AI comes with a lot of jargon. Here's what the most common words and phrases actually mean—no computer science degree required. Search a term or browse by category.
Software that does things we used to think needed a human brain.
A broad term for computer systems that can do tasks that normally require human intelligence—like understanding language, recognizing images, or making decisions. Most 'AI' you hear about today is really one specific flavor called machine learning.
Think of 'AI' like the word 'vehicle'—it covers everything from a bicycle to a jet. It's a category, not one specific thing.
Teaching software by showing it examples instead of writing rules.
Instead of a programmer writing out every rule, the software learns patterns from lots of examples. Show it thousands of photos of cats and it learns what a cat looks like—without anyone defining 'cat' line by line.
It's how a child learns a dog is a dog: not from a definition, but from seeing many dogs and being corrected.
A step-by-step recipe a computer follows.
A set of instructions for solving a problem or completing a task. In AI, algorithms are the procedures that let software learn from data and make predictions.
Like a cooking recipe: a fixed set of steps that turns ingredients (data) into a result.
The information AI learns from and works with.
Everything from documents and emails to spreadsheets, images, and customer records. AI is only as good as the data it learns from—clean, relevant data leads to better results.
Data is the fuel. A great engine still won't go far on bad fuel.
A math system loosely inspired by how brain cells connect.
A way of structuring machine learning that passes information through many connected layers, each one picking up on more complex patterns. It's the engine behind most modern AI.
Like a bucket brigade passing water down a line—each person (layer) does a small part of the job.
Machine learning using very large neural networks.
A powerful branch of machine learning that uses neural networks with many layers. It's what made recent leaps in language, image, and speech AI possible.
Same input, same answer every time—vs. a smart best-guess that can vary.
Traditional software is deterministic: give it the same input and it returns the exact same output, every single time. Most AI is probabilistic: it produces a likely answer based on patterns, so the same question can get slightly different responses. That's why AI feels flexible—and why it needs review for anything critical.
A calculator is deterministic—2+2 is always 4. Asking a knowledgeable friend to 'summarize this' is probabilistic—you'll get a great answer, but a bit different each time.
The trained 'brain' that produces AI results.
The finished product after an AI has learned from data. When you use a tool like ChatGPT, you're using a model. Different models have different strengths, sizes, and costs.
The recipe (algorithm) is the method; the model is the trained chef who has made the dish a million times.
AI trained on huge amounts of text to understand and write language.
The technology behind chatbots like ChatGPT, Claude, and Gemini. It has read enormous amounts of text and learned to predict and generate human-like writing, answer questions, summarize, and more.
Like an extremely well-read assistant who has skimmed much of the internet and can talk about almost anything.
The learning phase where AI studies examples.
The process of feeding a model large amounts of data so it can learn patterns. Training is expensive and done ahead of time; using the finished model afterward is cheap and fast.
Training is going to school for years. Using the model is asking the graduate a quick question.
The internal 'dials' a model tunes while learning.
The adjustable values inside a model that get fine-tuned during training. More parameters can mean a more capable model—but also a more expensive one. You'll see numbers like '70 billion parameters.'
Like the millions of tiny knobs on a mixing board that together produce the final sound.
Giving a general model extra training for your specific needs.
Taking an existing model and training it a bit more on your own examples so it gets better at a specialized task—like your industry's tone or a particular kind of document.
Hiring a skilled generalist, then training them on how your company specifically does things.
AI that handles more than just text—images, audio, video too.
A model that can understand and produce several types of content. A multimodal model might read a photo, listen to audio, and answer in text, all at once.
Like a person who can read, look, and listen—rather than one who can only read.
The small chunks of text AI reads and bills by.
AI breaks text into pieces called tokens (roughly a word or part of a word). Most AI tools charge based on how many tokens go in and come out, so tokens are how you measure and budget usage.
Like billing a phone call by the minute—tokens are the 'minutes' of AI usage.
A big, general-purpose model that everything else is built on.
A large model trained on broad data that serves as a starting point for many uses. Companies build specialized tools on top of it instead of training their own from scratch. The most advanced ones are sometimes called 'frontier' models.
Like a general-purpose engine that gets dropped into many different vehicles.
The moment AI actually gives you an answer.
