Glossary of Terms

A method to compare two versions of a prompt or AI system to determine which produces better results.

Autonomous systems that can interact with users and systems to achieve goals.

The gradual degradation of an AI model's performance over time as real-world data patterns change from those in its training data.

A benchmark test designed to measure an AI system's general intelligence and reasoning capabilities across various domains.

The process of considering something independently from its physical or concrete details.

The percentage of customers who successfully complete a certain milestone in your onboarding process.

The percentage of customers who move to the latest release or model. This is an indication of how attractive new capabilities are to the existing customer base or the difficulty associated with upgrading.

Having the capacity to act independently and make choices.

AI systems designed to exhibit agency or autonomous behavior.

Capabilities that allow AI systems to act independently and make decisions.

AI systems that perform specific tasks or make decisions.

A trial phase of a product’s launch where company employees test a release candidate in a customer environment. (See also "Beta Test".)

Attributing human characteristics or behaviors to non-human entities, animals, or objects.

A technique for sharing information between multiple products. For example, Zapier leverages APIs to connect different products. Zapier can update an online spreadsheet whenever a new subscriber signs up on your website.

Functioning independently without external control.

The average revenue generated from each customer contract, which can help assess the effectiveness of pricing and packaging strategies.

The average revenue generated per account (customer) within a specific time frame. For example, monthly.

A product backlog is a list of the new features, changes to existing features, bug fixes, infrastructure changes or other activities that a team may deliver in order to achieve a specific outcome. The backlog may contain items that will never be built because they never achieve a high priority—which explains why “let me put that on the backlog” is not a commitment to deliver. The product backlog is the single authoritative source for things that a team works on.

A trial phase of a product’s launch when customer employees test a release candidate in a customer environment. This is usually offered to a select group of early adopter users (and never sales prospects) and the product may still have bugs in it. (See also "Alpha test".)

Systematic errors or unfair preferences in AI systems that result from imbalanced training data or algorithmic design.

The presence of systematic and unfair discrimination in AI predictions or outputs, often resulting from biased training data.

Techniques employed to reduce or eliminate unfair preferences or discriminatory patterns in AI systems.

The percentage of visitors to a particular website who navigate away from the site after viewing only one page.

The deltas in various steps of the buyer’s journey. For example, the percentage of buyers that move from awareness to interest to decision to action.

The archetype of people in similar roles who share common problems, challenges, and buying preferences.

AI models designed to break down complex problems into intermediate reasoning steps, improving logical consistency and transparency.

The success of each acquisition channel in driving traffic, sign-ups, or purchases.

The percentage of users who stop using the product within a specific timeframe, such as monthly or annually.

A situation where some classes in a dataset have significantly fewer samples than others, impacting model performance.

An AI assistant developed by Anthropic designed with a focus on helpfulness, harmlessness, and honesty.

The percentage of people visiting a web page who follow a link to a particular advertisement.

Typically, an individual or organization who has purchased your product.

A statistical method used to group similar data points.

A prompt technique that asks an AI model to check its own reasoning for logical consistency and errors.

The challenge of building effective AI models with limited or no historical data.

The number of experiments performed and completed, regardless of the results, is an indication of continued validation efforts during the creation process.

An analytical method used by AI to determine how people value different features of a product or service by analyzing trade-offs.

An AI-driven academic search engine that extracts findings from scientific literature.

The amount of text an AI model can consider at once when generating responses, essentially its working memory.

The percentage of users who take a desired action, like signing up for a newsletter.

A term describing AI tools that assist users with tasks.

The total cost of acquiring a new customer through marketing and sales efforts.

The term 'customer' has two common meanings. In some contexts, it refers specifically to those who have purchased the product, also known as clients. For others, customer refers to the market full of existing and potential clients—a synonym for market.

The number of customers acquired in a timeframe such as monthly or annually.

Measures the ease with which customers can interact with your product or service. It is often determined by asking users to rate the effort required to accomplish a task or resolve an issue on a scale from very low to very high effort. A lower CES indicates a more user-friendly product, which can lead to higher user satisfaction and loyalty.

The overall impression and perception a customer has of a company throughout their entire relationship — from awareness, purchase, and onboarding to product usage, support interactions, and post-purchase engagement.

A composite metric that combines multiple indicators, such as usage, satisfaction, and support interactions, to provide an overall assessment of the customer’s relationship with the product.

