TL;DR: Model access is shared, so it is not a moat. Defensibility comes in five rungs, from weakest to strongest: prompt and UX polish, proprietary context (customers' data connected to the product), workflow ownership, an evaluation and data flywheel, and distribution and trust. Most AI products start on rung one; strategy is the deliberate climb, measured on each rung, with a simple test: what would customers lose by switching to a competitor using a better model?
Why a ladder
AI products are unusually easy to start and unusually easy to copy. Capabilities that look like breakthroughs (summarisation, drafting, image generation, code completion) become available to every competitor through the same model APIs, often within weeks. A strategy built on the capability alone has a short life.
The ladder gives a structured way to ask what, beyond the model, makes a product hard to displace, and where to invest next.
The five rungs
| Rung | What it is | Defends against | Fragility |
|---|---|---|---|
| 1. Prompt and UX | Good prompts, polished interface, fast iteration | Slow competitors | Copied in weeks |
| 2. Proprietary context | Connectors to customers' documents, systems, history; retrieval over their data | Generic assistants | Moderate: competitors can build connectors, but switching means reconnecting |
| 3. Workflow ownership | The product is where the work happens: collaboration, approvals, records, handoffs | Point tools | Strong: switching changes how teams work |
| 4. Eval and data flywheel | Domain failure taxonomies, labelled outcomes, evals that improve quality with use | New entrants | Strong if data is proprietary and actually used |
| 5. Distribution and trust | Brand, enterprise relationships, certifications, embedded procurement | Almost everyone | Strongest and slowest to build |
Read diagram description
The moat ladder as a staircase of five rungs above shared model access, which protects no one: prompt and UX, copied in weeks; proprietary context, where switching means reconnecting; workflow ownership, where switching changes how teams work; eval and data flywheel, strong if the data is used; and distribution and trust, strongest and slowest to build. The model-swap test asks what customers would lose by switching to a rival on a better model.
Climbing, with an example
A meeting-notes product:
- Rung 1: good summaries from transcripts. Easy to copy; every video platform can add it.
- Rung 2: connects to the CRM, project tools and documents, so notes link to the right deal or project and know the context of past meetings.
- Rung 3: follow-ups, action-item tracking, sharing and approvals happen inside the product; it becomes the team's system for decisions.
- Rung 4: every edit to a summary, every accepted or rejected action item, is a label. Domain evals (sales calls, interviews, board meetings) improve the product in ways a newcomer cannot replicate quickly.
- Rung 5: security certifications, enterprise contracts, references; it is the approved tool.
Measuring the climb
| Rung | Measure |
|---|---|
| 2 | Share of usage relying on connected customer data |
| 3 | Share of workflow steps completed in the product; collaboration per account |
| 4 | Growth of labelled examples per segment; quality lead over a generic model on your eval set |
| 5 | Enterprise share of revenue; gross retention; sales cycle win rates |
The model-swap test
Imagine a competitor launches tomorrow using a better model. What would your customers lose by switching? If the answer is "nothing", you are on rung one regardless of how good the product looks. If the answer is "our connected data, our team's workflow, accuracy on our documents, and an approved vendor", you have climbed.
Common misreadings
- "Our data is the moat." Only if it is used: labelled, turned into evals and improvements. Raw logs are storage.
- "Our own model is the moat." Only if evaluation shows a lasting lead on your tasks; frontier providers close gaps quickly.
- "Network effects." Rare in AI applications unless users' contributions directly improve the product for other users.
In the interview
Strategy questions about AI products ("what's your moat?", "a competitor launched the same feature") reward this structure. Strong answers concede that model access is shared, place the product on the ladder, propose specific investments for the next rungs, and suggest measures. The model-swap test is a memorable way to make the point concrete.