AI, in plain English.

A business process in which AI agents plan and carry out several steps toward a goal, using tools and making decisions within limits you set, instead of following a fixed script. Well-designed agentic workflows include checkpoints where a person approves important actions.

In practice: An agentic workflow for accounts payable might read incoming invoices, match them to purchase orders, flag mismatches for review, and queue approved bills for payment.

RelatedAI agentWorkflow automationHuman-in-the-loopTool use (function calling)

Software that uses an AI model to work toward a goal by deciding which steps to take, using tools such as email, a CRM, or a calendar, and checking its own progress. A chatbot answers questions; an agent can also take actions.

In practice: An AI agent could qualify a new web lead, look up the prospect's company, draft a personalized reply, and book a call on a salesperson's calendar.

RelatedAgentic workflowChatbotTool use (function calling)Guardrails

The policies, roles, and controls a company uses to decide which AI tools it adopts, how people use them, and how their results are checked. It covers data privacy, accuracy, vendor review, and who is accountable when something goes wrong.

In practice: A medical practice with basic AI governance knows which tools may touch patient data, who approved them, and how their output is reviewed before it reaches a chart or a patient.

RelatedAI policyShadow AIGuardrailsHuman-in-the-loop

When an AI model states something false or invented with the same confidence as a true answer, such as a fake citation, a wrong price, or a policy that doesn't exist. It happens because language models generate likely-sounding text rather than looking up verified facts.

In practice: A support assistant that hallucinates a refund policy can mislead a customer and start a dispute, which is why grounding, testing, and human review matter.

RelatedGroundingRAG (retrieval-augmented generation)Evals (AI evaluation)Large language model (LLM)

A written set of rules that tells employees which AI tools they may use, what information they can share with them, and how AI-assisted work must be reviewed before it goes out. For most small businesses, it is the most practical first step in AI governance.

In practice: A simple AI policy might approve one business-grade AI assistant for drafting emails while banning uploads of client financial records to free consumer apps.

RelatedAI governanceShadow AIAI readiness

How prepared a business is to adopt AI successfully, based on its goals, data, processes, people, and safeguards. A readiness assessment shows where AI can help now and what needs fixing first.

In practice: NextGen Code's free AI Readiness Assessment scores a business from 0 to 100 across five pillars: Strategy & Leadership, Data & Systems, Processes & Operations, People & Culture, and Governance & Risk.

RelatedAI governanceAI policyKPI (key performance indicator)Proof of concept (POC)

A defined way for one piece of software to request data or actions from another, such as your website sending a new order to your accounting system. APIs are how AI tools connect to the software a business already uses.

In practice: If your CRM has an API, an AI assistant can pull a customer's history before drafting a reply, so nobody has to copy and paste it.

RelatedTool use (function calling)Model Context Protocol (MCP)Workflow automationData pipeline

Software that performs tasks that usually take human judgment, such as understanding language, recognizing images, spotting patterns, and making predictions. Business AI ranges from simple forecasting models to large language models that can read and write.

In practice: Sorting support emails by topic, forecasting next month's sales, and reading scanned invoices are all everyday uses of AI in a small business.

RelatedMachine learningGenerative AILarge language model (LLM)Deep learning

Using software to complete a task without a person doing it by hand each time. Traditional automation follows fixed rules, while AI-assisted automation can also handle messy inputs like emails, documents, and phone calls.

In practice: Texting a review request 2 days after a job is marked complete is a simple, rule-based automation that runs the same way every time.

RelatedWorkflow automationRPA (robotic process automation)AI agent

The tools and practices that turn a company's data into reports and dashboards people use to make decisions. Good BI shows what is happening now, not just what happened last quarter.

In practice: A BI dashboard can show a restaurant group its daily sales, labor cost percentage, and food cost by location on one screen.

RelatedKPI (key performance indicator)Data warehousePredictive analyticsData pipeline

Software that holds a text conversation with people, usually on a website, in an app, or over messaging. Modern chatbots use large language models to handle open-ended questions, while older ones follow scripted menus and decision trees.

In practice: A chatbot grounded in your service menu and booking system can answer pricing questions and schedule appointments after hours.

RelatedAI agentVoice agentLarge language model (LLM)Grounding

The branch of AI that lets software interpret images and video, such as identifying objects, reading labels, counting items, or spotting defects.

In practice: A manufacturer can use computer vision on a production line to flag parts with visible defects before they ship.