The total revenue a customer generates during their relationship with the product.

The difference between the lifetime value of a customer (LTV) and the cost of acquiring them (CAC).

The percentage of customers who refer others to the product.

The percentage of users who renew their subscription or continue using the product after their initial contract period.

The percentage of users who continue using the product after a specific timeframe, such as monthly or quarterly.

A quantitative measure of customer satisfaction gathered through surveys, ratings, or reviews. Common techniques include simple “thumbs up,” star ratings, or numeric scales such as 1-5, or 0-10 as used in Net Promoter Score (NPS).

A group of buyers. Customer segmentation is the process of dividing customers into groups based on common characteristics so companies can market to each group effectively and appropriately. Also known as market segment.

Stands for daily active users or monthly active users. DAU and MAU are metrics used to determine how many users are using a product on a regular basis. Generally, companies want high DAU/MAU metrics.

The number of unique users who engage with the product daily.

When the statistical properties of data change over time, causing model performance to degrade.

The field of analyzing data to extract insights.

A subset of machine learning involving large neural networks with multiple layers, often used in image and speech recognition.

An AI search engine that generates answers using language models.

The number of new features deployed to the customer case for each release.

Deprecate refers to a program or feature set that is no longer supported or maintained by its developer. At some point, all software becomes deprecated as new versions are released and older ones are no longer supported.

A human-centered approach to innovation that integrates the needs of the customers (based on human-centered design), along with technical feasibility, viability, desirability, and business success requirements.

These users use the product as soon as it is available. These are often the product’s most dedicated fans and are likely to be beta testers.

Explain like I'm five; simplify complex concepts for easy understanding.

An AI tool for automating research workflows and literature reviews.

A technique for representing data, such as text or images, in a numerical format that machine learning models can process.

Large bodies of work that can be broken down into several smaller stories.

The practice of ensuring AI systems are fair, transparent, and aligned with societal values.

Additional revenue generated from existing customers, through upsells, cross-sells, or add-on purchases.

The ability to describe the steps an AI model took to arrive at its predictions or outputs.

An AI tool for real-time transcription and meeting summaries.

One capability or function within a product, designed to address a specific problem or requirement.

The frequency and depth of usage for specific product features.

"Feature factory" describes a business focused on building features rather than solving problems for customers. It is generally a negative term or even an insult.

A process of refining prompts based on output to improve results.

A training method where an AI model learns effectively from only a few labeled examples.

The process of refining a pre-trained AI model on specific data to improve performance for a particular task.

An AI tool for recording, transcribing, and collaborating on conversations.

An AI model designed to produce rapid yet effective insights.

A point in a workflow where users experience frustration or unnecessary work due to an inadequate or inefficient automation. This may be described by an epic or a user story depending on the complexity of the situation.

Google's multimodal AI model designed to work across different types of content including text, code, audio, and images.

AI models capable of creating content such as text, images, music, and videos.

AI models trained on large text data to generate text.

A go-to-market strategy is a blueprint for delivering the product from the company into the users’ hands.

An optimization algorithm used to train machine learning models by minimizing the error between predictions and actual results.

An AI model developed by Elon Musk's x.AI.

Safety mechanisms implemented in AI systems to prevent harmful, unethical, or dangerous outputs.

When an AI system generates false or misleading information that appears confident and accurate.

An approach that combines multiple AI reasoning methods to achieve better problem-solving capabilities.

Settings that control how a machine learning model is trained and optimized.

Proposed explanations or theories that can be tested through further investigation.

An instance of a pop-up or other Web advertisement being seen on computer users’ screens.

To instill an idea, attitude, or habit by persistent instruction.

The process of an AI model applying its learned knowledge to make predictions or generate outputs for new inputs.

Collections of epics that drive toward a common goal. (Contrast with themes). Initiatives break into epics which break down into stories.

Having the same structure or form despite potentially different origins.

A platform for data science competitions and dataset sharing that's frequently used to develop and test AI models.

A framework to help support prioritizing capabilities based on the degree to which they are likely to satisfy customers. Product teams can weigh a high-satisfaction feature against its costs to implement to determine whether adding it to the roadmap is a strategically sound decision.

A quantifiable measure of a company’s objective (e.g. KPI for growth might be an increased number of DAUs).