Using a trained model to produce a result—the 'thinking' that happens each time you ask a question. Training happens once and costs a lot; inference happens every time you use the model, which is what you pay for ongoing.
Training is the years of medical school; inference is the doctor answering one patient's question.
A model that 'thinks through' a problem step by step before answering.
A newer kind of model that works through harder problems in stages rather than blurting out the first answer. It's slower and costs more, but is better at math, logic, and complex multi-step tasks.
Like the difference between a snap answer and someone who shows their work on scratch paper first.
Turning words and documents into numbers that capture meaning.
A way of converting text into lists of numbers so a computer can tell what's similar in meaning. It's the quiet workhorse behind smart search and behind AI finding the right document to answer your question.
Like giving every idea a set of GPS coordinates, so related ideas end up near each other on a map.
The instruction or question you give an AI.
Whatever you type or say to an AI to get a response. A clear, specific prompt gets a better answer—vague questions get vague answers.
Like the brief you'd give a new assistant: the clearer the request, the better the result.
The skill of writing prompts that get great results.
Learning how to phrase requests, add context, and give examples so the AI reliably does what you want. It's less about coding and more about clear communication.
Like learning to ask a genie for exactly what you mean—so you don't get a literal, unhelpful wish.
How much the AI can 'keep in mind' at once.
The amount of text an AI can consider in a single conversation—your question plus any documents you share. Once you exceed it, the AI starts 'forgetting' the earliest parts.
Like short-term memory: you can only juggle so much at once before earlier details slip away.
When AI confidently makes something up.
AI sometimes states false information as if it were true—an invented statistic, a fake citation, a wrong date. It's the single biggest reason to review AI output for anything important.
Like a confident student who doesn't know an answer and guesses convincingly rather than admitting it.
Hidden instructions that set the AI's role and rules.
Behind-the-scenes guidance that tells the AI how to behave—its tone, its job, what it should and shouldn't do—before you ever type anything. Businesses use it to keep AI on-brand and on-task.
Like the training manual and dress code you give staff before they meet a customer.
A setting that controls how creative vs. predictable AI answers are.
A dial that adjusts how much the AI 'plays it safe.' Low temperature gives focused, consistent, predictable answers—good for facts and data. High temperature gives more varied, creative answers—good for brainstorming and writing.
Like a spice level: low is reliable and plain, high is adventurous but less predictable.
Asking AI to show its reasoning step by step.
A prompting technique where you ask the AI to work through a problem out loud rather than jumping to an answer. It noticeably improves accuracy on anything involving logic, math, or multiple steps.
Like a math teacher saying 'show your work'—the steps lead to fewer careless mistakes.
Asking with no examples vs. giving a few to copy.
Zero-shot means you just ask and hope the AI figures it out. Few-shot means you include a couple of examples of what a good answer looks like—which usually gets far better, more consistent results.
Zero-shot is 'write me a product description.' Few-shot is showing two you like first, then saying 'now do this one.'
Letting software handle repetitive work for you.
Setting up software to do routine tasks automatically—sending follow-ups, moving data between systems, generating reports—so people don't have to do them by hand.
Like a dishwasher: you set it up once, and it handles the tedious job on its own.
AI that can take actions and complete multi-step tasks, not just chat.
Instead of only answering questions, an agent can actually do things—look something up, use a tool, send an email, then decide what to do next—working through a goal step by step.
A chatbot answers your question; an agent is an assistant who goes and gets the whole task done.
Describes AI that acts on its own to reach a goal, not just responds.
A buzzword for AI that behaves like an agent—it plans, makes decisions, uses tools, and takes multiple steps toward a goal with little hand-holding. 'Agentic' is the adjective; the more of these traits a system has, the more agentic it is.
The difference between an intern who waits for each instruction and one you can hand a goal and trust to run with it.
Able to operate without a person driving every step.
Describes how much a system can do on its own. Fully autonomous AI acts without human input; most business AI is semi-autonomous, doing the work but checking in with a person at key moments.
Like self-driving cars: some need hands on the wheel, some don't—there's a spectrum, not a switch.
Coordinating several AI steps or tools into one smooth process.
When a task needs multiple models, tools, or steps working together, orchestration is the behind-the-scenes conductor that runs them in the right order and passes information between them.