RelatedMultimodal AIOCR (optical character recognition)Deep learning

The maximum amount of text, measured in tokens, that an AI model can consider at one time, including your instructions, any documents you provide, the conversation so far, and its reply. Anything outside the window is invisible to the model.

In practice: A large context window lets a model review a long contract in one pass, but you pay for every token you send, so sending only the relevant sections is often faster and cheaper.

RelatedTokenLarge language model (LLM)RAG (retrieval-augmented generation)

An AI assistant built into software you already use that suggests drafts, answers, or next steps while a person stays in control. Microsoft uses the name for its own AI products, but the word also describes the general pattern.

In practice: A sales copilot inside your CRM might summarize a customer's history and draft a follow-up email for the rep to edit and send.

RelatedAI agentHuman-in-the-loopGenerative AI

An automated series of steps that moves data from the systems where it is created to the places where it is used, cleaning and reshaping it along the way. Pipelines keep dashboards and AI tools supplied with current data without manual exports.

In practice: A nightly data pipeline might pull orders from Shopify, payments from Stripe, and expenses from QuickBooks into one database for reporting.

RelatedETL (extract, transform, load)Data warehouseBusiness intelligence (BI)API (application programming interface)

A central database built for analysis, where data from many business systems is stored together in a consistent format. Common examples include Snowflake, Google BigQuery, and Amazon Redshift.

In practice: With a data warehouse, you can see which marketing channel brings in your most profitable customers without stitching together spreadsheets from four different systems.

RelatedData pipelineETL (extract, transform, load)Business intelligence (BI)Structured data

A type of machine learning that uses neural networks with many layers to learn patterns from large amounts of data. It powers speech recognition, image recognition, and large language models.

In practice: The call transcription in a voice agent and the defect detection in a computer vision system both rely on deep learning.

RelatedNeural networkMachine learningLarge language model (LLM)

Lists of numbers that represent the meaning of a piece of text, an image, or other content, so that items with similar meanings end up with similar numbers. They let software compare content by meaning rather than by exact words.

In practice: Embeddings let a search for "broken AC" surface help articles about an air conditioner that isn't cooling, even though the words don't match.

RelatedSemantic searchVector databaseRAG (retrieval-augmented generation)

The process of pulling data out of source systems, cleaning and reshaping it into a consistent format, and loading it into a database or data warehouse. A related approach, ELT, loads the raw data first and transforms it inside the warehouse.

In practice: An ETL job might standardize customer names and addresses from your CRM and billing system so the same customer isn't counted twice.

RelatedData pipelineData warehouseStructured data

Structured tests that measure how well an AI system performs the tasks you care about, using a set of real or realistic examples with known good answers. Evals tell you whether a prompt change or a new model made results better or worse.

In practice: Before launching a support assistant, you might run it against 200 past customer questions and score how many answers were correct, complete, and within policy.

RelatedAI hallucinationPrompt engineeringProof of concept (POC)Guardrails

Including a few worked examples of a task in the prompt so the AI model can follow their pattern, format, and tone. It often improves consistency without any extra training.

In practice: Adding three sample product descriptions to your prompt helps the AI write new ones that match your brand voice and length.

RelatedZero-shot promptingPromptPrompt engineeringFine-tuning

Further training an existing AI model on your own examples so it adapts to a specific task, style, or format. Fine-tuning changes the model itself, unlike RAG, which leaves the model as is and supplies relevant information at the moment a question is asked.

In practice: A company might fine-tune a model to sort support tickets into its own categories after careful prompting alone still gives inconsistent results.

RelatedFoundation modelTraining dataRAG (retrieval-augmented generation)Few-shot prompting

A large AI model trained on broad data that can be adapted to many tasks, such as writing, summarizing, coding, or analyzing images. The models behind ChatGPT, Claude, and Gemini are foundation models.

In practice: Most business AI projects build on a foundation model through an API rather than training a new model from scratch, which keeps costs and timelines manageable.

RelatedLarge language model (LLM)Fine-tuningOpen-weight modelGenerative AI

AI that creates new content, such as text, images, audio, video, or code, based on patterns learned from its training data. Large language models and image generators are the best-known examples.

In practice: Drafting proposal sections, creating product images, and writing first drafts of social posts are everyday uses of generative AI.

RelatedLarge language model (LLM)Foundation modelMultimodal AIAI hallucination

The practice of making your business and content easy for AI assistants and AI-powered search, such as ChatGPT, Perplexity, and Google's AI Overviews, to find, understand, and cite. It builds on traditional SEO with clear facts, structured data, and quotable answers.

