When Harvey launched in 2022, most lawyers assumed it was another document search tool. By mid-2023, Allen & Overy had deployed it firm-wide. The difference wasn't the technology — it was that legal work contains all four signals that make a domain AI-native: high volume of repetitive expert judgment, tolerance for probabilistic output, high cost of human labor, and a structured underlying corpus. Harvey's founders didn't find a legal problem and attach AI. They found a pattern and built around it.
The clearest AI-native signal is a workflow that requires genuine expertise but applies the same reasoning pattern hundreds or thousands of times. Contract review, radiological image reading, credit underwriting, and customer support escalation triage all share this structure: a trained human applies a learned heuristic to a new case, produces a judgment, and moves on.
Why AI compounds here: Each judgment produces data. Data trains better models. Better models require less human review. The economics improve automatically as volume increases — the inverse of human-scaled operations where costs grow linearly with throughput.
When Waymo reported in June 2023 that its Phoenix robotaxi fleet had driven over 1 million autonomous miles without a life-threatening injury, analysts noted that each mile made the next mile safer. That compounding feedback loop — unavailable to any human driver fleet — is the defining signature of signal 1.
Recursion's platform takes fluorescence microscopy images of cells exposed to compounds — roughly 2.2 million images per week — and trains models on the resulting perturbation patterns. A human biologist evaluating each image would need centuries. The volume itself was the barrier; AI converted that barrier into an asset. By 2023, Recursion had the largest proprietary biological dataset of this kind in existence.
AI models learn from pattern in data. Industries sitting on large, structured, historically accumulated datasets are disproportionately AI-native. The data doesn't need to be clean — it needs to exist in sufficient volume and with sufficient labeling that a model can learn to generalize.
Bloomberg's 2023 release of BloombergGPT illustrated this precisely. Bloomberg had accumulated 40 years of financial news, filings, earnings transcripts, and market data. That corpus — unavailable to any general-purpose model — produced a financial language model that outperformed GPT-4 on financial NLP benchmarks despite being smaller. The opportunity was latent in the data before the model existed.
The implication for founders: Identify who sits on domain-specific historical data that hasn't been fully structured or exploited. The data moat often precedes the product opportunity by years.
AI economics are most favorable where the human alternative is expensive. This is why legal, medical, financial, and engineering domains attract disproportionate AI investment. When a U.S. corporate partner bills $1,200/hour for contract review, a model that completes first-pass review at $0.04 per page doesn't need to be perfect — it needs to be good enough to shift the burden.
Aidoc, which deploys AI for radiology triage, entered a market where a radiologist's time cost roughly $350,000/year in the U.S. and where emergency imaging backlogs routinely delayed treatment. The economic signal was unmistakable. By 2023, Aidoc's software was operating in over 900 hospitals globally. The business case required no exotic assumptions — only that an AI system could reliably flag the most urgent scans before a radiologist reviewed them.
AI models are probabilistic — they produce distributions, not certainties. Some workflows can absorb probabilistic output; others cannot. Identifying which is which is essential to viability assessment.
Tolerant workflows: Content drafting (human edits the draft), product recommendation (some mismatch is acceptable), fraud detection (false positives are investigated, not acted upon automatically), and marketing copy generation all tolerate a meaningful error rate because a human or downstream system catches failures.
Intolerant workflows: Autonomous surgical decisions, nuclear plant control, and autonomous financial execution at high velocity require near-perfect reliability. These are not AI-native opportunities today — they are future opportunities contingent on reliability thresholds that don't yet exist at scale.
Score any potential opportunity across all four signals: (1) Does the work involve repetitive expert judgment at volume? (2) Does structured historical data exist or accumulate? (3) Is human expert time costly relative to AI inference? (4) Can the workflow tolerate probabilistic output with human or system backstop? Strong candidates score yes on at least three. The strongest AI-native businesses score yes on all four.
