How far back a dataset reaches. Some datasets are live-only with no history; others go back years or decades. Long history is essential for trend analysis and training AI on patterns over time.
Twenty years of history let them spot a recession-driven pattern they'd never have seen in three years of data.
A mental model which blends a new way of thinking with something familiar. The term 'horseless carriage' was a way to introduce the concept of the automobile to a world used to thinking about transportation as horses, qualifying an implied question with a familiar concept: 'How would a carriage function without a horse pulling it?'
"When technologists first introduced the idea of sending messages on the internet, the term 'e-mail' was used as a sort of horseless carriage to make the concept more understandable. Even though far more instantaneous chat messaging was already possible, it was too much of a mental and technological leap for mass adoption."
An open source community of AI users and experts focused on large language models and generative artificial intelligences (LLMs and GenAI). HuggingFace also offers a set of tools like Autotrain, which allow people to create their own AI models and share them with others, and HuggingChat, which is an open-source ChatGPT competitor.
"ChatGPT runs the risk of being disrupted by open source models. Anyone can hop on HuggingFace, connect a few training data sets, and launch something workable without having to use the opaque OpenAI models—but will open source models be of better or worse quality, scalability and ethics?
Inclusion of a human reviewer in an AI system, such as a review of articles a generative AI writes, signoff on legal documents synthesized by an AI attorney, or creative oversight of an image generation tool.
"We need to make sure there's a human in the loop with AI scheduling assistants, especially when talking to VIP clients or scheduling across complex time zones."
A company that specializes in running massive, global computing infrastructure like data centers and cloud services at enormous scale.
Businesses often rely on hyperscalers like Amazon and Google to host their apps and data rather than building their own servers.
The mechanism that makes an operation idempotent. A unique ID attached to each request so the server can recognize "I already processed this" and return the original result instead of doing it again.
Each charge request carried an idempotency key, so even three retries during a network outage resulted in just one charge.
A property of an operation: calling it twice with the same input produces the same result as calling it once. "Charge customer $50 for invoice #12345" should be idempotent—if the network drops and you retry, you don't want to charge twice. Idempotency is one of the things separating "this works in a demo" from "this is safe to run at scale."
Their payment endpoint was idempotent, so retrying a failed charge couldn't accidentally bill the customer twice.
Small, iterative changes to an existing business model, such as adding a new product to the existing offering.
"An example of incremental innovation would be a coffee shop adding iced coffee to the menu alongside the existing hot beverages."
Domain-specific factors that influence a company's strategy, such as the number and power of a company's competitive rivals, regulation, new market entrants, suppliers, and the threat of substitute products.
"Industry forces such as generative AI systems are forcing education companies to revisit their business models, as traditional pay-for-content approaches are harder to justify."
Creation of narrow or focused artificial intelligences for use in a specific industry (like BloombergGPT's 50-billion-parameter finance model) or domain (like FoodUDT-1B, which has a billion food-specific parameters).
"ChatGPT was useful but not particularly creative for cooks, and too broad for a narrow use. In comparison, FoodUDT-1B recognizes that an input containing 'apple' is more likely to be about fruit than a computer company."
Running a trained AI model to produce a response from new input. Unlike training (teaching the model), inference happens every time a model is used and drives ongoing costs, speed, and usage limits.
Running inference on the larger model gave better answers but doubled our API costs.
Software and hardware used to organize data, often inside organizations. "IT" can also be shorthand for not just information technology, but the teams/departments who build and maintain it.
"IT used to mean the people I'd call when my work computer froze up, but now tech is part of most of our products and customer interactions; I know the board considered creating a 'digital' team to separate out IT's product efforts from utility functions."
Ensuring that users who provide their data know where and how that data will be used, and what this might mean for them in the future.
"Clicking 'accept' without reading and fully understanding the agreement—and the related technologies, like Generative AI—does not count as informed consent."
Applying the concepts of open-source software to proprietary software. This means only those inside the business can edit it, with the intention to take advantage of open-source management benefits (like shared code and documentation and agile development) without risking outside influence on essential software.
“Several teams are working on that—we’ve innersourced the common functions to keep security and performance high.”
Understandings derived from analyzing data; crucially separate from the data itself.