Like a conductor leading an orchestra so dozens of players produce one coherent piece of music.
A special storage system that finds things by meaning, not exact words.
A database built to store embeddings so AI can quickly find the most relevant information—even when the wording doesn't match exactly. It's the memory bank that powers RAG and smart search.
Like a librarian who finds books about your topic even if you don't know the exact title.
AI that assists a person rather than replacing them.
A style of AI tool that works alongside someone—suggesting, drafting, and speeding things up—while the human stays in control and makes the final call. Common in writing, coding, and customer support.
Like a co-pilot: helpful and skilled, but you're still the one flying the plane.
A fixed sequence of steps that runs automatically.
Connecting tools so a set process happens on its own—'when a form is submitted, create a record, notify the team, and send a confirmation.' Rules are defined in advance and followed exactly.
Like a domino run: predictable, ordered, and the same every time you tip the first one.
Letting AI look up your documents before it answers.
A technique where the AI first searches your company's real information—policies, product docs, past tickets—then uses it to answer. This grounds responses in your facts and cuts down on made-up answers.
Like an open-book exam: instead of answering from memory, the AI checks the source material first.
Connecting AI to the tools you already use.
Wiring AI into your existing systems—your CRM, email, help desk, or database—so it can actually work inside your business instead of living in a separate window.
Like plumbing: the fancy fixture only matters once it's connected to the water supply.
A standard way for software to talk to other software.
A connection point that lets one program request services from another. It's how your app 'calls' an AI model, and how AI tools plug into your other systems.
Like a restaurant waiter: you don't enter the kitchen; you order through them and get your dish back.
Safety limits that keep AI from going off the rails.
Rules and checks that stop the AI from saying or doing things it shouldn't—leaking private data, giving harmful advice, or going off-topic. Essential for any AI customers interact with.
Like bumpers in a bowling lane: they keep things headed in a safe direction.
Unfair patterns AI picks up from its training data.
If the data an AI learned from was skewed, the AI can reflect and repeat that skew—treating groups unequally or favoring certain outcomes. It's a real risk that needs active checking.
Like a student who only read one-sided sources—they'll confidently repeat a lopsided view.
Tests that measure whether AI is actually doing a good job.
A structured way to check an AI's accuracy and quality against known-good answers—so you can catch problems before customers do and know when a change made things better or worse.
Like quality control on a production line: you test samples instead of hoping for the best.
Keeping a person in charge of important AI decisions.
A setup where AI does the heavy lifting but a human reviews or approves the results before anything final happens. The safest way to use AI for high-stakes work.
Like a co-pilot: the autopilot flies, but a trained human is ready to take the controls.
Protecting sensitive information when using AI.
Making sure confidential data—customer details, financials, health records—is handled safely and not exposed or used to train someone else's model. A key question to ask any AI vendor.
Like a locked filing cabinet: convenience shouldn't mean leaving the drawer open.
A specific job you want AI to do for your business.
The concrete problem you're solving—'draft first-response emails to support tickets' or 'summarize sales calls.' Good AI projects start with a clear use case, not the technology.
Deciding you need to hang a picture (use case) before you go buy a drill (tool).
Whether the value AI creates is worth what it costs.
The payoff from an AI project—time saved, revenue gained, errors avoided—measured against what you spent to build and run it. The number that decides if a project is worth doing.
Like asking whether hiring an employee pays for itself in the value they add.
A small test to prove an AI idea works before going big.
A limited first version built quickly to check whether an AI solution actually delivers value—before you invest in a full rollout. A smart way to reduce risk.
Like a test kitchen dish before adding it to the whole restaurant's menu.
Whether a solution keeps working as usage grows.
A solution is scalable if it still performs well and stays affordable when you go from ten users to ten thousand. Something that works in a demo doesn't always scale.
A food truck recipe is great—scalability is whether it works for a 500-seat restaurant.
How prepared your business is to adopt AI successfully.
A measure of whether you have the data, processes, and clear goals in place to get real value from AI. Assessing readiness first prevents wasted spending on projects that can't succeed yet.
Like checking you have a foundation before building the house on top of it.
AI that creates new content—text, images, audio, code.
The category of AI that produces original output rather than just analyzing existing data. It's what writes the email, drafts the image, or generates the summary.
