The problem
Homes are quietly turning into little power stations — solar on the roof, a battery in the garage, an EV on the driveway, a heat pump in the cellar. Each device is happy to be “smart” on its own. The hard part is that they all share the same solar array, the same grid connection, the same prices — and the same comfort you never want broken. Making one device clever is a commodity. Orchestrating the whole home against those shared constraints is the real, valuable, difficult problem.
That’s a serious piece of engineering: optimisation, real hardware integration, forecasting, and a trust problem — people won’t hand over their home unless they understand what it’s doing.
The approach
We treated AI not as a code-vending machine but as a thinking partner across the whole build — architecture, domain modelling, code, copy, design. We stayed accountable for judgment: what to build, what to leave out, where the standard sits. AI carried an enormous amount of the how; we held the what and the why.
The discipline that made it work is the same one we bring to organisations: lean, ruthless about scope, clear about structure. Start with the smallest real instance of the problem — the EV today — while keeping the architecture open for the bigger vision. Don’t build what you don’t need yet; don’t paint yourself into a corner either.
AI collapsed the distance between an idea and working software. It raised the bar on judgment.
What we learned
The bottleneck moved — it didn’t disappear
When building gets cheap, the scarce thing becomes deciding what’s worth building. AI made the code fast; it didn’t decide which problem mattered. Direction, taste and judgment became more valuable, not less.
The hard part was never the code
It was the domain: shared-constraint optimisation, messy hardware, real homes. AI accelerates the typing. Understanding the problem deeply enough to model it correctly is still on you.
Rules without a gate are a wish
A rule only constrains if something enforces it. A human feels the weight of a written standard; an AI generating code feels nothing, and will break an ungated rule a hundred times an hour without noticing. So the standards that mattered got real gates — checks that fail, boundaries the code can’t cross — not just a line in a document. The rest we left fast and loose.
Trust is a feature, not a footnote
Automation only earns its keep if people believe it. So the product explains every decision it makes in plain language. The same is true inside teams adopting AI: visibility and explanation are what turn a black box into something people will actually rely on.
Working with AI is a human skill — with a hard floor
Clarity, decomposition, delegating the “how” while owning the “what”, holding a quality bar, knowing when to step in — judgment, not seniority. But there’s a floor: AI only amplifies judgment you already have. Where you can’t evaluate the output — not a qualified engineer, not a domain expert — you’re not delegating, you’re flying blind. It worked here because real expertise stayed in the loop to tell good from confident-but-wrong.
Why this sits on a coaching site
Because it’s the same story from two angles. The agility to perform — under pressure, with new tools — isn’t a slide. It’s behaviour. Building Energy AutoPilot was us practising, on a real product, exactly what we help leaders and teams do: stay effective when complexity rises and the tools keep changing.
Tools amplify. Behaviour decides.