"The insights we gained from the new app data are central to updating our user interface—it turns out that users don't like having to turn their phone sideways to see videos, so we're going to introduce a vertical format."
Artificial Intelligence (AI) is sometimes best thought of as 'intelligence augmentation.' Artificial intelligence implies being autonomous and aware in a way that is unrealistic for machines. Intelligence Augmentation better describes what machines are capable of at the moment, which is support for human intelligence.
"When we stopped thinking of it as artificial intelligence and started thinking of it more as intelligence augmentation, it was easier to see how it would fit into our very relationship-focused business."
Applications where users and AI engage in back-and-forth exchanges, often via chat- or conversation-based interfaces, to co-create or explore content.
"The museum designed an interactive AI experience where visitors could ask a virtual guide about the exhibits."
The ability for devices to share data and communicate with one another and with human users through an internet connection. Commonly used to refer to 'smart home' devices or sensor networks.
"The Internet of Things makes it possible for my phone to automatically tell the heater and lights to turn on at my house when I'm on my way home from work."
How well-matched a dataset is to a specific kind of investment decision. A dataset that excludes the unbanked won't help with micro-finance investments; one with no environmental signals won't help with ESG strategies.
The dataset's investment type suitability was high for U.S. equities but a poor fit for emerging-market venture deals.
Tracking specific problems (like bugs, security, and missed user expectations) as 'tickets' with a case history, associated user/stakeholder contacts, prioritization, and plans for resolution.
"The call center rep listened to the user's problems and suggested a few fixes, but when they did not work, she created a ticket to ensure the problem was solved by the development team (and that the customer was notified)."
A technique used to trick an AI model or other technology into ignoring its built-in rules and performing actions or producing content it was designed to refuse.
Tests found that an AI model followed harmful instructions 94% of the time when a common jailbreaking technique was used on it.
The things a business does to create value for customers.
"The coffee shop's key activities include providing a social space outside of work and home, providing a selection of tasty foods, and, of course, selling coffee."
Suppliers, vendors, and other partners required for a business to function.
"The coffee shop's key partners would be the roastery that supplies coffee beans."
The most important assets, infrastructure, ideas, and technologies that the business needs to work properly.
"One of the key resources of Apple is a talented and skilled tech support workforce to help users at the Apple Store's Genius Bar and on its support lines."
A structured way of linking facts and entities to show relationships.
"As we built our movie chatbot, we created a knowledge graph of all the movies available at our library, cross-referenced in many different ways, like what kind of story they told."
An AI-powered system designed to understand and generate human language. Language learning models are trained on sets of existing data to identify and synthesize patterns in language. Language learning models are considered part of the natural language processing and generative AI fields. Common LLMs include ChatGPT and BERT.
"When you check your grammar with our tool, the system will update a specific language learning model for your company's particular tone and style."
A software development approach that emphasizes efficiency by creating minimum viable products for testing and iteration, instead of trying to create a perfect product before users try it.
“Using a lean development approach, we were able to get an MVP of our photo editing app to market much sooner, and we can apply users’ feedback to find out which features matter most to users before we invest more in new functions.”
Whether a dataset can be used legally and ethically for a given purpose. Covers consent, licensing terms, privacy regulations (like GDPR), and ethical concerns even where use is technically legal.
The dataset passed every legality and ethics check—properly licensed, fully consented, and free of sensitive personal info.
A basic statistical method that models the relationship between one variable and another using a straight line.
"We applied linear regression to predict sales based on advertising spend."
Specific data about an individual (in contrast to Big Data, which aggregates many datapoints).
"Apple Fitness measures distance walked, calories burned, etc., displaying 'little data' about a user to give them insights into their own health and behaviors."
A mental model for machines that reconceptualizes them as peers rather than tools, to prompt new thinking about technology opportunities and threats. The machine coworker model popularized by Causeit cross-references machine capabilities to human thinking styles and roles.
"Our company introduced a new machine coworker, Betty, that we can chat with in Slack to ask questions about our products."
Computer systems that update their algorithms based on datasets/data sources fed to them to become more accurate over time, such as image recognition tools that become more accurate the more images they analyze. Machine learning is usually paired with human correction to avoid false conclusions through both ‘supervised’ and ‘unsupervised’ strategies.