Older AI sorts and labels the photos; generative AI paints a brand-new one.
Freely available models you can run yourself vs. paid ones you rent.
Open-source models can be downloaded and run on your own systems, giving more control and privacy but requiring more setup. Proprietary (closed) models are owned by a company and accessed for a fee—easier to start with, but you're renting access. The right choice depends on your privacy needs, budget, and team.
Open source is owning a car you maintain yourself; proprietary is a rideshare—convenient, but you don't own it.
A hypothetical AI as broadly capable as a human—doesn't exist yet.
The idea of an AI that could learn and reason across any topic as flexibly as a person, rather than being good at one narrow task. It's a long-term goal and a lot of hype—today's AI is not this.
Like the difference between a talented specialist and a well-rounded person who can pick up anything.
AI that's good at one specific job—which is all of today's AI.
AI built to do a particular task well, like answering questions or spotting fraud. Every AI tool you can actually buy today is narrow AI, even the impressive ones.
A world-class chess computer is brilliant at chess and can't do anything else.
The raw computer power AI needs to run.
The horsepower—usually specialized chips called GPUs—required to train and run AI. More compute means faster, bigger AI, but also higher cost. It's a major driver of AI expenses.
Like electricity for a factory: the more machines you run, the more power you burn.
Learning from labeled examples vs. finding patterns on its own.
Supervised learning trains on examples with the right answers attached ('this email is spam'). Unsupervised learning finds patterns in data with no labels. Reinforcement learning teaches through trial, reward, and error.
Supervised is studying with an answer key; unsupervised is sorting a pile of photos into groups yourself.
The family of models behind ChatGPT and many others.
Short for 'Generative Pre-trained Transformer'—a popular type of large language model. The name shows up in product names, but it just describes how the model is built and trained.
Like 'V8' in a car engine: a technical label that became a household brand.
The breakthrough design behind modern AI.
The architecture invented in 2017 that made today's language models possible. You don't need to know how it works—just that it's the engine design most modern AI is built on.
Like the invention of the assembly line: a behind-the-scenes breakthrough that changed everything built after it.
Teaching AI good manners using human ratings.
A training step where people rate the AI's answers, and the model learns to prefer the responses humans liked. It's a big reason modern chatbots feel helpful and polite instead of robotic.
Like coaching a new hire with feedback until their work matches what customers actually want.
A compact model that's cheaper and faster than the giants.
A smaller, more efficient model that can run at lower cost—sometimes even on your own device. Often good enough for focused tasks, without the price of a frontier model.
Like a scooter instead of a truck: less capacity, but cheaper and perfect for short trips.
When AI memorizes examples instead of learning the general idea.
A training problem where a model gets great at the exact data it studied but fails on anything new. It looks smart in testing and disappoints in the real world.
Like a student who memorized last year's exam answers but can't handle a new question.
A standard test used to compare AI models.
A shared set of tasks used to score and rank models against each other. Useful for comparison, but a high benchmark score doesn't always mean the model is best for your specific job.
Like standardized test scores: informative, but not the whole picture of a candidate.
The date after which a model knows nothing.
A model only knows what existed in its training data, up to a certain date. Ask about anything newer and it won't know—unless it's connected to live search or your documents.
Like a textbook printed last year: accurate up to a point, then frozen in time.
The field of getting computers to understand human language.
The broad area of AI focused on reading, understanding, and generating text and speech. It powers everything from spam filters to chatbots to voice assistants.
The umbrella skill of 'understanding language'—chatbots are just one thing built with it.
AI you interact with by chatting, like a helpful assistant.
Any AI designed to hold a back-and-forth conversation—answering questions, helping customers, or guiding someone through a task in plain language.
Like texting a knowledgeable colleague instead of digging through a manual.
Making AI answer in a tidy, predictable format.
Instructing the AI to return results in a fixed structure—like a form with set fields—instead of free-flowing text. Essential when the answer needs to plug into another system automatically.
Like asking someone to fill out a form instead of writing a paragraph, so filing it is easy.
Detecting the mood or opinion in text.
Using AI to tell whether a message is positive, negative, or neutral—handy for scanning reviews, support tickets, or social mentions at scale.
Like reading the room, but across thousands of messages at once.