Choose any industry or workflow you're interested in — healthcare administration, legal discovery, construction project management, insurance underwriting, or anything else. The AI advisor will guide you through scoring it against the four signals of AI-native opportunity from Lesson 1.
Complete at least 3 exchanges to finish this lab.
In 2021, a startup called Midjourney did not displace any professional illustrator at scale. It created a new market: people who wanted images but would never have hired an illustrator — either because they couldn't afford one, couldn't articulate what they wanted well enough to brief one, or needed images in volumes and speeds no illustrator could provide. By 2023, Midjourney had over 14 million Discord users. Most were not former illustration buyers. They were a new class of customer unlocked by a new capability.
AI applications fall into two fundamental categories. Displacement opportunities replace an existing workflow — often human labor — with an AI system that performs the same function more cheaply or quickly. Expansion opportunities do something that wasn't economically or practically feasible before AI, creating a market where none existed.
Displacement opportunities are real and often large. But they face a structural challenge: they compete directly with established workflows that have entrenched stakeholders, existing pricing, and procurement relationships. The incumbent is often the customer's current vendor. Replacing them requires not just superior performance but overcoming switching costs.
Expansion opportunities compete with nothing. There is no incumbent to displace because the thing didn't exist. The challenge shifts from competitive positioning to market education — convincing potential customers that the new capability solves a real problem they've been tolerating.
ElevenLabs launched voice synthesis that could clone a speaker's voice from a short audio sample and generate speech in any language at near-real-time speed. The displacement narrative would be: "replaces voice actors." But the expansion story is larger: audiobook production for titles that have never had audio editions (there are roughly 4 million English-language books with no audiobook), real-time localization for content creators in 29 languages, and personalized voice interfaces for apps that previously had no voice layer at all. The expansion market dwarfs the displacement market by an order of magnitude.
A practical diagnostic for expansion opportunities: ask whether the use case was technically possible before AI but economically prohibitive, or whether it was genuinely impossible before the capability existed.
Personalized nutrition coaching at scale is an example of the first type: a human nutritionist could theoretically monitor each person's dietary intake, sleep, exercise, and biomarkers and provide daily guidance. The problem was cost — $300/hour experts cannot serve 300 million people. AI makes it economically viable without making it technically novel. This is expansion through democratization.
Drug molecule generation — where models like DeepMind's AlphaFold predict protein structures that took years of lab work to determine — is genuinely impossible-to-possible. Before AlphaFold's 2020 breakthrough, computational protein structure prediction had an accuracy ceiling that made most practical applications non-viable. The expansion didn't come from cheaper biology; it came from a new capability that didn't exist.
Displacement opportunities are bounded by the size of the market being displaced. If the total spend on U.S. legal document review is $12 billion, the maximum addressable market for AI document review is approximately $12 billion — and capturing even a fraction requires displacing law firms' existing workflows and billing structures.
Expansion opportunities can exceed the size of any existing market because they create new demand. GitHub Copilot didn't expand into the "IDE market" — it created a new category of AI-assisted development that generated new spending by developers and organizations that had no prior budget line for that function. GitHub reported in 2023 that Copilot users completed tasks 55% faster — the value wasn't replacing another tool, it was unlocking time previously unavailable for other work.
When evaluating an opportunity, map both stories. What is the displacement narrative and what is the expansion narrative? The strongest AI businesses have a displacement story that generates early revenue (replacing something that exists) and an expansion story that justifies venture-scale growth. The displacement pays the bills; the expansion justifies the valuation.
Many AI-native expansion opportunities follow a democratization pattern: capabilities previously available only to large organizations or wealthy individuals become accessible to individuals, small businesses, or underserved markets. This pattern has structural advantages — it doesn't require displacing existing customers' vendors, and it addresses markets that are often much larger in aggregate than the premium markets where the capability originated.
Jasper.ai's growth to $75M ARR in its first 18 months (reported 2022) didn't come primarily from replacing agency copywriters. It came from founders, marketers, and small businesses that previously produced no long-form content at all — or produced it at low quality — because they lacked access to professional writers. The market being served didn't previously exist in the form Jasper addressed it.