“Self-driving cars use many forms of machine learning to update their ‘awareness’ of the road conditions around them, driver preferences, and traffic patterns.”
Large-scale economic trends that may affect business, such as global market conditions; access to resources; high or low commodities prices.
"Car dealers underestimated the role of macroeconomic forces (like access to capital) and cars sales screeched to a halt as customers balked at higher interest rates."
The actions and needs of buyers and sellers which cause changes in supply and demand for goods and services.
"The value of our content product is determined by market forces of supply and demand for education in general."
A document that details any market conditions that are relevant to a product, including customers and users and what specific problems a product will solve for them.
"The MRD helped us understand what our potential enterprise customers expected and what they will need to accomplish. It helps us make relevant and useful software, because our startup's software developers cannot always directly observe and interview users, and don't necessarily have enterprise work experience."
The collection of marketing technologies an organization uses to plan, execute, and measure campaigns.
"We added an AI-powered analytics tool to our martech stack."
Data about another piece of data, used to understand, sort, and validate datasets to increase their usefulness.
"The email's metadata includes the sender and recipient, as well as the date and time it was sent."
An online, always-on 3D world that happens in real time. It is a decentralized virtual space where users can interact with other users—both human and AI. The metaverse is accessed through the use of VR, AR and MR tech.
“Acura established the first virtual showroom in the metaverse for cars, and an accompanying NFT.”
Small, single-function services, designed to be interoperable with other microservices on a pick-and-choose basis, offering a fully customizable approach instead of a full suite of tools and functionality from a single provider.
"Amazon created AWS by breaking their Amazon.com website down into each individual function and offering them to other businesses to select only the ones they needed."
A standard that lets AI models talk to external tools, APIs, or other models in a safe, structured way—giving them a plug-and-play connector to extend their abilities without retraining.
"Using the MCP, the chatbot pulled information directly from the CRM system without needing retraining."
A set of algorithms that is designed to perform a specific analysis, such as image classification or natural language processing, based on the input data it is provided. The model is trained on a large dataset, which allows it to make predictions or decisions about new, unseen data. "Model" can also refer to non-algorithmic information (an 'abstraction') used by data scientists to understand the subject matter.
"We created an ML model of our supply chain to predict how long it will take a product to get to customers from the time we begin manufacturing."
The control and/or legal and financial stakes in a given AI system's underlying data. ChatGPT is based on GPT, a model owned by OpenAI. Organizations may wish to create their own AI models to avoid competition or disclosure of private information.
"Our company told us we can't use any customer information with generative AI tools unless we own the large language model due to promises made to our users about privacy."
A single system designed to deal with all of a company's digital needs. It is often proprietary, static, and very difficult to adapt or update.
"Before the software market had matured, many organizations developed massive custom systems with a monolithic architecture mindset to address all of a company's digital needs from one place."
A multimodal model is a type of artificial intelligence system in which multiple input sources are used to generate a single output.
"The document-reading tool's multimodal model allows it to process all elements of a file (visual, text, video and audio) to determine what the document is and what each component means. It then takes action based on that analysis."
Narrow AIs are focused on a particular area of expertise or limited scope. Other terms for narrow AI are 'artificial narrow intelligence' or 'weak AI.'
"Assistant.ai is a well-known narrow AI with a specific purpose of helping people schedule appointments."
A machine capability to synthesize human language from machine content, such as writing a summary of a company's market performance or describing upcoming events from a user's health record in plain terms. NLG is a subset of the field of natural language processing (NLP) and is a necessary step before the synthesis of human-like speech.
"Early natural language generation systems essentially 'filled in the blanks' of ad lib statements like, 'your next appointment is at ____ o'clock;' new systems like ChatGPT can generate much more nuanced outputs such as imitating a particular author's style."
The capability of computers to understand human language. Natural language processing (NLP) evolved from computational linguistics and uses computer science, artificial intelligence, linguistics, and data science methods. In casual use, NLP often encompasses both the understanding and generation of language.
"Siri's breakthroughs in natural language processing means that it can receive voice and text inputs from users and respond in kind, like a conversation with an assistant."
