── ── AI & agents
Proprietary Data Moat
In the AI era, models and compute can be bought. What can become an advantage is data others cannot get that makes the product better the more it is used — and most claimed "data moats" fail that test. Casado & Lauten (a16z, 2019): most data defensibility "actually boil[s] down to scale effects," and "the cost of adding unique data…
Run Proprietary Data Moat on a real problem
Bring something you're actually deciding — free, in the browser.
How it works
1. Inventory the data. For each data class: source, volume, update frequency, and how it arises (product use / manual collection / purchased / public). 2. Uniqueness test. Could competitors or general-model vendors buy it, scrape it, synthesize it, or substitute public data? Gate: obtainable → not a moat; downgrade to "operating asset" → stop for this class. 3. Compounding test. Does new data keep making the product better, or do marginal gains saturate quickly? Gate: gains have saturated or cannot be measured → exclude from the moat. 4. Design the feedback loop. Use → label or outcome signal → improved model or rules → better experience → more use. Name the signal and how fast it flows back. Gate: the signal needs heavy manual work → estimate unit cost; redesign if too high. 5. Rights and consent. Do contracts and the privacy policy permit use for training and improvement? Is cross-customer use authorized? Which regimes apply (GDPR, CCPA, etc.)? Gate: no clear right → exclude the data until authorization is obtained. 6. Quality. Accuracy, coverage, freshness, bias; assign owners for sampling and cleaning. 7. Productize. Which customer-visible improvement does the data produce (accuracy, speed, personalization, benchmarks), and how is it measured? 8. Stop-rule: no data class passes steps 2, 3, and 5 together → concede there is no data moat and compete on distribution, workflow, or brand instead.
Output template: data class · source · uniqueness (obtainable?) · compounding (measurable lift?) · loop signal + cycle time · rights basis · quality metric · product advantage + measure · verdict (moat / operating asset / drop)
→ Method in Action: Everalbum and the FTC model-deletion order (2017–2021)
When to use it
- user says 'is our data a moat?', 'how do we build up proprietary data?', 'how do we set up a data flywheel?', 'investors keep asking about our data defensibility', 'data network effects', or must decide whether to invest in collecting, labeling, or partnering for data
When not to use it
When the decision is routine and reversible, applying a formal method costs more than it returns.
Worked example
Everalbum and the FTC Model-Deletion Order (2017–2021)
Everalbum, Inc. ran "Ever," a consumer photo-storage app. By the FTC's account, it had what looks on paper like a textbook data moat: millions of user-uploaded photos no competitor could scrape, a feature that had users label faces by name, and an enterprise facial-recognition business built on top. In 2021 the FTC's settlement required the company to delete not only the data but the models and algorithms built from it. The case shows why step 5 (rights) is a hard gate rather than a compliance…
Install this skill (free, MIT)
npx skills add deciqAI/knowledge-skillsUseful? Star the repo — stars help other builders find it.
Related mental models
Answer engines — featured snippets, "People Also Ask" boxes, voice assistants, AI answer panels — lift one passage and present it as the answer.
Teams usually decide how much to hand to an agent by watching a demo.
An agent that is live without logs, quality metrics or alerts fails quietly.
AI agents can research every prospect before writing to them, at close to the cost of a mass blast.