Pick any AI business idea — real or hypothetical. The advisor will help you articulate both the displacement narrative (what it replaces) and the expansion narrative (what new market it creates or what previously impossible thing it enables).
Complete at least 3 exchanges to finish this lab.
When Stripe launched Radar, its fraud detection system, in 2016, it didn't try to automate Stripe's entire payment operations. It decomposed the workflow: authorization, settlement, dispute resolution, customer support, and fraud review were all separate layers. Only one layer — real-time transaction scoring — had the characteristics that made AI structurally superior: millisecond decision windows, high volume, historical labeled data (fraudulent vs. legitimate transactions), and a tolerance for some false positives. Radar became one of the most effective fraud prevention systems in payments precisely because it was scoped to the right layer.
Workflow decomposition is the practice of breaking a multi-stage process into discrete steps and evaluating each step independently against the four-signal framework from Lesson 1. Most complex professional workflows contain three to seven distinct stages. Typically, one or two of those stages are AI-ready; the rest remain human-dependent for structural reasons.
A standard corporate M&A transaction illustrates this clearly. The workflow includes: target identification, initial valuation modeling, due diligence document review, legal structuring, regulatory analysis, negotiation, and post-merger integration. Each step has a different AI-readiness profile.
| M&A Stage | AI Readiness | Reason |
|---|---|---|
| Document due diligence | High | Repetitive, structured corpus, probabilistic tolerance |
| Financial modeling | High | Historical data patterns, formula-driven, auditable |
| Target identification | Medium | Data-driven screening but relationship context missing |
| Regulatory analysis | Medium | Structured statutes but novel precedent matters |
| Negotiation strategy | Low | Relationship-dependent, real-time, non-computable |
| Post-merger integration | Low | Organizational politics, cultural nuance, change management |
Companies like Kira Systems (acquired by Litera in 2021) and Luminance built exactly on this insight: they addressed only the document review layer of legal and M&A workflows. They didn't try to automate negotiation or post-merger integration. That scope discipline was the product strategy.
Nabla, a French AI startup, decomposed the clinical workflow and identified that physician documentation — the act of writing clinical notes from a patient encounter — consumed approximately 37% of a doctor's time (per AMA studies) but had zero patient-facing value. The diagnosis, treatment decision, and bedside manner remained entirely human. Nabla built an ambient AI that listened to encounters and produced structured clinical notes. By mid-2023, Nabla reported 45,000 healthcare providers using the platform across the U.S. and Europe. The scope was exactly one layer of a much larger clinical workflow.
A common failure mode in AI product development is expanding into adjacent workflow layers before the core layer is defensible. This happens for understandable reasons — customers ask for more features, the adjacent layer seems natural, and scope creep feels like growth. But it often dilutes the core advantage.
When an AI document review tool tries to also handle negotiation strategy, it loses the precision that made the document layer valuable and adds a layer where AI has no structural advantage. The product becomes mediocre at everything rather than excellent at one thing.
The discipline of identifying one AI-ready layer and staying precisely there — at least until that layer is deeply defensible — is one of the most consistently observed patterns in successful AI-native companies during the 2020–2024 period.
One consistently overlooked layer in workflow decomposition is the interface between workflow stages — the handoff from one step to the next. In many complex workflows, significant friction and information loss occur at these handoffs. An AI layer that sits at the interface — summarizing, translating, re-formatting, or routing information between stages — can create substantial value without owning any core stage.
Notion AI's initial insertion into professional workflows was largely at this interface layer: turning meeting notes into structured action items, summarizing long documents for quick review, and converting rough drafts into polished text. None of these replaced the core creative or analytical work — they reduced the friction cost of moving between stages of that work.
List every step in a target workflow. For each step, evaluate: (1) Is this step primarily information processing or relationship/judgment dependent? (2) Does historical data exist to train a model on this step? (3) Is this step's output verifiable without full human re-execution? (4) Would speeding or cheapening this step change downstream economics materially? Steps that score yes on 3–4 of these are the AI-ready layer. Focus there first.