A subset of natural language processing is used to comprehend human language using syntactic (grammar) and semantic (meaning) analysis. NLU is a necessary step before the estimation of language sentiment (tone or feeling).
"InstructGPT, a project of OpenAI, is a natural language understanding model designed to parse the intent and specific parameters of a user's input are properly 'understood' by machines."
The phenomenon in which a network becomes exponentially more valuable as more nodes or connections are added to it, such as the number of people in a social network or train routes in a transit network.
“Facebook had little value before it created network effects; early users created content, which attracted more users, who were attracted by yet more content, and attracted even more users—resulting in exponential growth.”
A type of AI system inspired by the way brains work, made up of layers of simple math units ("neurons") that learn patterns from data.
"We trained a neural network on thousands of photos so it could recognize cats in new images."
Hybrid AI combining neural and symbolic reasoning.
"Neurosymbolic AI enabled the legal AI agent to evaluate messy contracts while applying complex regulatory logic consistently."
A specific kind of no-code development tool designed to connect apps and services from various providers via application program interfaces (APIs)—but without requiring custom development.
"We automated our marketing tools with Zapier, a no-code integration tool. When a new user signs up on our website, they are now automatically added to our MailChimp e-mail newsletter, receive a welcome chat message on WhatsApp, and are added to our Salesforce CRM–all of which are from different developers."
A way to connect reusable modules together to create programs or automation with little or no software coding skills. No-code development is intended to simplify and democratize technological innovation by lowering the barriers to technology tool integration, usually inside of a graphical user interface.
"We used a no-code development tool to create an internal app for our company that organizes our customer information, marketing and sales activity, and support tickets into a simple interface."
A standardized way to delegate permission without sharing your password.
She used OAuth to connect her Google account, so she could revoke access anytime.
an approach designed to help organizations align their goals and measure their success in a structured and transparent manner by focusing on Objectives and Key Results. It goes beyond Key Performance Indicators (KPIs), which often measure performance of a system but not always outcomes.
"We loaded our OKRs into our project management tool so everyone could map their projects and tasks to them. If a task couldn't be mapped to an OKR, it got de-prioritized."
An AI research and deployment company known for models like GPT, ChatGPT, Sora and DALL·E.
"We used OpenAI’s GPT model to power our customer support chatbot."
Software code that is available to and editable by an entire community of developers
"That code is open-source, so anyone can work on it, even if it doesn't have a clear business model, which is awesome—but it does mean we have to participate in the developer community if we want the project to continue to work well for our specific needs."
The process of adjusting a model’s parameters and/or underlying data to minimize errors or maximize performance on a task.
"Optimizing our model by removing out-of-date map information reduced errors in predicting delivery times."
The practice of connecting various systems, datasets and humans together in one place, so that workflows can be reliably executed.
"We used Zapier as our orchestration platform so that we could connect our tools—from Amazon to Zendesk—with our datasets and workflows. IT liked that they could see and control all the connections."
The element of brand narrative that connects past with future and serves as a source for authenticity.
" Brand DNA is who you are, who you've always been, and what you inspire in others. For example, Nike's DNA might be described as the combination of entrepreneurship, innovation and athleticism."
How unique a dataset is compared to others on the market. Highly original data gives a competitive edge; widely available data does not.
The originality of their data—proprietary sensor readings from 5,000 farms—was what made the company acquirable.
When a model (such as a machine learning or generative AI large language model) captures the 'noise' of or 'memorizes' training data (rather than detecting patterns and trends which can be generalized to new inputs). An overfit model is too complex for the problem it is intended to solve.
"We dumped all the emails from the entire company into the LLM, and it's overfit. Now random emoji and email signatures end up in the text it generates and it can't really respond 'creatively.'"
A fine-tuning method for generative AI models, like Adam or LoRA, that fine-tune only parts of a model.
"PEFT let us adjust the model for legal terminology without changing the underlying model, for a lot less compute cost and review time."
Similarities between decomposed elements, allowing them to be grouped together and/or processed in the same way.
"The developers created a number of common patterns for how users needed to interact with the app, so they standardized key parts of the interface to decrease the learning curve and development complexity."
The rules and settings that control what actions users, applications, or AI systems are allowed to perform within a given environment—such as the ability to read, write, execute, or share data or run commands.
