Questions every AI startup founder should ask before AGI arrives

Executive overview

Most founders plan six months ahead for AI. That horizon is too short. AGI may arrive in two to three years, and it will reshape not just products but team structures, trust models, and whether software-as-a-service makes sense at all.

The questions that matter now are structural: what is your defensible moat when a mega-corp can replicate your product with a prompt, who controls the AI agents your users rely on, and how do you build trust without human whistleblowers inside your company?

The founder's edge is asking the right questions before the rules change again.

Products in a world of on-demand software

• Enterprises may stop buying SaaS and simply build in-house with tools like Claude Code.
• Consumers may stop downloading apps and generate on-demand software for themselves.
• Counter-pressure: AI also raises the quality ceiling — exceptional products still require great teams working with AI.
• On-demand code generation raises the possibility of apps that extend themselves at runtime for individual users.
• Runtime code generation requires deep model trust — current models are not reliable enough for this.
• AI-native products built from scratch may not always beat retrofitted incumbents with existing distribution.

Team structure and culture

• AI-native teams built from scratch may have an advantage over large companies downsizing with AI.
• What "AI-native" looks like changes every 12–18 months — yesterday's native is tomorrow's legacy.
• Smaller teams mean fewer internal checks; a single bad actor can reshape a product with no one to object.
• Human diversity inside companies is a trust mechanism — it erodes as teams shrink or automate.

Trust, agents, and alignment

• Personal and professional agents need to collaborate without leaking private context across boundaries.
• An agent built by an ad-based company may optimise for the company's interests, not the user's.
• As models become more capable, misaligned agent incentives become more dangerous, not less.
• Long-horizon agents (running for days or weeks) require alignment progress just to be economically viable.
• Economic pressure to solve alignment is real: agents that go off the rails over long tasks are unusable.

New guardrails and auditing models

• AI-powered audits could verify company behaviour against public mission statements.
• An AI auditor can delete its findings after a clean audit — no risk of IP or sensitive data leaking.
• Binding commitments to neutral AI audits are more credible than public values statements.
• These mechanisms do not exist yet but will likely be necessary as human guardrails disappear.

Defensible advantage after AGI

• Custom proprietary data still matters in domains where tacit knowledge has not reached the internet (e.g. semiconductor fabrication, materials science).
• Technical moats around capacity, routing, and fine-tuning are real but temporary — models improve and the advantage closes.
• Hard physical problems — infrastructure, energy, chips, manufacturing — remain hard even as software commoditises.
• Some tasks will hit an intelligence ceiling where quality saturates; commoditisation pressure accelerates there.
• Ask: can a mega-corp replicate your startup by throwing more tokens at the next model? If yes, rethink defensibility.

Neutrality and societal stakes

• A handful of corporations currently decide what AI models will and will not do.
• If AI becomes infrastructure, the question of neutrality — like electrical grid neutrality — becomes critical.
• Founders have a narrow window to build products that are genuinely good for users, not just engaging for 20 seconds.
• Build something people want means building something society needs — trust, long-term wellbeing, not just consumption.