AI terms in plain English
AI proposals and contracts are full of words that sound clear and are not. This page explains the common ones in a few sentences each.
No jargon walls. Where we have written more on a topic, the term links to it.
AI agent
An AI agent is software that uses an AI model to work through a task in steps and take actions, such as reading an email, looking up a record and drafting a reply. For AI agents for business, the useful question is not "how smart is it" but "what is it allowed to do, and who approves its actions." A well-built agent works inside clear limits, with a person approving anything that matters.
AI acceptable use policy
An AI acceptable use policy is a short company document that says which AI tools staff may use, what data they may put into them, and what must be checked by a person. It works best when it is short and gives people an approved path that is easier than the unofficial one. It is a direct answer to shadow AI.
AI governance (and AI agent governance)
AI governance is the set of rules, roles and records a company uses to decide which AI it uses, for what, and how it checks the results. AI agent governance applies the same idea to agents that take actions: which systems an agent can touch, what needs human approval, and how every action is logged. For a small company it can fit on one page.
AI workflow automation
AI workflow automation means a recurring task runs on its own from start to finish, with AI used only for the steps that need judgment, such as reading a document or sorting a request. The rest is plain rules. The difference from a chat tool: a person no longer has to copy the answer into the next system. More on moving from chatbot habits to automation.
Business associate agreement (BAA)
A BAA is a contract HIPAA requires when an outside company handles protected health information for a HIPAA covered entity (a health plan, a health care clearinghouse, or a health care provider that conducts certain billing and payment transactions electronically), or for another business associate. HHS says a cloud provider that stores or processes that data is a business associate, even if the data is encrypted and the provider has no key. A BAA is a contract, not a guarantee of safety. See how to protect proprietary data when using AI.
Cloud AI (vendor cloud)
Cloud AI is an AI service that a provider runs on its own servers for many customers. You send prompts or documents over the internet and get answers back. It is fast to start, but your data goes to an outside party under that party's terms. Compare the options in on-premise, private cloud or vendor cloud.
Data retention
Data retention is how long a provider keeps your prompts, files, outputs and logs, and what happens to them after. Each type of data can have a different period, and backups can keep copies longer. Ask for it in writing. See what "no training" and retention claims really mean.
Fine-tuning
Fine-tuning means training an existing AI model further on your own examples so it follows your style or format more closely. It changes the model itself. For answering questions from company documents, retrieval (RAG) is often the simpler choice, because the documents can change without retraining.
Hallucination
A hallucination is when an AI model gives an answer that sounds confident but is wrong or made up, such as a fake citation or a number that is not in the source. It is a known limit of these models, not a rare bug. Citations to sources and a person checking the output are the practical defenses. See what RAG is.
Human in the loop
Human in the loop means a person reviews or approves the AI's work at set points before it reaches a customer, a record or a decision. The skill is choosing those points well: review what carries risk, and let routine, low-risk steps run. Keeping data private does not remove the need for review, as we cover in what private AI doesn't fix.
Intelligent document processing
Intelligent document processing (IDP) is software that reads documents such as invoices, receipts and forms, pulls out the fields you need, and sends them to the right system. AI handles messy layouts and handwriting. Rules and a person handle the exceptions. Here is an example of receipt processing based on our work.
Large language model (LLM)
A large language model is the AI engine behind many chat tools. It was trained on large amounts of text, and it predicts the next words to write an answer. It is good with language and weak at knowing whether it is right, which is why the setup around it matters.
On-premise AI
On-premise AI runs on hardware at your own site, such as a server in your office. In normal use, prompts and documents stay on systems you control. You still need someone to patch, back up and secure the machine. See on-premise, private cloud or vendor cloud.
Open source and open-weight models
Many models described as "open source" are open-weight: the trained model files are published so others can download and run them. The training data and code may not be published, and the license terms differ from model to model. Read the license before commercial use.
Pilot vs production
A pilot is a small, time-limited test with a few users to see whether the tool works. Production means the tool runs every day, inside real work, with support, monitoring and an owner. Many AI projects stall between the two, so agree on what "it works" means before the pilot starts. See what the first 30 days should deliver.
Private cloud
A private cloud is computing set aside for one organization. It can sit in your building or in a hosting company's data center, run by you or by a third party. "Private" does not always mean "on our own servers," so ask who runs the hardware and where it is.
Private deployment
A private deployment is an AI setup used only by your organization, on hardware you control or lease, instead of a shared outside AI service. In normal use, prompts and documents stay inside your environment. It controls where data goes; access rules, logs and review still need work. More in how to protect proprietary data.
Prompt
A prompt is the instruction or question you give an AI model, plus any text or files you include with it. Clear prompts give better results. In an automated workflow, prompts are written once, tested and reused, so results do not depend on who is typing.
RAG (retrieval-augmented generation)
RAG means the AI first searches your own documents for the relevant passages, then writes its answer from them, often with citations. It keeps answers tied to your current documents without retraining the model. Our explainer: what is RAG, and why not just upload files to a chat tool.
Self-hosted AI
Self-hosted AI means you, or a firm working for you, run the AI model on servers you control, in your office or on dedicated servers in a data center. No shared outside AI service sees your prompts in normal use. You take on the upkeep in return. See on-premise, private cloud or vendor cloud.
Shadow AI
Shadow AI is staff using AI tools the company has not approved or does not know about, often with real company data. It often starts with good intent: people want to work faster. The fix is rarely a ban. A clear AI acceptable use policy and an approved tool that is easy to use work better.
Tokens
Tokens are the small chunks of text an AI model reads and writes, often a word or part of a word. Many AI services charge by the token, and each model has a limit on how many tokens it can handle at once. More documents and longer answers mean more tokens.
Zero data retention
Zero data retention is a provider promise not to store your prompts and outputs after the request is done, or to keep them only very briefly. The exact terms vary, and some providers still keep limited data for purposes such as abuse monitoring. Read the terms for what is and is not covered. See what retention claims really mean.
Before you sign anything
Terms are the start. The questions come next. Our checklist, Before you sign: where will our AI run?, lists twelve to ask any AI provider.
This page is general information, not legal advice.
If you want to talk through which of these terms apply to your work, start with a free consultation.