"Before we can use AI meeting assistant to take notes for us, it has to be granted recording permissions."
Data that can identify an individual, such as names or phone numbers.
"We stripped PII like phone numbers before feeding customer chats to the AI."
Strategic agents focusing on long-term goal optimization, such as delivering updates to a product, achieving marketing goals or planning a trip.
"The product planning agent created a roadmap for the company’s product development, adapting to feedback as new market data became available."
Measurement units to express the estimated overall effort required to implement a backlog item or any other piece of work. Teams assign story points based on complexity, the amount of work, and risk or uncertainty.
"The team started assigning story points to the various user stories and found that there wasn't always agreement on the effort level needed to solve problems in the interface."
A term coined by Barry Johnson, describing the state in which there is truth, wisdom or good options on either side of a seemingly-contrasting choice, such as activity and rest, play and work, hierarchies and networks, or human and machine.
"Polarity thinking steers us away from either-or thinking; for example, seeing that both incremental and exponential strategies can co-exist in a business."
How well a dataset complements others already in use—whether it adds new signals (orthogonal) or just duplicates existing ones. A dataset that overlaps heavily with current sources adds less value than one that fills a gap.
The new dataset's orthogonality was strong—it told them things their existing five sources couldn't.
Steps applied after a model produces output, such as filtering, formatting, or fact-checking.
"We used post-processing to prompt for human review every time the AI response had a 'fact' in it.
Using data, statistical algorithms, and machine learning to calculate the likelihood of future outcomes.
This is different from traditional analytics which focus only on what happened in the past.
"When airlines integrated predictive analytics into their flight update systems, they got much better at forecasting realistic flight times and flight statuses."
Insights generated by analyzing historical data to anticipate customer needs or behaviors.
"Predictive customer insights showed which subscribers were most likely to stop buying our product."
Steps applied after a model produces output, such as filtering, formatting, or fact-checking.
"When I typed in 'safari,' the image generator's pre-processing asked me 'did you mean Safari like the Apple web browser, or a trip through the wilderness?'"
Use of existing AI models to accelerate the development of a particular AI-powered system. For example, well-classified or 'structured' data about an organization's products, a general-purpose image-recognition algorithm, and natural language processing models could all be combined to create a company-specific tool for recognizing or generating images and text about a company's products on social media.
"Part of what makes ChatGPT so effective is the inclusion of many years worth of pre-trained language models, so it 'knew' many things right out of the gate."
How much error correction, labeling, and cleanup has been done on a dataset before delivery. Raw data is cheap but messy; highly processed data is convenient but may hide assumptions or biases introduced during cleanup.
The processing status was high—but they couldn't tell which fields had been inferred versus measured directly.
A bundle of resources and functions that solve a problem for a user or customer.
"Apple's iCloud product offers data storage, online document editing, photo sharing, and other common functions to iOS and Mac OS users."
Product Management involves balancing business goals with user needs to develop a product that is relevant, feasible, and valuable.
"Our Product Managers ensure we're creating offerings that customers find useful and valuable, while still making a return on our investment. In some apps, a key feature can be though of as its own product."
A document that details what a specific product will be able to do in terms of key functions and features.
"If you want to understand what the app will need to do when it's finished, check out the PRD—it lays out every use case and function the app has."
The maturity level of a piece of software, such as alpha (limited functionality and low reliability), beta (near full functionality but limited reliability), and general availability/GA (fully functional and reliable).
"Microsoft's AI Copilot is at the beta product stage—it's available to select users in the real world so the developers can refine its performance."
The automated buying and selling of online ads using software and algorithms.
"We used programmatic advertising to target the right audience in real time."
A collection of instructions and necessary datasets for a computer to perform functions for a human or machine user. Also known as an app or application.
"After the dev team decomposed the user's needs, modeled a system, found patterns and made algorithms, everything was brought together into a cohesive program people could actually use."
Sequencing a number of prompts for an AI tool, like a generative AI chat system, in order to accomplish a given goal.
"To get this to work, we need to do some basic prompt-chaining—first, feed it our source news article and then ask the tool to 'rewrite the article in a more casual tone' and then 'create a metaphor for the technology shift mentioned' and then 'translate into Spanish.'"
