In the early 2000s, the phrase internet-first company described something specific: a business whose core operations, distribution, and product lived online by design, not as a retrofit. Amazon was internet-first; Barnes & Noble was not, even after it built a website. The difference mattered for a decade and shaped a generation of tech.
AI-first is the same kind of distinction. You can bolt AI onto an existing business — most companies are doing exactly that right now. Or you can build a business whose unit economics, product experience, workforce structure, and competitive moat all assume AI as a foundation. The second is much harder. It's also where the next generation of category-defining companies is being built.
This course is about that second path. It covers what an AI-first business actually looks like in different industries, how to design one from scratch, how to retrofit an existing company when the second path is the right answer, the economics of AI at scale, the founder profiles of current AI-first success stories, and the honest failure patterns of the AI-first companies that didn't make it.
If you finish every module, here's who you become:
In 2023, two insurance companies both announced they were "using AI." The first bolted a ChatGPT plugin onto its existing claims portal and called it an AI strategy. The second — Lemonade — had already been running a system called AI Jim since 2016: an AI that handled the entire first notice of loss, cross-referenced claims against policy data, and in one documented case, paid a claim in three seconds. Same technology era. Completely different businesses.
Every business today sits somewhere on a spectrum between AI-Added and AI-First. The distinction is not about how much money you spend on AI tools, how many subscriptions your team has, or whether your CEO mentions AI in earnings calls. It is about where AI sits in the architecture of your value creation.
An AI-Added business has human-designed workflows at its core, with AI sprinkled in at the edges: a Grammarly plugin here, a ChatGPT summary there, maybe a dashboard that uses machine learning to surface anomalies. These tools reduce friction. They do not reshape the fundamental economics of the business.
An AI-First business was designed — or fundamentally redesigned — so that AI sits at the center of value delivery. The workflows, the cost structure, the scalability assumptions, and often the product itself are impossible without AI at the core.
Lemonade (Insurance, 2015–present). Lemonade was not an insurance company that added AI — it was incorporated with the explicit design that AI would handle underwriting signals, fraud detection, and claims processing. By 2020, its loss ratio and operating costs were structured differently than any traditional insurer of comparable size because the variable cost of adding a new policyholder was near zero — a structural advantage unavailable to AI-Added competitors.
Duolingo (EdTech, 2023 pivot). Duolingo's April 2023 layoffs of contract content writers were accompanied by a public statement from CEO Luis von Ahn: the company was moving to AI-generated content for its language exercises. This was not a cost-cutting measure dressed up as strategy — the company's existing content production bottleneck meant it could only support a limited number of languages and exercise types. AI removed that bottleneck structurally. By Q3 2023, Duolingo reported a 62% increase in daily active users partly attributed to the speed of new course deployments made possible by the AI content pipeline.
Klarna (Fintech, 2024). In February 2024, Klarna's CEO Sebastian Siemiatkowski announced that its AI assistant — built on OpenAI — was doing the work of 700 customer service agents in its first month, handling 2.3 million conversations. Klarna did not add AI to an existing contact center. It rebuilt the contact center around AI and resized its human workforce accordingly. This is a structural, not additive, deployment.
AI-Added businesses improve margins incrementally. AI-First businesses change the shape of their cost curves. When Klarna adds 100,000 new customers, its customer service cost barely moves. When a traditional bank adds 100,000 customers, it hires proportionally. That asymmetry compounds over time into a structural competitive moat.
A common mistake — made by executives, consultants, and investors alike — is to conflate AI usage rate with AI-first posture. A company where every employee uses Copilot daily may be profoundly AI-Added: humans still design the process, AI assists execution. A company with lower measured AI usage but where the core product literally cannot exist without a model running at its center is AI-First.
The question to ask of any organization is not "do we use AI?" but "if AI were switched off tomorrow, would our core value proposition still function?" For an AI-Added business, the answer is yes — slower, but yes. For an AI-First business, the answer is no.
If your AI tools went dark at 9am Monday, would your core product still ship? Would customers notice a degraded experience, or would the product simply not exist? Your honest answer locates you on the spectrum.
You'll describe a real or hypothetical company's AI usage to the assistant. The assistant will help you classify it on the AI-First / AI-Added spectrum, identify where AI sits in the value-creation loop, and suggest what a genuine AI-First redesign might look like.