Tying AI answers to real, verifiable sources.
Connecting the AI's responses to actual facts—your documents, a live database, trusted references—so answers are backed by real information instead of guesses.
Like requiring a report to cite its sources instead of stating things from memory.
Letting AI use apps and tools to get real work done.
Giving an AI the ability to actually do things—look up an order, send an email, run a calculation—by calling other software, instead of only talking about it.
The difference between an assistant who describes how to book a flight and one who books it.
Several AI agents working together on a bigger job.
Instead of one AI doing everything, multiple specialized agents collaborate—one researches, one writes, one checks the work—coordinated toward a shared goal.
Like a project team where each person has a role, rather than one person doing it all.
A common standard for plugging AI into your tools and data.
A shared 'language' that lets AI assistants connect to different apps and data sources in a consistent way—so you don't need custom wiring for every single tool.
Like a USB port: one standard plug that lets many different devices connect easily.
Software 'bots' that mimic clicks and keystrokes.
Older-style automation where a bot repeats the exact steps a person would take in software—copying data between screens, filling forms. Rule-based and rigid, increasingly paired with AI to make it smarter.
Like a player piano: it reproduces the keystrokes precisely, but doesn't improvise.
Building tools by clicking and dragging instead of programming.
Platforms that let non-programmers create apps and automations visually. They make AI and automation accessible to business teams without waiting on developers.
Like building with LEGO instead of machining each part from raw metal.
The event that kicks off an automation.
The 'when this happens' that starts an automated process—a new email arrives, a form is submitted, a deal closes. A webhook is the technical messenger that announces the event.
Like a motion sensor that switches the lights on the moment someone walks in.
Making sure AI's goals match what people actually want.
The work of getting AI to behave in line with human intentions and values—being helpful, honest, and safe rather than technically correct but unhelpful or harmful.
Like making sure a new hire understands the spirit of the job, not just the literal instructions.
A trick that sneaks hidden instructions past an AI's rules.
A security risk where someone hides malicious instructions in text the AI reads—getting it to ignore its guidelines or leak information. A key thing to defend against in customer-facing AI.
Like slipping a fake note into someone's stack of orders to make them do something they shouldn't.
When you can't fully see why AI gave an answer.
Modern AI often can't neatly explain its reasoning, which is a problem for regulated or high-stakes decisions. 'Explainability' is the effort to make that reasoning clearer.
Like a brilliant expert who gives great answers but can't tell you how they got there.
When an AI's performance quietly gets worse over time.
The real world changes—customer behavior, language, products—and a model trained on older patterns slowly becomes less accurate. It's why AI needs monitoring, not just a one-time setup.
Like a map that's accurate today but slowly goes stale as new roads get built.
Using AI in a way that's fair, safe, and accountable.
An approach that considers fairness, privacy, transparency, and impact on people—not just whether the AI works. Increasingly expected by customers and regulators.
Like food safety standards: it's not enough that the meal tastes good, it has to be safe too.
Data that identifies a specific person.
Information like names, emails, phone numbers, or account details that can identify someone. Handling it carefully with AI is both an ethical duty and often a legal requirement.
The stuff you'd shred rather than toss in the recycling—sensitive by nature.
AI-generated fake images, video, or voice that look real.
Synthetic media that convincingly imitates a real person's face or voice. A growing risk for fraud and misinformation, and a reason to verify unexpected requests.
Like an extremely convincing impersonator—now available to anyone with the right software.
Modernizing how a business runs using technology.
The broad effort to update processes, tools, and culture with digital technology—AI is often the latest chapter. A buzzword, but it points to real change in how work gets done.
Like renovating a house room by room, rather than just buying one new appliance.
The full cost of an AI solution, not just the sticker price.
Everything it takes to run AI over time—licenses, usage fees, setup, maintenance, and staff time—not only the upfront price. The honest number for comparing options.
A cheap printer with expensive ink can cost more than a pricier one over its life.
Getting stuck with one provider because switching is hard.
When your setup depends so heavily on one company's AI that moving away later would be costly or painful. Worth thinking about before you commit deeply to any single vendor.
Like building your whole kitchen around one brand's proprietary parts.
Paying for how much AI you actually use.
Instead of a flat fee, you're billed based on volume—often by tokens or number of requests. Costs scale with use, which is flexible but can be unpredictable without monitoring.