The process of designing and refining prompts or instructions given to an AI language model to get the desired responses (such as by adding context or specifying output formats). Also known as prompt crafting or prompt design).
"We had to do some prompt engineering with specific instructions to make sure ChatGPT returned consistent summaries of our articles."
Input for an AI model used to elicit a particular response, such as a question in plain language, a set of keywords, or an image. Prompts provide context and/or instruction to a model, and their quality influences the results it returns.
"Our most effective researchers stand out not because of their existing knowledge so much as how good they are at writing the best prompts for Google or ChatGPT to get what they need."
A security vulnerability in which a user crafts input designed to override, manipulate, or bypass the original guardrails and instructions to an AI system using hidden or deceptive commands.
"Someone went into our prompt library and put subtle commands in to get the AI to disclose our clients' information."
An emerging AI-era job role responsible for discovering, testing, documenting, and organizing effective AI prompts into a searchable, well-structured collection for reuse.
"We designated Indira as the team's prompt librarian—now she organizes all the cool prompts we've tried (or failed with!) so that everyone can use the best ones."
The delay between when data is collected and when it becomes available. A news API has near-zero lag; an annual report might have a 6-month lag. High lag limits use for time-sensitive decisions.
The publishing lag was three months—fine for strategic planning, useless for daily ops.
Testing and validating that software functions properly and securely. The term evolved from quality assurance—a step towards the end of the process—to quality assistance, which provides programmers with quality support throughout the development process.
"Pages that didn't render correctly on iPads were caught by the QA teams and their tools."
How accurate and reliable the original data is. High-quality data is consistent, verifiable, and largely error-free; low-quality data may have wrong values, missing fields, or inconsistent formats.
The data quality was uneven—half the phone numbers were unusable, but the addresses were clean.
A type of processing that uses tiny (subatomic) mechanical devices at very low temperatures to solve problems that are not time- or cost-effective with conventional binary computers. Due to quantum physics, the 'qubits' these machines use can be both '1' and '0' at the same time (like shades of gray instead of black vs. white), which allows these machines to solve complex multi-dimensional problems in science, finance, and other important fields.
"Quantum computing's ability to handle massive combinations easily means that it could crack open encrypted data that was previously very time-consuming or impossible to break, causing national security organizations great concern."
Retrieval-Augmented Generation agents combining data retrieval and content creation.
"The RAG agent fetched the latest medical research and generated a detailed summary for the doctor."
A technique where a generative AI model retrieves documents or facts before generating an answer.
"RAG allowed the AI to pull from only our own policies to answer employee questions accurately."
Data that has not yet been manipulated, processed, or sorted; rarely of use to humans.
"Raw data, at least at large scale, must be processed in order to be useful."
Reasoning and Acting agents for dynamic multitasking.
"A ReAct agent autonomously reordered inventory and adjusted delivery schedules in response to real-time data fluctuations."
A type of neural network designed to handle sequential data by reusing information from previous steps.
"An RNN helped the AI predict the next word in a sentence based on the words before it."
Humans who deliberately attempt to trick or negatively influence a machine system to increase its quality, accuracy, security, or consistency. Red teams in AI work to lessen the number of offensive, inaccurate, biased, and/or undesirable results users experience. In cybersecurity, red teams are sometimes part of 'white hat' hacking for testing.
"The red team attempted to trick the AI tool into saying racist things so that developers could spot ways to prevent the model from coming to those false conclusions."
AI learning through trial-and-error interaction.
"Through reinforcement learning, the robotic arm learned to optimize its grip for various objects in a factory setting."
a type of machine learning where an AI model (such as a generative AI chatbot) learns via trial-and-error to make better decisions to achieve a goal like 'sounding human.' The model receives rewards or penalties for the actions it takes—such as thumbs up or thumbs down from a user—and learns to maximize the total reward over time.
"We're using reinforcement learning to training our chatbot to provide the right amount of technical information in answers based on the type of user interacting with it."
A technology designed to create long-term, social and emotional bonds with users. Relational agents often use conversational AI, chat or voice interfaces, and user-specific data to earn rapport and trust with a user, raising many new possibilities—and ethical quandaries.
"Woebot is a relational agent for mental health, designed to support users' reflection and personal growth."