In practice: When someone asks an AI assistant to recommend a roofing company in their city, GEO is the work that improves your odds of being named in the answer.

RelatedStructured dataLarge language model (LLM)Grounding

Connecting an AI model's answers to trusted sources, such as your documents, databases, or live search results, so its responses rest on real information instead of only what it learned in training. Grounded answers can usually point to the source they used.

In practice: A grounded HR assistant answers from your actual employee handbook and shows which section it used.

RelatedRAG (retrieval-augmented generation)AI hallucinationSemantic search

Rules and checks placed around an AI system to keep it within safe, accurate, and on-brand limits, such as blocking certain topics, filtering sensitive data, or requiring approval before an action is taken.

In practice: Guardrails can stop a sales assistant from offering discounts above a set limit or sharing one customer's information with another.

RelatedAI governanceHuman-in-the-loopPrompt injectionSystem prompt

A design in which a person reviews, approves, or corrects an AI system's work at key points before it takes effect. It is a common safeguard for decisions that affect money, health, legal matters, or customer relationships.

In practice: An AI might draft every collections email, but a staff member approves each one before it's sent.

RelatedGuardrailsAgentic workflowAI governanceCopilot

Running a trained AI model to produce an output, such as an answer, a prediction, or a classification. When you use an AI model through an API, you typically pay for inference by the token or by the request.

In practice: Every time your website chatbot answers a customer, that's an inference call, so a high-volume use case needs a cost estimate per conversation before launch.

RelatedTokenLarge language model (LLM)Small language modelReasoning model

A metric a business tracks to judge progress toward an important goal, such as revenue per employee, average response time, or customer retention rate. Every AI project should name the KPI it is meant to move before anything is built.

In practice: For an AI intake assistant, the KPI might be the share of after-hours leads that get a response within 5 minutes.

RelatedBusiness intelligence (BI)Proof of concept (POC)AI readiness

An AI model trained on enormous amounts of text to predict the next token, or small piece of text, which lets it read, write, summarize, translate, and answer questions. ChatGPT, Claude, and Gemini are products built on LLMs.

In practice: An LLM can turn a rambling call transcript into a clean summary with action items in a few seconds.

RelatedFoundation modelTokenSmall language modelGenerative AI

A branch of AI in which software learns patterns from historical data instead of following hand-written rules, then applies those patterns to make predictions or decisions about new data.

In practice: A machine learning model trained on past jobs can estimate how long a new job will take based on its size, location, and crew.

RelatedArtificial intelligence (AI)Deep learningPredictive analyticsTraining data

An open standard, introduced by Anthropic in 2024 and since adopted by other major AI providers, that gives AI applications a common way to connect to tools and data sources such as file storage, databases, CRMs, and calendars. Instead of building a custom integration for every AI tool, a system can offer one MCP server that any compatible AI app can use.

In practice: With an MCP server for your project management system, an AI assistant can check a job's status or create a task without a one-off integration.

RelatedAPI (application programming interface)Tool use (function calling)AI agent

AI that can work with more than one type of input or output, such as text, images, audio, and video, within the same task.

In practice: A multimodal model can look at photos of a damaged roof, read the inspector's notes, and draft a summary for the repair estimate.

RelatedComputer visionGenerative AISpeech-to-textFoundation model

The field of AI focused on understanding and generating human language, including tasks like classifying emails, extracting names and dates, gauging sentiment, and translating. Large language models now handle most of these tasks well.

In practice: NLP can sort a shared inbox by request type and pull the order number out of each message automatically.

RelatedLarge language model (LLM)Semantic searchSpeech-to-textChatbot

A type of AI model loosely inspired by the brain, made of layers of connected mathematical units whose connections are adjusted during training until the network recognizes patterns. Neural networks with many layers are the basis of deep learning.

In practice: The models that transcribe phone calls, read handwriting, and generate text are all neural networks, even though each is trained for a very different job.

RelatedDeep learningMachine learningTraining data

Technology that converts text in images or scanned documents into machine-readable text. Modern document tools pair OCR with language models so they can also understand what the text means, such as which number is the invoice total.

In practice: OCR plus AI extraction can turn a stack of scanned delivery tickets into rows in your accounting system, complete with dates, amounts, and job numbers.