Like a utility bill: you pay for the electricity you actually consume.
Employees using AI tools without official approval.
When staff quietly use public AI tools for work the company hasn't vetted—risking data leaks and inconsistent quality. A sign you need a clear, sanctioned AI option.
Like employees using their own personal apps for company work, off the books.
How quickly an AI project starts delivering real benefit.
The stretch between starting a project and seeing actual results. Shorter is better—smart teams look for quick wins before committing to long, expensive builds.
Like planting fast-growing vegetables alongside the slow ones, so you eat sooner.
Buying a ready-made AI tool vs. building your own.
Off-the-shelf tools are fast and cheap to start but generic. Custom solutions fit your exact needs but cost more to build. Many businesses start off-the-shelf and customize once they know what works.
Like a suit off the rack versus one tailored to measure—speed and price versus fit.
A model whose trained 'brain' is downloadable, even if the recipe isn't.
The model's learned parameters (its weights) are released for anyone to download and run, so you can use it privately on your own systems. It's not always fully 'open source'—the training data and code may stay private—but you get the working model itself.
Like getting the finished, working engine to install in your own car—even if the factory blueprints stay secret.
The numbers that store everything a model has learned.
The millions or billions of adjusted values inside a model that hold its knowledge. When people talk about 'the weights,' they mean the actual trained model file you'd load and run.
Like all the fine-tuned settings in an instrument that together produce the right sound.
Shrinking a model so it's cheaper and faster to run.
A compression technique that trims the precision of a model's numbers to make it smaller and quicker, with usually only a small hit to quality. It's how big models get squeezed onto laptops and phones.
Like saving a photo at slightly lower resolution—much smaller file, still looks fine.
Training a small model to copy a big one.
A technique where a large, expensive model teaches a smaller, cheaper one to behave almost as well. You get most of the quality at a fraction of the running cost.
Like an expert writing a concise cheat sheet so a junior can handle most of the same questions.
A model that only wakes up the parts it needs for each task.
A design that splits a model into specialized 'experts' and activates just the relevant ones per request. This keeps a very large model fast and cheaper to run than using all of it every time.
Like a hospital routing you to the right specialist instead of every doctor seeing every patient.
A lightweight, cheap way to customize a model.
A popular fine-tuning shortcut that teaches a model new behavior by adding a small add-on layer instead of retraining the whole thing. Fast and inexpensive, and you can swap add-ons in and out.
Like clip-on lenses for a camera—you adapt it for a new shot without buying a new camera.
AI-generated fake data used to train or test AI.
Made-up but realistic data created to train models when real data is scarce, sensitive, or expensive. Useful for privacy, but it has to be built carefully so it reflects reality.
Like a flight simulator: not a real flight, but realistic enough to train on safely.
Search that understands meaning, not just matching words.
Finding results based on what you mean rather than exact keywords, so 'ways to cut costs' can surface a doc titled 'reducing expenses.' It's powered by embeddings and underlies a lot of modern AI search.
Like a helpful librarian who gets your intent, versus a search that needs the exact title.
AI showing its answer word-by-word as it writes.
Instead of waiting for the whole response, the answer appears live as it's generated. It's why chatbots feel fast and responsive even on longer replies.
Like watching someone type back in real time instead of waiting for a finished letter.
How long you wait for the AI to respond.
The delay between asking and getting an answer. Low latency feels snappy; high latency feels sluggish. It matters a lot for customer-facing and real-time uses.
Like the lag on a phone call—short is natural, long makes conversation awkward.
Reusing repeated instructions to cut cost and speed things up.
When the same big chunk of instructions or documents gets sent repeatedly, caching lets the AI reuse it instead of reprocessing every time—saving money and time.
Like a barista remembering your usual order instead of asking every single visit.
Tricking an AI into ignoring its safety rules.
A clever prompt that gets the model to do something it's supposed to refuse. Related to prompt injection, and a key thing to test for before putting AI in front of the public.
Like talking your way past a rule by finding a loophole in how it was worded.
Deliberately attacking your own AI to find weaknesses.
Having people try to break, trick, or misuse an AI system on purpose before launch, so you can fix the holes first. A best practice for anything high-stakes or customer-facing.
