for advice."
Vision & robotics

The same maths that finishes your sentence is now driving your car.

Optimus didn't suddenly become possible because someone invented walking. It became possible because neural networks, trained on cheap cameras, finally learned to see the real world fast enough — on chips small enough — to fit on a robot's spine. The same is true for Tesla's full self-driving, the Roomba under your couch, and the robotic mower trimming a lawn next door. Vision + neural prediction + finally-affordable compute.

The world, tokenised in real time.

A camera streams pixels. A neural network — trained on billions of frames — carves the scene into bounding boxes and labels: person, kerb, ladder, gas line.

The decision system then predicts the next safe action the way an LLM predicts the next word: continuously, with probability, with no human in the loop.

That's why FSD works. That's why a $300 Roomba can finally map your house. That's why Optimus walks up stairs. The hardware caught up with the maths.

person · 0.97
ladder · 0.94
condenser · 0.91
safe path · 0.88

Optimus

Bipedal worker. Sees the floor, picks the part, places it. Same NN family as FSD.

Tesla FSD

Eight cameras. One neural net. Drives home in traffic without a single line of hand-coded rules.

Robot vacuums

The humble Roomba finally maps a room properly — 20 years after the prototype.

Robotic mowers

Vision-guided, GPS-fused, no boundary wire needed. The yard mows itself by 2027.

Grok Imagine · in-motion concepts

Drop the rendered MP4 clips into /docs/grok-imagine/ and they autoplay here. Each slot is captioned so you can match prompt → file → placement.

slot 01 · /grok-imagine/optimus-walk.mp4
Optimus walking through a workshop, neural-bbox overlay
slot 02 · /grok-imagine/fsd-night-drive.mp4
Tesla FSD POV at dusk, lane segmentation lit up purple
slot 03 · /grok-imagine/data-centre-orbit.mp4
Solar-powered data centre orbiting Earth, Starship visible
Where it pays back

HVAC and Fire are the perfect first home for AI.

Both industries run on paperwork — bills, work orders, service reports, compliance certificates, regulator-defined inspection regimes — wrapped around physical work that breaks predictably. Both have rich sensor and image data. Both have margin pressure that rewards the operator who can quote faster, dispatch smarter, and never miss a compliance test. AI compounds the advantage on every one of those.

Vertical · HVAC

Cooler, smarter, faster — without hiring an analyst.

HVAC bleeds value when techs spend hours on paperwork, when faults aren't predicted, and when quotes go out cold. AI fixes all three.

  • Predictive maintenance from vibration & sound NN models flag failing bearings and refrigerant leaks weeks before they trip — turning emergencies into scheduled work.
  • Energy optimisation per site Feed BMS data to a model. Get a weekly playbook of setpoint, duct, and run-time changes per building.
  • Bill capture & job matching Snap a Reece invoice, AI reads every line, allocates to the right job, posts to Xero. Hours back to the ops team — every week.
  • Quote & service report drafting Voice notes + a couple of photos in. Branded, line-itemised quote or service report out — ready for the customer in minutes, not days.
  • Smart dispatch Schedule the right tech, with the right parts, on the right truck — using skills, location, traffic, and SLA.
Vertical · Fire protection

Compliance without the clipboard.

Fire is regulator-driven. Every door, hose, sprinkler and detector has a test cadence. Miss one — face penalties or worse. AI turns the entire AS 1851 lifecycle into a quiet, automatic background process.

  • AS 1851 inspection scheduling Asset registers parsed once, tested forever. Auto-schedules the right test at the right cadence. No expired tags.
  • Defect classification from photos Tech snaps a defect, vision model classifies severity, attaches the right code, drafts the rectification quote.
  • Compliance certificates & ESM reports One-click building-by-building PDF for council, insurer or owner. Pre-signed, dated, attached to the asset register.
  • Drone roof & high-rise inspections Vision NN reads frames in real time, flags spalling, cracked sprinklers, blocked vents — all before the tech climbs.
  • Evacuation modelling Floor plan + occupancy + exits → AI simulates a full evacuation, finds bottlenecks, recommends signage and door changes.
"Both industries run on paperwork wrapped around physical work that breaks predictably. That's the perfect shape for an AI to help."
And for you

It's also the best personal coach you've ever had.

A Project for your health. A Project for your kids. A Project for your finances. A Project for the next five-year plan. Each one becomes a private, growing knowledge base — written in your voice, attuned to your context — that you can ask hard questions of any time. Just remember: cross-reference the politically sensitive ones with Grok, and never trust any single model on a high-stakes call.

H
Health intelligence Histamines, sleep, supplements, cardio zones — keep it in one project, ask the science (cross-check on Grok).
B
Business decisions Three macro risks to your business this quarter? Always Opus. Treat it like a $400/hr consultant who never sleeps.
F
Family & finances Statements in. Cashflow forecasts out. School-fee scenarios. Estate planning prompts. One Project per topic.
M
Memory you can trust Long-term, this lives in your own agent. Local. Backed up. Yours forever — not the model provider's.
Action plan

This week. This month. This year.

AI compounds. The earlier you start, the bigger the gap with the operator who waits. None of these steps require a developer. All of them require taste.

This week

Get the best brain you can buy

  1. Sign up to Claude on the Max plan and switch the default model to Opus 4.7.
  2. Sign up to Grok (premium tier) for cross-referencing health, science, controversial questions.
  3. Create your first three Projects: Business strategy, My health, The next 90 days.
  4. Drop in real source material — ABN docs, latest P&L, blood results, calendar.
This month

Move from chat to build

  1. Stand up a Cursor account. Walk through one small build (e.g. a website refresh).
  2. Pick one HVAC or Fire workflow that hurts most and prototype it: bill capture, AS 1851 schedule, quote drafting.
  3. Run a "before / after" — measure hours saved, errors reduced.
  4. Brief the team. Give one or two power users seats. Watch what they invent.
This year

Build your own Jeeves

  1. Dedicate a Mac mini or Mac Studio to be the agent host.
  2. Install Open Claw / Hermes. Wire it to email, calendar, accounting, pipeline.
  3. Let it run a daily morning brief — pipeline movement, P&L delta, alerts.
  4. Back the whole thing to GitHub so the agent and its memory survive any laptop change.
"AI got 10× smarter in 18 months. Now imagine 10× more, again. The owners who get on early are the ones who'll still be here in five years."