RelatedComputer visionUnstructured dataAutomationMultimodal AI

An AI model whose trained parameters, called weights, are published so anyone can download and run it on their own hardware or private cloud, subject to its license. Meta's Llama and Google's Gemma are examples. Open-weight is not always the same as open source, because the training data and code may stay private.

In practice: A firm with strict confidentiality rules might run an open-weight model on its own servers so sensitive documents never leave its network.

RelatedFoundation modelSmall language modelFine-tuning

Using historical data with statistical or machine learning models to forecast what is likely to happen next, such as demand, customer churn, or late payments.

In practice: Predictive analytics can flag the invoices most likely to be paid late, so your team can follow up before they're overdue.

RelatedMachine learningBusiness intelligence (BI)KPI (key performance indicator)

The instructions, questions, and context you give an AI model to get a response. The clearer and more specific the prompt, the more useful and consistent the output.

In practice: Asking the AI to "summarize this call in five bullet points and list every promised follow-up with a date" gets far better results than "summarize this."

RelatedPrompt engineeringSystem promptFew-shot promptingZero-shot prompting

The practice of designing, testing, and refining prompts so an AI system produces reliable, accurate output for a specific task. In business systems, it includes writing system prompts, choosing examples, setting output formats, and checking changes with evals.

In practice: Careful prompt engineering is often the difference between a demo that works once and a tool your staff can trust every day.

RelatedPromptSystem promptEvals (AI evaluation)Few-shot promptingTemperature

An attack in which hidden or malicious instructions are slipped into content an AI system reads, such as an email, web page, or document, to trick it into ignoring its rules or leaking data. It is one of the main security risks for AI agents that read outside content and can take actions.

In practice: An email with hidden text telling an AI assistant to forward the inbox to an outside address is a prompt injection attempt, which is why agents need tight permissions and approval steps.

RelatedGuardrailsAI agentSystem promptHuman-in-the-loop

A small, time-boxed build that tests whether an idea works with your real data and workflows before you commit to a full project. A useful POC has a success metric agreed on in advance.

In practice: A 2-week POC might test whether AI can correctly pull the key fields from 100 of your actual purchase orders before you automate the whole process.

RelatedEvals (AI evaluation)KPI (key performance indicator)AI readiness

A technique that lets an AI model answer from your own documents and data by first searching for the most relevant passages, then handing them to the model along with the question. It keeps answers current and specific to your business without retraining the model.

In practice: With RAG, a new employee can ask about your return policy and get an answer drawn from your actual policy documents, with a link to the source.

RelatedGroundingEmbeddingsVector databaseSemantic search

A large language model built to work through a problem step by step before giving its final answer, which tends to improve results on math, planning, coding, and multi-step analysis. Reasoning models are usually slower and more expensive per request than standard models.

In practice: A reasoning model is a good fit for checking a complex quote against contract terms, but it's overkill for tagging support tickets.

RelatedLarge language model (LLM)InferenceToken

Software bots that mimic the clicks and keystrokes a person makes in applications to complete repetitive, rule-based tasks, often between systems that lack APIs. Traditional RPA breaks easily when screens or inputs change, and AI can make it more flexible.

In practice: An RPA bot might log in to a supplier portal every morning, download new invoices, and enter them into your accounting system.

RelatedAutomationWorkflow automationAPI (application programming interface)OCR (optical character recognition)

Search that matches on meaning rather than exact keywords, usually by comparing embeddings. It finds relevant results even when the person searching and the document use different words.

In practice: Semantic search lets staff type "customer wants their money back" and find your refund policy, even though the policy never uses that phrase.

RelatedEmbeddingsVector databaseRAG (retrieval-augmented generation)Natural language processing (NLP)

Employees using AI tools for work without the company's knowledge or approval, often by pasting business information into free consumer apps. It creates privacy, security, and quality risks that a clear AI policy and approved tools can reduce.

In practice: A team member who uploads a client contract to an unapproved chatbot for a quick summary may expose confidential information without realizing it.

RelatedAI policyAI governanceGuardrails

A language model with far fewer parameters (the internal values it learned in training) than the largest models, designed to be faster, cheaper, and able to run on modest hardware such as a laptop, a phone, or a company's own server. Small models work best on narrow, well-defined tasks.

In practice: A small language model running on your own server could classify incoming documents or redact personal information without sending data to an outside provider.

RelatedLarge language model (LLM)Open-weight modelInference

Technology that converts spoken audio into written text, also called automatic speech recognition or transcription. It is the first step in most voice agents and call analytics tools.

