Guide

A plain glossary of AI, software and web terms

14 min read

By . Published . Last updated . Reviewed by Mavrin Labs on . How this guide was researched and checked.

The Mavrin Labs glossary defines the technical words that come up when a small business or nonprofit plans AI, automation, software or a website, each in one or two plain sentences. Where a guide covers a term in depth, its entry links to that guide, so a definition leads straight to the decision it affects.

The entries are grouped by where they come up, starting with AI terms explained in plain words, then systems and data, the web and search, nonprofit finance, and the project itself. Where a public body sets the meaning of a term, the entry links to that body's own text.

Look at the taskWorkflow assessment, baselineSettle the data and thesystemsSystem of record, CRM, CRMintegration, API, middleware, ETL,least privilegeAdd the AI stepsLarge language model, AIautomation, AI agent, agentic AI,hallucination, prompt injectionTest before launchAcceptance testPublish and be foundStructured data, llms.txt,robots.txt, search console, WCAGReport from the booksForm 990, restricted gift
Where the twenty-three terms come up on one project, from the first look at a task to the website and the books; each step lists the entries below that belong to it.

Why it matters: plain words make better decisions

Plain definitions matter because dense text leaves many readers behind. Between 2017 and 2023 the share of U.S. adults aged 16 to 65 at the lowest literacy level rose from 19 to 28 percent, in the federal survey of adult skills (federal highlights of the 2023 adult skills survey, checked Tuesday, November 17, 2026).

A term an owner can read in one sentence is a term the owner can question. That makes it easier to ask the right thing before a project starts, and harder for a vague promise to hide behind a technical word.

AI terms

  • Large language model

    A large language model (LLM) is the kind of AI model behind chat assistants, trained on large amounts of text to predict and write language. The model reads and writes words well, and it can state something untrue with the same confidence as something true.

    In a business build, the model handles the steps that need reading or writing, such as sorting a request or drafting a reply, while fixed rules and a person handle the rest. The plain definition of AI automation shows where the model sits in a workflow.

  • AI agent

    An AI agent is software that reads a request, decides the next step within limits set for it, and carries the step out, with a person approving what matters. Unlike a chat assistant, an agent acts as well as answers.

    An agent's limits are written down before it is built: its goal, the records it may read, the tools it may use and the point where it stops and asks. The plain definition of an AI agent sets out each part and the safeguards around it.

  • Agentic AI

    Agentic describes AI that takes steps on its own toward a goal rather than only answering a question. Agentic AI is the broad label, and an AI agent is one working form of it inside a business.

    Public guidance treats agentic systems with extra care. CISA and international partners advised on Friday, May 1, 2026 that first uses be low-risk and kept away from sensitive data and critical systems (CISA announcement on agentic AI). The custom AI agents guide shows where those limits sit in a build.

  • AI automation

    AI automation is software that follows a set workflow and calls an AI model only for the steps rules cannot handle, such as reading an email or drafting a reply, with a person approving the result. The workflow stays fixed, and only those steps use AI.

    A fixed sequence is easier to test than an agent, which makes AI automation the safer first project for a small team. The guide to what a small business should automate first sorts the candidates by department.

  • Hallucination

    A hallucination is an AI answer that sounds confident but is not true. The model is not lying on purpose: it writes the most likely words, and sometimes the most likely words are wrong.

    NIST's generative AI profile of July 2024 calls the same problem confabulation and lists it among the risks a business should manage (NIST's generative AI profile). The counter is a person who checks names, figures and dates before anything leaves the business.

  • Prompt injection

    Prompt injection is hidden or crafted instructions in text an AI reads that try to make it ignore its rules, such as a line in an email telling an assistant to forward the whole inbox.

    OWASP's 2025 list of risks for language-model applications ranks prompt injection first (OWASP's list of top risks). The defense is to limit what the AI can do, not only what it is told, and to test it with hostile text before launch.

Systems and data terms

  • CRM, customer relationship management

    A CRM is the program that holds client and prospect records, with every call, email and deal against each name. The letters stand for customer relationship management.

    Follow-up work starts from the client record, so the CRM is often the first program an automation connects to. The definition of CRM integration explains how it joins the accounting software, the calendar and the website form.

  • CRM integration

    CRM integration is the connection of a CRM to the other programs a business runs, such as accounting software, a calendar or the website form, so a record entered once reaches every place that needs it.

    A sound integration names the program that owns each field and says what happens when a step fails. The integration checklist covers both, along with duplicate records and upkeep after launch.

  • API, application programming interface

    An API is a published doorway one program offers so another program can read or send its data without a person copying it. The letters stand for application programming interface.