When Klarna rebuilt its customer service function around OpenAI's models in 2023, it did not simply automate existing scripts. It redesigned the entire interaction architecture: the AI had direct access to transaction data, could initiate refunds, update account states, and escalate to human agents based on sentiment signals it detected in real time. The humans who remained were no longer first-line responders — they were edge-case specialists and AI supervisors. The organizational chart changed shape. The cost per resolved ticket dropped by roughly 80 percent. The company did not become more efficient. It became a structurally different type of company.
AI-first businesses differ from AI-added ones across four structural dimensions:
The most consequential structural difference is what happens to quality over time. In a traditional business, quality is a function of the skill and experience of your workforce — it is costly to improve and easy to degrade through turnover. In an AI-first business, quality is increasingly a function of data and model iteration. Every customer interaction is, in principle, training signal.
Waymo's autonomous vehicle program illustrates this starkly. By 2023, Waymo had accumulated over 20 million miles of real-world autonomous driving data. This data is not a report — it is an infrastructure asset that makes the next mile safer than the last. A competitor entering the space cannot simply hire better drivers; it must accumulate comparable data, which requires comparable time and scale. The flywheel of an AI-first business compounds in ways that traditional competitive advantages do not.
Contrast this with a taxi company that uses AI to optimize surge pricing. That is AI-added: useful, profitable, but the pricing model does not make the rides themselves better, and the advantage does not compound through usage.
Ask: does each customer interaction make the product better for the next customer? If yes — if usage feeds improvement — you have the structural flywheel of an AI-first business. If usage just generates revenue but not capability, the flywheel is absent.
The structural shift in AI-first companies creates a new type of org chart. Anthropic, OpenAI, and Cohere are extreme examples — the majority of their headcount is researchers and engineers, not operators. But even non-AI-native companies that make the AI-first transition show predictable patterns:
None of this means AI-first businesses have no humans. It means the type of human work changes. Klarna's remaining customer service staff handle disputes where emotional intelligence and legal judgment are required — tasks where the model's failure rate remains high. The humans are not gone; they are repositioned at the edges where AI still fails.
Map your company's core workflow. At each step, ask: is AI inside this step (AI-First signal) or does it assist a human who owns this step (AI-Added signal)? The ratio of inside to outside gives you your structural posture.
Walk through a company's core workflow with the assistant. Together you'll identify which steps have AI inside them vs. assisting humans, whether a data flywheel exists, and what the headcount architecture looks like.
IBM spent over a decade and billions of dollars trying to make Watson a commercial AI platform. The technology was not the failure point. Watson could process language. It could analyze medical records. The failure was structural: IBM sold Watson into organizations where the workflows were not redesigned around AI. Hospital systems that licensed Watson for Oncology kept their existing diagnostic processes and asked Watson to fit inside them. By 2022, IBM had sold or wound down most Watson Health divisions. The lesson was not about AI capability — it was about the impossibility of grafting AI onto unchanged organizational structure and expecting AI-first outcomes.
Companies attempting the transition from AI-added to AI-first consistently encounter three categories of resistance:
The companies that have successfully transitioned share a pattern. They do not attempt to become AI-first everywhere simultaneously. They identify one core workflow where AI-first architecture is structurally possible — where data is available, the process can be redesigned, and the economic incentive is large enough to override organizational resistance — and they go fully AI-first in that one place first.
JPMorgan Chase's COiN program (Contract Intelligence, 2017) is the most cited example. The bank identified commercial loan agreement review as a target: 12,000 hours of lawyer time per year reviewing standard documents. COiN reviewed those same documents in seconds. Crucially, JPMorgan did not try to make its entire legal operation AI-first. It picked the most tractable workflow, made it AI-first, demonstrated the economic case, and used that to build organizational will for the next workflow. By 2023, JPMorgan had over 300 AI use cases in production — but the path started with one.
Stitch Fix (fashion subscription, 2011–present) provides a different pattern. Its founding premise was that AI would drive clothing selection — every box a customer receives is selected by a combination of algorithm and human stylist, with the algorithm doing the first-pass filtering from millions of SKUs. The human stylist adds the final editorial judgment. This is a hybrid AI-first architecture: AI owns the core filtering function; humans add the edge-case judgment. This was designed in from the start, not retrofitted.
IBM Watson's failure and dozens of similar enterprise AI deployments share a common pattern: organizations expected AI to improve outcomes without changing the processes that determined outcomes. AI-first architecture requires process redesign. There is no shortcut through this.