In practice: Speech-to-text turns recorded sales calls into searchable transcripts that AI can summarize and review for coaching.

RelatedVoice agentNatural language processing (NLP)Multimodal AI

Information organized in a fixed format, such as rows and columns in a spreadsheet or database, where every field has a defined meaning. On websites, the term also refers to schema markup that tells search engines and AI systems exactly what a page describes.

In practice: Adding structured data to your location pages helps search engines and AI assistants read your hours, services, and service area correctly.

RelatedUnstructured dataData warehouseGenerative engine optimization (GEO)

Artificially generated data that mimics the patterns of real data without containing actual records about real people or transactions. It is used to test software, train models, or share data safely when real data is scarce or sensitive.

In practice: A clinic could test a new scheduling tool with synthetic patient records so no real patient information is exposed during development.

RelatedTraining dataMachine learningAI governance

The standing instructions an application gives an AI model before any user message, setting its role, rules, tone, and limits. Users usually don't see it, but it shapes every response.

In practice: A system prompt might tell your website assistant to answer only questions about your services, never quote prices, and hand complaints to a person.

RelatedPromptPrompt engineeringGuardrailsPrompt injection

A setting that controls how predictable or varied an AI model's output is. Lower temperatures produce more consistent, focused answers, and higher temperatures produce more varied, creative ones.

In practice: You'd use a low temperature for extracting invoice totals and a higher one for brainstorming marketing taglines.

RelatedPromptLarge language model (LLM)Inference

The unit of text an AI model reads and writes, usually a word or part of a word. As a rough rule, 100 tokens equal about 75 English words. Pricing, speed, and context window limits are all measured in tokens.

In practice: When an AI vendor prices usage per million tokens, you can estimate your monthly cost from the number and length of the documents or conversations you expect to process.

RelatedContext windowInferenceLarge language model (LLM)

The ability of an AI model to ask the surrounding application to run a specific function, such as searching a database, checking a calendar, or creating an invoice, and then use the result in its answer. It is how AI agents take actions in other software.

In practice: With tool use, an assistant asked when a customer's delivery will arrive can look up the order in your system instead of guessing.

RelatedAI agentAPI (application programming interface)Model Context Protocol (MCP)Agentic workflow

The examples an AI model learns from, such as text, images, or labeled business records. The quality, coverage, and bias of the training data shape what the model gets right and wrong.

In practice: A model trained only on summer jobs may underestimate how long winter jobs take, because its training data never included them.

RelatedMachine learningSynthetic dataFine-tuningFoundation model

Information that doesn't fit neatly into rows and columns, such as emails, PDFs, contracts, call recordings, photos, and chat logs. Much of a company's knowledge lives in this form, and modern AI is especially good at reading it.

In practice: AI can pull dates, amounts, and renewal terms out of hundreds of vendor contracts and load them into a spreadsheet you can sort and filter.

RelatedStructured dataOCR (optical character recognition)Natural language processing (NLP)RAG (retrieval-augmented generation)

A database built to store embeddings and quickly find the items closest in meaning to a query. It is a common building block for RAG and semantic search, and standard databases like PostgreSQL can add the same capability through extensions such as pgvector.

In practice: A vector database lets a support assistant search thousands of help articles by meaning and return the best matches in a fraction of a second.

RelatedEmbeddingsSemantic searchRAG (retrieval-augmented generation)

An AI system that holds spoken conversations, usually by phone, to answer questions, book appointments, or route calls. It either chains speech-to-text, a language model, and text-to-speech together, or uses a model that works with speech directly.

In practice: A voice agent can answer after-hours calls for a plumbing company, collect the problem and address, and book an emergency visit.

RelatedSpeech-to-textAI agentChatbotLarge language model (LLM)

Connecting the steps of a business process across apps so work moves forward on its own through triggers, conditions, and actions. Tools like Zapier, Make, and Microsoft Power Automate are common, and AI adds the ability to read, classify, and draft along the way.

In practice: When a client signs a proposal, workflow automation can create the project, send the welcome email, schedule the invoices, and notify the team.

RelatedAutomationAgentic workflowRPA (robotic process automation)API (application programming interface)

Asking an AI model to perform a task using instructions alone, with no worked examples. Modern models handle many tasks well this way, and when results are inconsistent, adding a few examples usually helps.

In practice: Asking a model to label each incoming email as sales, support, or billing, without showing it any labeled samples, is zero-shot prompting.

RelatedFew-shot promptingPromptPrompt engineering