Original practice question. The scoring rubric is our teaching guide. How we research and score answers
TL;DR: Model access is not a moat when everyone has it. Defensibility comes in layers, roughly from weakest to strongest: prompt and UX polish, proprietary context (customer data connected to the product), ownership of the workflow, an evaluation and data flywheel built from real failures and outcomes, and distribution and trust. Most AI applications start on the bottom rung; the strategy is to climb deliberately, measuring progress on each rung.
How to approach it
Concede the premise: anyone can call the same model, and models improve for everyone. Then lay out where defensibility comes from, assess where the product sits today, and propose the specific investments that move it up, with how you would know they are working.
A strong answer
The ladder.
| Rung | What it is | Why it defends | How fragile |
|---|---|---|---|
| 1. Prompt and UX | Clever prompts, nice interface | Speed to market | Copied in weeks |
| 2. Proprietary context | Connectors to customers' data; retrieval over their documents, history, systems | Answers competitors cannot give without the same access | Moderate; competitors can build connectors, but switching means reconnecting |
| 3. Workflow ownership | The product is where the work happens: approvals, collaboration, records | Switching means changing how teams work | Strong |
| 4. Eval and data flywheel | Failure taxonomies, labelled outcomes, domain-specific evals that improve the product with every use | Quality compounds; new entrants start from zero | Strong, if the data is proprietary and actually used |
| 5. Distribution and trust | Brand, enterprise relationships, certifications, embedded in customers' procurement | Hard to dislodge | Strongest |
Where most AI applications start. Rung 1, sometimes 2. That is fine for a start, but a company that stays there is one model release away from irrelevance, because the model provider or a well-funded competitor can match it.
Climbing, with an example. Take an AI assistant for contract review used by in-house legal teams.
- Rung 2: connect to the customer's document management system, their past contracts and their playbook of preferred clauses. Now answers reflect "how we negotiate", not generic legal knowledge.
- Rung 3: become where negotiation happens: redlines, approval routing, version history, clause library. Leaving means retraining the team.
- Rung 4: every accepted or rejected suggestion is a label. Build a failure taxonomy specific to contract types and jurisdictions, evaluation sets no competitor has, and use them to improve weekly. The product gets measurably better in ways a competitor cannot replicate quickly.
- Rung 5: security certifications, references from respected legal teams, integration into procurement and compliance processes.
Measuring the climb. Share of usage that depends on connected customer data, share of workflow steps completed inside the product, number and growth of labelled examples per customer segment, quality lead over a generic model on your eval set, and gross retention.
A public reference point, with its rung 4 visible. Cursor began as a coding editor built on other companies' models and climbed by owning where developers work. It then built the flywheel this ladder describes: in September 2025 it described retraining its Tab suggestion model on live accept and reject signals, rolling out a new checkpoint every 1.5 to 2 hours, and reported the new model made 21% fewer suggestions with a 28% higher accept rate (Cursor, Tab RL, company-reported). In October 2025 it launched Composer, its own agent model trained on real software tasks (Cursor, Composer). The workflow came first and made the data possible; the proprietary model came last, trained on that data. SpaceX's agreement in June 2026 to acquire it at an implied $60 billion equity value, completed in August 2026 (SEC filings), shows how the market priced that position.
Follow the flow, one stage at a time.
Illustrated sequence; animation speed does not represent processing time.Read diagram description
Model access is shared by everyone, so defensibility is climbed rung by rung: prompt and UX polish, proprietary customer context, ownership of the workflow, an evaluation and data flywheel, and distribution and trust. Each rung is illustrated for a contract-review assistant used by in-house legal teams.
The model-swap test. If a competitor using a better model launched tomorrow, what would customers lose by switching to them? If the answer is "nothing", the product is still on rung 1.
What interviewers probe next
"Shouldn't we train our own model as the moat?" Only if evaluation shows a lasting quality advantage on your tasks that justifies the cost. Cursor's path is the pattern: years of owned workflow produced the usage data first, and its own models followed. Frontier labs improve general models quickly; a proprietary model's lead can vanish in a release cycle.
"Is proprietary data automatically a flywheel?" No. Raw usage logs rarely improve a product unless they are labelled, analysed and turned into evals and fixes.
How it's scored
| Signal | Weak | Hire | Strong Hire |
|---|---|---|---|
| Diagnosis | "Better prompts and UX" | Data as moat | Concedes model access is shared; places the product on a ladder |
| Strategy | Generic "network effects" | Some rungs | Specific investments per rung for the product, in order |
| Rigour | No measures | Retention | Measures for each rung plus the model-swap test |
Common mistakes
Claiming prompt engineering as a moat.
Assuming usage data is automatically a flywheel.
Betting on a proprietary model without evidence of a lasting lead.
See the moat ladder and data flywheels.