AI vendors talk in a language built for engineers. Owners and ops leads end up nodding through demos, unsure whether "agent" and "model" mean the same thing, or whether a "context window" is something they need to buy.

This glossary is for owners and operations leads at manufacturers and industrial companies. It gives each term a plain definition and an example from the work you already do, grouped in the order you meet them: what you are buying, how it gets your information, how you test it, how the work stays safe, and how it is run and paid for. The terms for testing and launching an AI employee (test set, shadow mode, go-live threshold) match the ones in our AI Employee Handbook, so you can read one next to the other.

The basics: what you are actually buying

An RFQ with six drawings lands in the estimator's inbox at 4:55 on a Friday, and a vendor says their AI can draft the quote. Before you ask how well it works, you need to know what you would be paying for: a chat window someone types into, or an agent with its own login, one job, and a person who checks its work. These terms name the parts.

Agent
a model put to work, taking steps such as reading an email, looking up an order, and drafting a document. An agent can read an RFQ, pull last year's price for the same part from the ERP, and draft the quote for the estimator to check.
AI (artificial intelligence)
software that reads, writes, and makes first-pass judgments from documents and instructions, where older software could only follow fixed rules. It can read a customer PO that arrives as a PDF in a new layout and pull out the part numbers and quantities.
AI employee
an agent set up like a new hire, with its own login, access to specific systems, one defined job, and a named person who reviews its work. An agent that drafts cert packets under its own login, for the quality lead to check and sign, is one.
Automation vs. AI
automation follows fixed rules, like a CNC program running the same path every time. AI handles work that varies, like reading a drawing it has never seen, and needs a person to check it.
Chatbot
an AI you type questions into, which answers in a chat window and then waits for the next question. ChatGPT's chat window is one. It has no login to your ERP and finishes no job on its own, so someone still has to copy its answer into the quote.
First pass
the initial draft the AI produces. It is like a first-article part: it still needs inspection before it counts.
Frontier model
one of the most capable AI models available at a given time, built by a handful of large labs. In September 2026 that list included Claude Opus 5.5 and GPT-6 Astra.
Generative AI
AI that writes new text, such as a quote, a reply, or an order entry, instead of only sorting or scoring what it is given. An agent drafting a cert packet from the job record is generative AI at work.
Large language model (LLM)
a model trained on huge amounts of text, code, and images so it can read and write language. ChatGPT and Claude are built on large language models.
Model
the AI software a company like Anthropic, OpenAI, or Google trains on huge amounts of text, code, and images so it can read and write. It predicts, and knows nothing about your company until you show it. Claude, GPT, and Gemini are models.
Prompt
the instructions you give an AI. A clear prompt works like a clear work instruction: what to do, with what, and what good looks like.
Second org chart
a chart of your AI employees laid out the same way as your people, with each one's job, reviewer, and permissions. The agent that drafts cert packets sits under the quality lead, and the one that drafts orders sits under customer service.
Workflow
the sequence of steps a piece of work moves through, from the RFQ arriving to the quote going out. AI usually handles steps within a workflow, not the whole thing.

How the AI gets your information

The estimator quoting that RFQ needs the last price paid for the material, the customer's spec, and the drawings. An AI drafting the quote needs the same things, and how it gets them decides whether anyone still retypes numbers from one screen into another. These terms describe the connections between the AI and your inbox, files, and ERP.

API (application programming interface)
a standard doorway that lets one piece of software ask another for data or send it instructions. When your quoting tool pulls material prices from the ERP without anyone retyping them, it is usually going through an API.
Computer use
an AI's ability to operate a screen the way a person does, clicking and typing in software that has no API. It can key an order into an old ERP screen the same way your order clerk does, only slower than a direct connection.
Context
what the AI has in front of it while it works, such as the parts master, the order record, and the customer's requirements. Leave the customer's cert requirements out of the context, and the AI cannot follow them.
Context window
the maximum amount of text and documents a model can take in at once. If a bid package for a flow line job with 40 drawings is bigger than the window, the AI cannot see all of it at the same time.
Data extraction
pulling specific values out of documents and into fields. Reading heat number, grade, and chemistry off a scanned MTR, or part numbers and quantities off a circuit board BOM, and putting each in its own box is data extraction.
ERP (enterprise resource planning)
the system that runs orders, inventory, purchasing, and accounting. AI tools usually work alongside it, reading from it and drafting entries a person approves, instead of replacing it.
Integration
a working connection between the AI system and your existing software, like your ERP or email, so data moves without anyone retyping it. An integration can drop a drafted order straight into the ERP as a pending entry for a person to release.
MCP (Model Context Protocol)
an open standard that lets AI systems connect to other software and data sources in a consistent way. It works like a standard hydraulic fitting: if your ERP vendor offers an MCP connection, an AI can plug in without a custom adapter.
OCR (optical character recognition)
software that turns images of text, like a scanned MTR, into text a computer can work with. Current AI models read scans directly, and a person still checks every value against the page.
Onboarding pack
50 to 200 real past examples of the job, with inputs, approved outputs, and notes, plus the SOPs and templates. For order entry, that is past POs next to the orders your team entered and a note on each odd one.
RAG (retrieval-augmented generation)
a method where the AI looks up relevant documents, like past quotes or specs, before answering, instead of relying on what it learned in training. It is the AI checking the job file before answering a customer.
Structured data
information stored in labeled fields, the same way every time, like a heat number in a box marked Heat Number.
Tool use
an AI's ability to call other software to get something done, like looking up a price in your ERP or pulling a drawing from a shared drive.
Traceability
the ability to follow any part back through every step of the job to the material and documents it came from, such as a fabricated part back to its mill heat, or a board assembly back to its component lots.
Unstructured data
information without fixed fields, like a scanned page, a photo of a traveler, a customer's marked-up sign proof, or the body of an email.

