AI-Native Companies Will Outperform Everyone Else | Harvard Business School, Rem Koning
In a Nutshell
AI-native companies outperform others by allocating intelligence—deciding what tasks go to tools like Claude, ChatGPT, or humans—reimagining workflows with agentic AI agents that scale via compute rather than headcount, as seen in Gamma's presentation tool and global founder sprints yielding 20% more weekly output in Nigeria. High-performers with judgment amplify AI's value, while strugglers decline (e.g., 10% profit drop in Kenyan study), emphasizing the need for earned insight over raw tool power. AI equalizes routine tasks but drives leapfrogging in emerging markets, proliferating small SaaS and entrepreneurship, though builders must validate user demand to avoid overengineered failures.
These notes were generated by AI and may contain inaccuracies.
Wrote a paper 3 years ago examining whether ChatGPT delivered over WhatsApp could help small business entrepreneurs in Kenya. Struggling entrepreneurs (lower baseline profits and revenues) saw a 10% decline in profits and revenues from talking to the AI. High-performing entrepreneurs (revenues and profits above median) did better. Even with advanced models like GPT-5.2 or Claude, results would likely be the same.
History of business shows edges from allocating resources: Warren Buffett excels at allocating capital at Berkshire Hathaway; companies like McKinsey at allocating talent. Now, key is allocating intelligence for market edge.
Hi, I'm Rem Koning, professor at Harvard Business School, studying entrepreneurship and AI to help entrepreneurs worldwide, from Silicon Valley moguls to coconut sellers in Indonesia, bring new products to market, compete, grow firms, and unlock entrepreneurial potential with AI.
AI changes firm-building: questions on who builds AI-native firms, scaling, strategy. AI Founder Sprint from INSEAD involved over 500 entrepreneurs globally (quarter each from Africa, Asia, Americas, Europe), all building AI solutions like AI for maternal health in Africa, edtech in India, startups in Kenya, Europe, US.
Tracked AI usage; preview results: teaching founders to be AI-native—using generative AI, ChatGPT, Claude, vibe coding tools, multimodal tools, agents—helps everywhere. In Nigeria, 20% more done weekly, more customers, product launches, revenue. Despite success and faster growth, founders want to raise $250,000 less capital.
Founders reimagine workflows, build custom AI for bottlenecks like customer acquisition: AI auto-builds marketing strategy, go-to-market plan, executes it. Used off-the-shelf tools to build own agents, scaling with on-demand compute instead of headcount, changing business economics, especially outside Silicon Valley.
Gamma example: value from process (AI in coding, customer support) and embedding AI in product so AI works directly with customer, removing humans from loop. Humans don't scale well; pre-generative AI Gamma would need tens/hundreds of thousands of graphic designers, collapsing economics. Instead, scales with compute.
Key for AI-native founders: find loops where AI works with user/AI/website without team involvement.
Edge from allocating capital (Warren Buffett, Berkshire Hathaway), talent (McKinsey: hiring, partnering, client matching). Now allocate intelligence: decide tasks for Claude, Lovable, Grok, DeepSeek; blend/orchestrate models. Also allocate between AI and humans—humans have edges in thinking differently for strategy/market edge.
Blend human intelligence where it adds unique value over models. All struggle with personal allocation (e.g., what for ChatGPT vs. self), now strategic for firms: what entrepreneur does vs. AI, which AI. Better allocation gives market edge.
AI equalizes existing work: all code with Lovable, decks with Gamma, editing with ChatGPT/Claude. But for new businesses/products/firms, returns greatest for those with judgment (prior companies, technical background)—their agency amplified.
Kenyan study repeat: struggling entrepreneurs declined 10% profits/revenues—asked many questions but lacked judgment to pick good advice. High performers asked similar but followed good advice. With Claude Opus over WhatsApp, same results: AI gives 4-5 plausible options, need judgment to choose.
Mistake: stuck in chatbot world (e.g., chatbots for Shopify, entrepreneurs, Harvard). Chatbot advice like "update website" unhelpful without coding skills.
Today, use agentic AI: tools to build websites, launch campaigns. For time-constrained entrepreneurs, virtual employees expand strengths. Current work explores this.
Shift from conversations to AI taking actions/interfaces. Startups edge: contextual knowledge of workflows/situations/world regions, not beating OpenAI/Anthropic at frontier models.
Skills in Claude code: context snippets teaching AI tasks, shared globally, highly effective.
More people become entrepreneurs/entrepreneurial due to easy building. World more entrepreneurial. Developing markets leapfrog like India's UPI payment infrastructure (phone payments everywhere, unlocked billions, many startups).
AI opportunity in knowledge work if models have emerging market context (e.g., Kenyan entrepreneurs, education). Application layer coming; inference costs dropping exponentially to near-free, opening low-income markets.
Concrete: virtual agents as experts (e.g., NYC/SV marketer in Nairobi), marketing, exporting, skilling workforce.
Proliferation of software for unimagined problems: deep (AlphaFold), mundane (CRM for Thai restaurants via Lovable/Codex). Solves millions/billions problems; software scalable, low marginal costs, turns knowledge to help billions. World like software economy.
Downside: past software concentrated wealth (mega-billionaires, platforms like Facebook with network effects). Vibe coding may lead to small SaaS, bootstrapping without VC, spreading prosperity—but policy needed for wealth concentration.
Most dangerous assumption: building with AI means people want it. Loop: AI tools fun, build overengineered software fast, launch to crickets. Traditional rules apply: get in front of users.
Little AI goes long way: Gamma core was sentences-to-deck; rest traditional software—one unlock built business. Find smallest AI application unlocking workflows.
AI gives unearned insight without judgment/taste—focus on earned insight, knowing where/how to apply tools, not next model (last gen fine).
Drew Bent, leads education at Anthropic: learning changing; accelerate by giving AI complex problems, elevate ambition. Push further as tools smarter. Future: inversion of control—AI does high-level strategy, delegates to humans for taste/agency.
Keep EO in your library
Save the videos and channels worth coming back to, and find them again in one place.