Choose a professional workflow — hiring and recruiting, insurance claims processing, content production, customer onboarding, or anything with multiple stages. Walk through the decomposition protocol with the advisor to identify which specific layer is AI-ready and which layers should remain human.
Complete at least 3 exchanges to finish this lab.
In October 2023, OpenAI released GPT-4 with vision capabilities. Within 72 hours, analysts counted at least 14 VC-backed startups whose core product had been building a GPT-4-vision wrapper that now faced direct competition from OpenAI itself. Some of those companies had raised $5–20M. Their moat — access to a powerful API — turned out to be a shared resource, not a proprietary one. The question of competitive moat in AI is not academic. It determines whether a business has a 24-month window or a 10-year one.
Durable competitive advantage in AI-native businesses derives from five distinct sources, which vary in strength and in how quickly they can be established. Understanding which source your business relies on — and whether that source is currently protected — is essential to opportunity evaluation.
The most durable AI moat is data generated by the act of operating the business. Each transaction, each model inference, each user correction becomes training signal that makes the next inference better. This creates a flywheel that accelerates rather than decays over time — and it is structurally unavailable to a new entrant.
Palantir's competitive position illustrates this precisely. The company's Foundry and Gotham platforms generate operational data through use that continuously improves their ontology models. By 2023, Palantir had processed operational data from over 100 large enterprises and government agencies. A competitor building a technically superior product in 2024 would start with zero of that operational history. The data moat is not about the size of the training set — it's about the specific operational context encoded in years of real-world use.
Scale AI built what appears externally to be a data labeling business. The actual competitive moat is the evaluation framework they've built across millions of annotation tasks — understanding exactly where human raters disagree, what edge cases exist in each domain, and how to calibrate models against real-world human judgment. That calibration knowledge accumulates with each project. A competitor starting a data labeling business in 2024 can hire the same number of annotators but cannot replicate Scale's 8 years of understanding about what makes annotation reliable in each specific domain.
Deep workflow integration creates switching costs that are often more durable than product quality advantages. When an AI system is embedded in a customer's core operations — with custom models trained on their data, outputs feeding into their downstream systems, and teams trained to review its work — the cost of switching exceeds the marginal benefit of a technically superior alternative.
Veeva Systems, which serves pharmaceutical companies, illustrates the principle. Veeva's CRM and regulatory data management tools are embedded so deeply in pharmaceutical workflows that switching costs are estimated at 12–24 months of implementation time for large enterprises. When Salesforce entered the pharmaceutical CRM market in 2022 with a competitive product, market share change was minimal — not because Veeva's product was technically superior on every dimension, but because the integration depth made any alternative prohibitively disruptive.
The weakest AI competitive position is a product that provides UI and user experience around a third-party model API without accumulating any of the five moat types. These products are vulnerable on three fronts: the underlying model provider can add the same features natively (the OpenAI vision case), a competitor can build the same UI in weeks, and the model provider can change API pricing unilaterally.
This does not mean API-based products are unviable — it means viability requires that something other than the API is doing the competitive work. That something is usually workflow integration (sticky customer relationships), domain-specific fine-tuning data that improves performance on the specific task, or a go-to-market motion that reaches customers the model provider cannot reach directly.
For any AI opportunity you're evaluating, explicitly answer: (1) What data does operating this business generate that a new entrant cannot access without operating the same business? (2) What does deep customer integration look like, and how many months/years does it take to unwind? (3) What regulatory, certification, or expert network positions can be built that create barriers independent of model quality? If you cannot clearly answer at least one of these questions, the competitive position is fragile.
Describe an AI business — one you're building, one you're evaluating for investment or partnership, or a real company you want to analyze. The advisor will walk you through the five moat types and help you assess which sources of durable competitive advantage are present, absent, or buildable.
Complete at least 3 exchanges to finish this lab.