You Don't Need Millions. You Need Guardrails.

I went to a wedding and ended up in the most useful AI governance conversation of my month.

My table was young Asian entrepreneurs, half a dozen of them, none older than thirty-five. Every one of them was building a technology business. Not one of them was job-hunting. Between the starter and the speeches, the conversation wasn't about funding rounds or growth hacks. It was about oversight, about where the guardrails sit, about who signs off when the AI gets something wrong. I didn't bring the subject up. They did.

That surprised me more than it should have. Here's why it shouldn't have surprised me at all.

The New Builders

For the past month I've been at events, Civo Navigate in London, a handful of others, and the pattern has repeated every time. I keep meeting founders who chose to build rather than apply. Not because they had capital behind them. Because they didn't need it.

AI platforms have quietly removed the barrier that used to gate who got to start a technology business. Ten years ago, a direct-to-consumer retail business needed a developer for the storefront, an agency for the ads, and a marketer to run demand generation, and that was before you'd sold anything. Today one founder with an AI agent can do all three jobs at once, reasonably well, from a laptop.

Here's the mechanism, because "AI makes it easier" is the kind of claim everyone nods at and nobody unpacks. Every business idea lives or dies on the same cycle: innovate, test, fail, reiterate, succeed. The expensive part was never the innovating. It was the testing, because you needed a working prototype before you could show it to a real customer, and building that prototype meant hiring a developer or an agency before you had any evidence anyone wanted the thing at all. That's the stage that used to cost tens of thousands and months of runway, and it's the stage that killed most ideas before they ever met a customer.

Agentic AI collapses that stage to almost nothing. An agent can write the landing page, generate the storefront, draft the ad creative, launch the campaign, and answer the first customer messages, in roughly the time it used to take to brief one freelancer. The loop that used to run once a year now runs several times a week, because each cycle costs pounds instead of thousands.

Here's the shape of it, an illustration, not a real case I'm citing. A founder has an idea for a subscription box for a narrow hobbyist niche. Day one, an agent builds the landing page and a short explainer clip and launches a small ad campaign, spend under fifty pounds. Day two, the agent runs the sign-up flow and answers questions in a chat widget, and the founder watches whether anyone converts. Nobody does: that's a fail costing an afternoon, not a mortgaged year. A handful convert: the agent reruns the test overnight at a different price point. By the end of the week, the founder has run three or four full cycles of a loop that used to take a whole business a year, and a bank loan, to attempt once.

Bret Taylor, cofounder of Sierra and chair of OpenAI, made the same point in a McKinsey interview this month: "I think it will help entrepreneurs in every state, in every community." He wasn't talking about Silicon Valley. He was talking about the same thing I saw at that wedding table, someone starting a direct-to-consumer business with an AI agent handling demand generation, the storefront, the ad buying, the things they'd never have been able to hire for. His interviewer put the sharper version of it: the cost to start a business today is so low that we don't yet know what the resulting wave of entrepreneurship actually looks like.

I'll be blunt about what this means for SMEs specifically. The businesses that used to compete on access, on who could afford the agency, the developer, the consultant, are now competing on something else entirely. Capital is no longer the barrier. Trust is.

And trust doesn't come from a pitch deck. It comes from being able to show, plainly, that you know what your AI is doing and that someone is watching it.

Every Prompt Runs in a Data Centre

Here's the part almost nobody building a five-person AI-powered business has priced in: every prompt you send, every agent you run overnight, is electricity and water in a building somewhere, and that building's footprint is no longer a background detail.

The International Energy Agency's Electricity 2024 report said data-centre electricity consumption could reach more than 1,000 terawatt-hours in 2026, roughly what Japan uses, and that demand from data centres, AI and the cryptocurrency sector could double by 2026. Its Electricity 2026 report forecasts global electricity demand growing at an average of 3.6% a year over 2026 to 2030, with the expansion of data centres among the drivers. That isn't a hyperscaler problem you can shrug off because you're not Amazon. It's the reason your token bill and your carbon footprint are now the same conversation, not two different ones.

Governments have noticed. Australia's Senate Environment and Communications References Committee has an inquiry into AI and data centres under way. Submissions closed on 1 September and it is due to report in November. The Climate Council's submission raises concerns about what AI-driven data-centre growth means for the shift to renewables, the climate and water resources, and other submissions make the same point from another angle: the industry is scaling without measuring the water and energy it uses.

If a five-person team can't answer "what does our AI actually cost the planet" in an afternoon, that's not a compliance gap. It's a sales gap, because your next enterprise customer's procurement team is going to ask, and "I don't know" isn't an answer, it's a stalled deal. I wrote about exactly this scenario, a supplier questionnaire that stalls a deal over a single carbon question, in The Green Mitigation, Book 4 of the Safe AI SME Series.

What Civo Navigate Told Us

Three things stayed with me from that day at Civo Navigate London, and all three point the same direction.

Sovereignty is now a question of control, not a hosting checkbox: who can reach your data, under which laws, and who's accountable when an AI system acts on it. The US CLOUD Act means data stored in the UK isn't automatically governed in the UK, a distinction most founders haven't thought through.

Oversight is becoming the central issue, not a side conversation. As agentic AI moves into production, the question every board and every customer will ask is where humans sit in the loop and how decisions can be traced.

And sovereignty and sustainability are converging faster than either gets discussed alone. Where your data centre sits decides both whose law applies to your data and what energy mix powers your AI. The organisations that govern these two questions together, rather than treating them as separate departments' problems, will be much better placed when regulators and investors start asking about both at once.

Guardrails Are the Product

Here's the thing that separates the founders I met at that wedding table from most of the SME AI adoption I see in my advisory work: they're building oversight in from day one, not bolting it on after something goes wrong.

That's the emphatic point of this whole piece. In a world where anyone can build a technology business with an AI agent and a laptop, the product isn't the AI anymore. Everyone has access to the same intelligence. The product is the proof that someone is watching it, that a human made the call on what the AI can do unsupervised, and that you can produce the evidence when a regulator, an investor, or an enterprise customer asks for it.

This is the same five-stage sequence I laid out in the previous issue of the newsletter: Discover what's running, Govern it with a named owner, Operationalise the human-oversight boundary, Measure the evidence and the value, then Mature once the earlier stages actually hold. Provable oversight isn't a compliance tax on the new builders. It's what makes their product saleable to the people who write the biggest cheques.

The Builder's Checklist

Seven things you can do this week, whatever size your team is.

1. Know where your models run and what powers them. If you can't answer this in one sentence, find out before your next customer asks.
2. Name an owner for every AI tool in use. Not a team. A person.
3. Set a boundary on what each AI tool can do unsupervised. Write it down, even if it's one line.
4. Log the evidence that oversight happened. A screenshot beats a memory every time.
5. Cap your token spend. An unwatched agent left running overnight is a budget line waiting to become a headline.
6. Choose the smallest model that does the job. Bigger isn't a strategy, it's a cost.
7. Publish a one-paragraph responsible AI statement. It costs you an afternoon and it answers the question before your customer has to ask it.

Build fast. Build accountable. Build clean. The barrier to building a technology business has never been lower. The barrier to being trusted with one has never been higher, and that's the right trade.

If you're building something right now, guardrails and all, I'd like to hear what it is. And if you want the honest, unscored version of where your own AI governance stands today, the free 8-question AI Governance Readiness Quiz takes two minutes.