How you test it before you trust it

Fifty cert packets your quality lead already approved sit in a locked folder. Before the AI drafts a single live packet, it drafts those fifty, and someone checks every field against what the quality lead signed. These terms are how you decide, in writing and ahead of time, whether it is good enough to work on live orders.

Assisted mode
live work in which the AI drafts and a person approves every output. The AI drafts the order entry, and the customer service rep checks it against the PO before saving it.
Benchmark
a standard test used to compare AI models, the way a weld cert compares welders on the same coupon. A high score says the model can do hard work in general, not that it will read your MTRs correctly.
Critical error
a wrong value in a field where a mistake causes harm, such as a heat number or a country of origin, defined in writing before testing. One critical error on the test set means the AI does not go live, whatever its overall accuracy.
Evaluation
testing an AI system on your own real work and checking its answers against what a person got right. Running the AI on your 50-case test set and scoring it field by field is an evaluation.
Field-level accuracy
the share of individual fields the AI got right on the test set, counted field by field. If 50 cert packets have 12 fields each, that is 600 fields, and 588 correct is 98%.
Go-live threshold
the written bar the AI must clear before live work, a field-level accuracy figure plus zero critical errors. For example: 98% of fields correct on the test set and no wrong heat numbers.
Hallucination
when an AI states something false with full confidence, such as a heat number that is not on the MTR. This is why every value should link to its source and a person should verify it.
Shadow mode
two weeks in which the AI drafts on live work, nothing it drafts is sent, and someone compares its drafts with what people sent. The customer service team works as usual while the AI drafts the same orders on the side.
Test set
50 real past cases with known answers, sealed before building starts and used to judge every version of the AI. Fifty past cert packets that your quality lead approved, locked in a folder nobody uses for setup, make a test set.
Verification
a person checking the AI's work against the source documents and correcting what is wrong before signing off. The reviewer opens the MTR next to the drafted cert and checks each heat number, or opens the pressure chart next to a drafted test record and checks the test pressure.

How the work stays safe

That cert packet carries your heat numbers, your customer's name, and your job record. Before an AI touches it, an owner should be able to point to where the data goes, who can see it, and who signs before it leaves the building. These terms are the answers, and each one is something you can check.

Approval step
a point in a workflow where a person must say yes before anything goes further. The quality lead's sign-off before a cert packet goes to the customer is an approval step.
Audit log
a record of every action the software took, when, and on whose authority. It works like the sign-off column on a traveler, for software.
Cloud
running software on a provider's computers over the internet, instead of on computers in your building. Most AI models run in the cloud, so your text travels to the model provider's servers and back.
Data privacy
the rules for who can see your data and what a vendor may do with it. Get it in writing from every vendor that your drawings, prices, and job records are not used to train its models or anyone else's.
Guardrails
limits built into an AI system that stop it from doing things it should not, like sending an email to a customer without approval. They do the job a machine guard does.
Human in the loop
a setup where a person reviews and approves the AI's work before it counts. The AI drafts the quote, and the estimator decides the price.
On-premises
running software on computers in your own building, instead of in the cloud. A server in your office that runs an open-weight model is on-premises. You get more control and more to maintain.
Open-weight model
a model whose trained settings are published, so a company can download it and run it on its own computers. A shop with ITAR drawings can run one in-house so the drawings never leave the building.
Permissions
the rules for what an AI system is allowed to see and do, set on its own login. A new hire does not get the keys to everything on day one, and neither should an AI: an order-entry agent can read POs and draft orders but cannot change prices.
Prompt injection
instructions hidden inside a document or email the AI reads, written to make it do something its owner never asked for. A supplier email with a buried line telling the AI to send your price list to an outside address is prompt injection. Limited permissions and an approval step are the defense: the AI's login cannot reach the price list, and a person sees every outgoing message before it is sent.
Reviewer
the named person who checks the AI's work and approves it before it counts. For cert packets, the reviewer is usually the quality lead.

How it is run and paid for

At month end, the model provider's bill lands on the controller's desk, itemized in tokens. Once an AI employee drafts cert packets every day, it has running costs like any machine, and vendors will offer extra work on top. These terms help you read the bill and judge the upsell.

Digital twin
a live digital model of a machine, a line, or a whole facility that mirrors what the real one is doing, such as a screen showing a press line's cycle times as they happen. Our view: you do not need one to start with AI, because most first AI jobs are paperwork.
Fine-tuning
further training a model on a specific set of examples so it behaves a certain way. Our view: most manufacturers do not need it, because examples and instructions kept in files the model reads each time do the job and move with you when you switch models.
Inference cost
what you pay each time the AI does work, usually billed by the amount of text it reads and writes. It is the cost per part, not the price of the machine.
Latency
how long the AI takes to respond. A minute is fine for a drafted quote. It matters more for anything a person waits on, like an answer for a customer on the phone.
Token
the small chunk of text an AI reads and writes, often part of a word. Pricing and context windows are measured in tokens, so a long RFQ package costs more to process than a one-line email.

Three questions for your next vendor call

Ask these three questions and listen for plain answers:

  1. What is the go-live threshold, and which fields count as critical errors?
  2. Will you test on our own sealed test set, and will we see the field-level accuracy?
  3. Who is the reviewer in assisted mode, and what can the AI do without approval?

A vendor who cannot answer in the terms above has not planned how you will know the AI works.