    Whether a program offers a usable API decides whether it can be connected at all. OWASP's 2023 list of API security risks ranks broken object level authorization first: a request that returns a record the caller should not see (OWASP list of API security risks).

  • Middleware

    Middleware is software that sits between two programs and passes data from one to the other, translating formats and retrying when a connection drops.

    Middleware helps when two programs cannot talk to each other directly, and it becomes one more thing to watch after launch. The framework for choosing a system weighs it against a direct connection or a new tool.

  • System of record

    A system of record is the one program whose copy of a figure or record counts when two programs disagree. Contact details often belong to the CRM, and invoice totals to the accounting software.

    Naming the system of record for each field is the first decision of any integration or report, because without it two copies drift apart and nobody can say which is right. The reporting guide applies the idea to every figure in a report.

  • ETL, extract, transform, load

    ETL is copying data out of one system, cleaning and reshaping it, and loading it into another for reporting. The letters stand for extract, transform, load.

    ETL is the plumbing behind a report built from several systems, which the guide to reports whose figures agree plans figure by figure. Each run happens on a schedule, and each should be safe to repeat, so a retry after a dropped connection never counts the same sale twice.

  • Least privilege

    Least privilege means giving each person or program only the access its job needs, and no more. A connection that only reads invoices gets read access to invoices, not to payroll.

    NIST's glossary defines the principle the same way for users and for processes (NIST glossary entry on least privilege). The principle limits the damage when a password, a key or an AI tool is misused.

Web and search terms

  • Structured data

    Structured data is labels in a page's code that tell search engines what the page is about, such as the firm's name, a question and its answer. It is also called schema markup.

    Google's guidelines say structured data must describe what the page visibly shows (Google's structured data guidelines). The AI search visibility guide explains what the labels can and cannot do for a business.

  • llms.txt

    A file named llms.txt is a plain-text file at the root of a website that lists its main pages for AI tools to read first, with a line about each.

    The file's authors describe it as a proposal, and Google's guide for site owners says such files are not needed to appear in its AI features (Google's guide to AI features in Search). Publishing one does no harm.

  • robots.txt

    A file named robots.txt is a plain-text file that tells search and AI crawlers which parts of a site they may read. The file is a request, not a lock, and a crawler can ignore it.

    The robots standard, RFC 9309, says that when the file cannot be fetched because of a server error, a crawler should treat the whole site as off limits (RFC 9309, checked Monday, November 16, 2026). A broken file can hide a whole site, as the guide to being found in AI answers explains.

  • Search console

    The search console is Google's account tool that shows which searches bring a site's pages up and any problems Google found while reading them.

    A business uses the search console to check that its pages are indexed and to see the queries that reach them. The tool counts what happens in Google's own results, so other engines and AI assistants need their own checks.

  • Accessibility guidelines (WCAG)

    WCAG (the web content accessibility guidelines) are the international guidelines for making web pages usable by people with disabilities, written by the web's main standards body.

    The current version, dated December 2024, sets testable thresholds such as text contrast and the size of tap targets (the published accessibility guidelines, checked Tuesday, November 17, 2026). The website planning guide shows how to check a finished site against them.

Nonprofit finance terms

  • The annual IRS return for nonprofits

    Form 990 is the annual information return most US tax-exempt organizations file with the IRS, and its figures come from the books (IRS instructions for Form 990, checked Monday, November 16, 2026).

    The return draws on the same records as every funder report, which is why grant tags and net asset classes matter long before filing season. The guide to grant reports from the books shows how to keep them ready all year.

  • Restricted gift

    A restricted gift is a donation the donor has limited to a stated purpose or period, such as a gift for one program or for next year's work.

    Nonprofit statements show restricted money apart from the rest, under the classes of net assets FASB set in August 2016 (FASB's update on nonprofit statements). The donor database to accounting guide shows how a restriction travels between the two systems.

Project terms

  • Workflow assessment

    A workflow assessment is the first stage of a Mavrin Labs engagement: the chosen task is mapped as it runs today, timed and written down, so there is a baseline to judge any build against and a fixed plan for the first build.

    The sample workflow assessment report, built on made-up data, shows what the stage hands over and what the client decides at the end of it.

  • Acceptance test

    An acceptance test is a check written before a build that says what a correct result looks like for a real case, so everyone can see whether the finished system passes.

    For an AI system, the tests include the hard cases: a message in a customer's own words, an attempt to make it break its rules, and a request it should refuse. The guide to testing an AI agent lists them in full.

  • Baseline

    A baseline is the measure of a task as it runs today, recorded before anything changes: how long it takes each week, how often it goes wrong and how long items wait.

    Without a baseline, nobody can say whether a build gave hours back. The hour-value tool turns a weekly baseline into hours a year, and the page on agreeing a measurable result before the build explains how the target is set.

