Fifteen terms, defined the way I use them when shipping AI products with non-technical founders and cross-functional teams. I add terms as the work turns them up.
Add his coinages as entries using existing site language, e.g.: “Abstention — the model saying ‘I’m not sure about this one’ instead of bluffing. Pitched too eager, users learn to skim past it; too shy, one fluent wrong claim undoes earned trust. The threshold has to fail toward humility — the cheaper mistake.”
- ML Explainability
- Showing users why an algorithm made a decision, not just what it decided. Example: instead of “Risk: High,” explain “Risk: High — due to 6 failed trades in the last 18 months and 2x normal deal size.”
- Feature Importance
- Which inputs to the model drove the prediction most? Tools like SHAP and LIME show which factors influenced the decision. Non-technical users care: “This recommendation is based on X, Y, Z.”
- Distribution Shift
- When a model is trained on one type of data but deployed on a different type, it fails silently. Example: model trained on 90-second trades suddenly sees trades that take 30 minutes. It doesn’t know this is different.
- Human-in-the-Loop
- AI suggests, humans decide. Rather than fully automating a decision, the system shows the recommendation and lets users confirm, edit, or override. Builds trust and preserves human judgment.
- Confidence Layer
- The interface that communicates uncertainty. “92% confident” paired with “what data drove this” and “how would the decision change if X changed.” The difference between a user trusting and rejecting the same recommendation.
Product & Design
- WCAG AA
- Web Content Accessibility Guidelines Level AA is the practical baseline standard for web accessibility. It ensures your product works for users with disabilities: colour-blind users, keyboard-only users, screen reader users, and users with low vision.
- Cognitive Friction
- Unnecessary mental effort. “What does this button do?” or “Why is this here?” are friction points. Remove a confusing step and adoption rises.
- Time-to-Value (TTV)
- How long before a new user feels “Oh, I see why this is useful.” Shorter TTV = faster adoption. If TTV is 1 hour but users abandon at 15 minutes, you have a design problem.
- Product-Led Growth (PLG)
- No expensive sales team needed — the product itself drives adoption. Figma and Slack did this: so useful, users tell other users, virality follows.
- User Experience (UX)
- Not just beauty — clarity, speed, and trust. On AI products it is the layer where a right answer becomes an acted-on one.
Business & Go-to-Market
- Customer Acquisition Cost (CAC)
- Total marketing spend ÷ new customers acquired. If CAC is $1,000 but the customer only generates $500 in lifetime value, your unit economics are broken.
- Lifetime Value (LTV)
- If a customer pays $100/month for 3 years, LTV ≈ $3,600. Healthy businesses have LTV > 3× CAC. Below that, you’re losing money on every customer.
- Painkiller vs. Vitamin
- A painkiller is a must-have: users are already in enough pain to pay for relief. A vitamin is a nice-to-have, so adoption is slower and you have to manufacture the urgency. Painkillers are easier to sell, which is why I build those first.
- Product-Market Fit (PMF)
- When your product solves a real problem for a specific market and users want it badly enough to tell others. Without PMF, growth is unsustainable. With it, growth becomes effortless.
- Programmatic Advertising
- Instead of humans buying individual ad placements, algorithms bid on impressions in real time. It’s faster, cheaper, and more precise than manual buying — and, because it’s automated at scale, far easier to get wrong.
Last updated September 2026. I add terms as the work turns them up.