Like hiring someone to try to break into your building to expose the weak locks.
A short label describing what a model is and isn't good for.
A plain summary published with a model covering its intended uses, limits, and known risks. Handy for deciding whether it fits your needs responsibly.
Like a nutrition label—at a glance you see what's inside and what to watch out for.
Low-quality, mass-produced AI content.
The flood of generic, low-effort text and images churned out by AI with little human care. A reminder that using AI well still requires judgment, editing, and standards.
Like junk mail: cheap to produce in bulk, but nobody actually wants it.
Running AI on your own servers vs. someone else's.
On-premise (or self-hosted) means the AI runs on hardware you control—more privacy, more upkeep. Cloud means you use a provider's servers over the internet—easier and faster to start, but your data leaves your walls.
Like owning a generator versus buying power from the grid: control versus convenience.
Feeding the AI the right information to do a job well.
The growing practice of assembling exactly the right documents, data, and instructions an AI needs for a task—beyond just wording a prompt. Often the difference between a mediocre and a great AI result.
Like briefing a contractor with the full plans and site details, not just a one-line request.
A map of facts showing how people, places, and things connect.
A way of storing information as a web of connected facts—'Acme employs Jane,' 'Jane manages the Denver office'—so AI can follow the relationships between things, not just read isolated sentences. It gives AI a structured, connected view of your world.
Like a detective's cork board with photos linked by string, showing who's connected to whom.
A database built to store and explore connections.
A type of database designed around relationships between items rather than rows in a table. It shines when the connections matter as much as the data itself—org charts, supply chains, fraud rings.
Like a subway map: the value is in seeing how every stop links to the others.
RAG that pulls answers from a knowledge graph, not just documents.
A smarter version of retrieval where the AI draws on a knowledge graph of connected facts. It's better at questions that require joining several dots—'which of our clients are affected by this supplier's delay?'—than plain document search.
Instead of handing the AI a stack of reports, you hand it a wired-up map of how everything relates.
An agreed-upon vocabulary for how your information is organized.
The defined set of categories, terms, and relationships a system uses—what counts as a 'customer,' a 'product,' an 'order,' and how they relate. It keeps AI and data consistent instead of everyone inventing their own labels.
Like a shared filing system everyone in the company agrees to use.
The 'things' and the links between them that graphs are built from.
An entity is a distinct thing—a person, company, or invoice. A relationship is how two entities connect. AI can read your documents and pull these out automatically to build a knowledge graph.
Entities are the nouns; relationships are the verbs that tie them together.
Whether the AI remembers past interactions or starts fresh each time.
Stateless means the AI has no memory between conversations—every session starts from scratch. Stateful means it carries information forward. Most raw models are stateless; 'memory' features are added on top to make them stateful.
Stateless is a stranger who forgets you instantly; stateful is a regular who remembers your usual.
What the AI remembers within the current conversation.
The information an AI keeps track of during a single session—what you just said a few messages ago. It lasts only as long as the conversation and fits within the context window.
Like keeping a phone number in your head just long enough to dial it.
Information the AI keeps across many sessions, over time.
A stored memory that survives between conversations, so an assistant can recall your preferences or past decisions weeks later. It's usually built by saving key facts to a database the AI can look up.
Like a colleague who remembers your projects and preferences from months ago.
Remembering specific events vs. remembering general facts.
Episodic memory is the record of particular interactions—'last Tuesday you asked me to draft a refund email.' Semantic memory is distilled general knowledge—'this customer prefers email over phone.' Advanced assistants use both.
Episodic is your diary of events; semantic is the lessons you've learned from them.
Storing memories by meaning so the AI can recall the relevant ones.
A common way to give AI long-term memory: past information is saved as embeddings in a vector database, so when something similar comes up, the AI can pull back the most relevant memories—even if worded differently.
Like a mind that surfaces the right past experience the moment a similar situation appears.
An assistant's ability to remember you between chats.
A feature that lets an AI hold onto facts about you or your business across conversations—names, preferences, ongoing projects—so you don't have to re-explain every time. Built by saving and re-loading key details behind the scenes.
The difference between a new temp every day and an assistant who already knows how you work.
We translate the jargon into a plan. Let's talk about what AI can actually do for your business—in plain language.