How do AI automation, an AI agent and a chat assistant differ?

AI automation, an AI agent and a chat assistant differ in who chooses the steps and whether anything happens outside the conversation, which also decides how much testing each one needs before anyone relies on it. A chat assistant answers what it is asked and acts nowhere else.

AI automation runs fixed steps and uses a model for some of them. An AI agent picks its own next step toward a goal, inside limits a business sets.

How do AI automation, an AI agent and a chat assistant differ?
Chat assistant AI automation AI agent
What it does Answers a question in a conversation Runs a fixed workflow with a model in some steps Works toward a goal, one step at a time
Who sets the steps Nobody: it replies The builder, in advance The agent, within written limits
Acts in other programs No Yes, in the steps it was built for Yes, with the tools it is allowed
Where a person stays Reads every answer Approves what leaves the business Approves what matters and handles hand-overs
Easiest to test Not a build to test Yes, against agreed examples Harder: the goal, the limits and every tool

The order of the columns is also the usual order of risk. A first project for a small team is most often AI automation, because its fixed steps can be tested before launch, while an agent suits work that changes from case to case.

How were the terms chosen, and what has changed?

The terms were chosen because they come up in the guides on this site: the site's own term list, plus five words its guides lean on (AI automation, CRM integration, workflow assessment, acceptance test and baseline). A term joins when a guide starts to depend on it, and an entry links that guide where one covers the term.

Corrections and changes to this page, newest first:

  • Wednesday, November 18, 2026: each term is set in bold where its entry defines it; the search console heading now names its term; three entries link the guide that covers them; the comparison table and the project diagram were added.
  • Tuesday, November 17, 2026: the page was published, then retitled that evening to name its search phrase.

Key takeaways

  • Each term here is defined in one or two plain sentences, with a link to the guide that covers it where there is one.
  • AI automation follows fixed steps, while an AI agent chooses its own within set limits.
  • A system of record and least privilege settle whose data counts and who may see it.
  • Public bodies own several terms, and their entries link to the source text.
  • A baseline and an acceptance test make any build's result checkable.

Where to go next

The next step from a definition is a decision about your own week. The six-question project finder asks where the hours go, points to a first project and asks for no contact details.

Every guide in the set is listed in the library of guides. To talk through one term or one task, start a conversation.

Frequently asked questions

Who is the glossary for?

The glossary is for owners, office managers and nonprofit leaders who meet technical words while planning a project and want a plain answer before a meeting. Each entry gives the definition in a sentence or two, then points to the guide that explains what the term changes for a business like theirs.

What is the difference between AI automation and an AI agent?

AI automation follows a fixed sequence of steps and calls an AI model only for the parts rules cannot handle. An AI agent is given a goal and chooses its own steps toward it within set limits. Automation is easier to test, while an agent suits work that changes from case to case.

Which terms should I understand before hiring an AI firm?

Before hiring an AI firm, understand five terms: system of record, least privilege, acceptance test, baseline and prompt injection. Together they cover which program's data counts, how much access the build gets, how it is tested, how the result is measured and how it can be attacked through the text it reads.

Where do the definitions come from?

The definitions come from the guides on this site, written in plain English for a busy owner. Where a public body sets the meaning, such as a standards group or a government agency, the entry links to that source, so each definition can be checked against the body that owns the term.

How often is the glossary updated?

The glossary is updated whenever a guide adds a new term or a public source changes a definition, and each update moves the date shown at the top of the page. Corrections follow the same rules as every guide, which the editorial policy sets out, including how a reader can report an error.

The Mavrin Labs mark

Who writes these guides

Client work at Mavrin Labs is delivered by a team in mixed roles, led by our founder. AI tools help with research and drafting, and a person at Mavrin Labs checks each fact against its named source before a guide is published.

Enquiry form

Start with one workflow

Tell Mavrin Labs what is stuck and which systems it touches. A person at the firm reads every enquiry and replies by email.

Start a conversation

Fields marked with an asterisk are the ones a reply needs.

So a reply can address you properly.

The address the reply goes to.

Optional. Give one only if you would rather talk than write; the reply goes by email either way.

Optional, and useful when it says what the work has to fit around.

The result you want, and when you would want it working. Two or three sentences are plenty.

The task as it actually happens, roughly how often it runs, and which systems it touches. Two links at most, please.

Optional.

What you send goes to the Mavrin Labs inbox and is used to answer you, and a phone number, if you give one, is used for the same thing and nothing else.

A person reads it and replies by email.

If the form does not send, write to info@mavrinlabs.com instead; it reaches the same inbox.

Which hours would you take back first?

Name the one workflow that eats the most of your week. Mavrin Labs reads it, and the reply comes from the people who would build the fix.

Start a conversation