For an existing business attempting the transition, the practical question is not "how do we become AI-first?" but "where should we start?" The answer involves three filters:
| Filter | What to Look For | Red Flag |
|---|---|---|
| Data availability | High-volume, structured, historically consistent data about the workflow | Data locked in PDFs, email threads, or human memory |
| Economic size | Large enough cost or revenue impact to justify redesign investment | Marginal workflows that won't move the needle if transformed |
| Organizational will | A senior champion who benefits from the transformation and can absorb the resistance | Mid-level sponsorship with no executive backing and no budget authority |
The best starting workflow for an AI-first transition is not the most exciting one — it is the most tractable one. High volume, clean data, large economic impact, and a senior champion who won't be fired for disrupting it. JPMorgan picked loan agreement review for exactly these reasons.
Describe your business (or a business you're analyzing) and its major workflows. The assistant will apply the three-filter framework — data availability, economic size, organizational will — to help you identify which workflow is the best candidate for an AI-first transition.
In early 2023, a leaked internal Google memo titled "We Have No Moat, And Neither Does OpenAI" argued that open-source AI models were closing the capability gap with proprietary ones at a rate that would eliminate model quality as a durable advantage. Whether that specific prediction proves accurate, the underlying logic is sound: if the model is the moat, and models are becoming commodities, the moat does not hold. The companies with durable AI-first advantages are not winning on model quality alone. They are winning on data, distribution, and workflow lock-in.
In an AI-first competitive landscape, durable moats form in four places. Model quality is not the primary one:
| Moat Type | How It Works | Real Example |
|---|---|---|
| Proprietary Data | Data that cannot be replicated — generated by your customers, your operations, or your unique access — and that makes your model better than anyone else's for your specific domain. | Bloomberg's financial data corpus; Veeva's pharmaceutical CRM data; Waymo's miles driven |
| Workflow Lock-In | AI embedded so deeply in customer workflows that switching requires rebuilding the workflow itself, not just changing a tool. | GitHub Copilot embedded in developer IDE; Harvey AI in law firm document workflows |
| Network Effects + AI | More users → more data → better model → better product → more users. The AI amplifies a traditional network effect into a faster flywheel. | Waze traffic routing; Duolingo exercise quality improving with more learners |
| Brand Trust in High-Stakes Domains | In domains where AI errors have serious consequences (healthcare, legal, finance), established trust and regulatory credentials create a barrier that technical quality alone cannot overcome. | Epic Systems in healthcare AI; established accounting firms in AI-assisted audit |
Three traditional competitive advantages are being rapidly eroded in AI-first markets:
Access to talent alone. In traditional software, hiring the best engineers was a durable advantage. In AI-first, talent matters but is rapidly commoditized through open-source models, APIs, and tooling. A startup with three engineers and access to GPT-4 can build capabilities that would have required 50 engineers five years ago.
Feature richness. AI capabilities can be added to products so rapidly that feature gaps close within quarters, not years. The pace of iteration in AI-enabled products makes traditional product moats fragile.
Scale without data differentiation. Being large used to mean better resources and distribution. In AI-first markets, scale only matters if it generates proprietary data. A large company processing generic transactions has no AI advantage over a smaller company processing the same generic transactions. Scale + unique data = moat. Scale alone does not.
Bloomberg built a financial data terminal moat not on technology but on data and workflow lock-in. Traders integrated Bloomberg so deeply into their workflows that the $24,000/year terminal became functionally irreplaceable. Bloomberg GPT — launched in 2023 — extends this: a 50-billion parameter model trained on Bloomberg's proprietary financial corpus. The model is not the moat. The corpus is the moat. The model just makes the corpus more accessible.
For founders and builders, the implication is clear: if you are building an AI-first business, ask from day one what your proprietary data source will be. The answer drives architecture decisions. If you generate unique behavioral data through your product (like Lemonade's claim patterns), protect that data and build your model on it. If you serve a domain with no existing structured data corpus, consider whether your product can generate that corpus through operation — and whether the resulting data is defensible.
Harvey AI, the legal AI platform backed by Sequoia and OpenAI, illustrates this pattern. The company's 2023–2024 growth was not primarily a function of model quality — law firms can access the same base models. Harvey's advantage comes from fine-tuning on legal documents generated through its platform's operation, and from workflow integration that creates switching costs. The moat is in the data and the embeddedness, not the base model.
AI-first is not a technology posture — it is an architectural, organizational, and strategic posture. The technology is increasingly available to everyone. What separates durable AI-first businesses is what they build on top of the technology: proprietary data, embedded workflows, and network effects that compound with every interaction.
Describe a business you're building, running, or analyzing. The assistant will help you identify which of the four moat types (proprietary data, workflow lock-in, network effects + AI, brand trust) applies to your situation, assess the durability of your current advantages, and surface moat-building moves you may not have considered.