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.

Use this piece as a cheat sheet, and do not try to read it straight through. Keep it by your desk for the next demo. When a vendor says a word you do not know, find the entry, read it, and go back to the demo.

The basics: what you are actually buying

An RFQ with six drawings lands in the estimator's inbox at 4:55 on a Friday. A vendor says their AI can draft the quote.

Before you ask how well it works, know what you would be paying for. It could be a chat window someone types into. It could be an agent with its own login, one defined job, and a named person who checks its work.

AI (artificial intelligence)
Software that reads, writes, and makes first-pass judgments from the documents and instructions you give it. Older software could only follow fixed rules. When a customer PO arrives as a PDF in a new layout, the AI can still pull out the part numbers and quantities.
Automation vs. AI
Automation follows fixed rules, like a CNC program that runs the same path every time. AI handles work that varies from job to job, such as reading a drawing it has never seen, and a person checks its work.
Model
The AI software itself, which a company like Anthropic, OpenAI, or Google trains to read and write on huge amounts of text, code, and images. Vendors also call it a large language model, or LLM. Claude, GPT, and Gemini are models, and ChatGPT and Claude are built on large language models. A model predicts, and it knows nothing about your company until you show it your work.
Prompt
The instructions you give an AI. Like a clear work instruction, a good prompt names the task and the material to use, and shows what good looks like.
Chatbot
An AI you type questions into. It answers in a chat window, then waits for your next question. The chat window in ChatGPT is one example. It has no login to your ERP and finishes no job on its own, so a person still copies its answer into the quote.
Agent
A model put to work on a task. It can read an email, look up an order, and draft a document. Take the RFQ from that Friday inbox. The agent reads it, pulls last year's price for the same part from the ERP, and drafts the quote for the estimator to check.
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 a quality lead who checks and signs them, is an AI employee.
First pass
The initial draft the AI produces. Like a first-article part, it still needs inspection before it counts.

How the AI gets your information

The estimator who quotes that RFQ needs the last price paid for the material, the customer's spec, and the drawings. The AI needs the same records when it drafts the quote, and the way it gets them decides whether a person still retypes numbers from one screen into another.

Some of those records, like an email or a scanned MTR, have no fixed fields. Vendors call that unstructured data, which means the facts are not in labeled boxes yet.

Data extraction
Pulling values out of documents and putting each one in its own labeled box. Documents with no fixed fields, such as an email, a scanned MTR, or a photo of a traveler, are what vendors call unstructured data. The answers go into labeled fields that stay the same every time, such as a box marked Heat Number, and vendors call that structured data. Reading heat number, grade, and chemistry off a scanned MTR is data extraction, and so is reading part numbers and quantities off a circuit board BOM.
OCR (optical character recognition)
Software that turns images of text, such as a scanned MTR, into text a computer can use. Current AI models read scans directly. A person still checks every value against the page.
Context
The records the AI has in front of it while it works. The parts master, the order record, and the customer's requirements can all be part of it. Leave the customer's cert requirements out, and the AI cannot follow them.
Context window
The most text and documents a model can take in at one time. A bid package for a flow line job may carry 40 drawings. If the package is bigger than the window, the AI cannot see all of it at once.
RAG (retrieval-augmented generation)
Retrieval means looking up records, and generation means writing. So the AI looks up records such as past quotes or specs first, then writes from them instead of from training alone. It checks the job file before it answers a customer.
Onboarding pack
A pack of 50 to 200 real past examples of the job, each showing the input, the approved output, and notes, along with the SOPs and templates. For order entry, it pairs past POs with the orders your team entered, and your team adds a note on each odd one.
API (application programming interface)
A standard doorway between two pieces of software, through which one can ask the other for data or send it instructions. When the AI looks up a price in your ERP or pulls a drawing from a shared drive, vendors call that tool use. The working connection between the AI and your software is an integration. It lets the AI pull material prices and drop a drafted order into the ERP as a pending entry, so no one has to retype it.
MCP (Model Context Protocol)
Vendor talk for a shared connection rule that any vendor can use. The rule lets an AI connect to other software and data the same way each time, like a standard hydraulic fitting. If your ERP vendor offers an MCP connection, an AI can plug in without a custom adapter.
Computer use
The AI uses 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 way your order clerk does, though a direct connection works faster.

How you test it before you trust it

Your quality lead has already approved 50 cert packets, and they sit in a locked folder. Before the AI drafts a live packet, it drafts those 50. A person checks each field against what the quality lead signed. You write the bar for passing before the test starts.

Benchmark
A standard test people use to compare AI models, the way a weld cert compares welders on the same coupon. A high score shows the model can do hard work in general. It does not prove the model will read your MTRs correctly.
Hallucination
The AI states a false fact and sounds sure of it, such as a heat number that is not on the MTR. Each value should link to its source, and a person should verify it.
Test set
A set of 50 real past cases with known answers. Your team seals the set before building starts and uses it to judge every version of the AI. Fifty past cert packets your quality lead approved can make a test set, once your team locks them in a folder and keeps them out of the setup work.
Evaluation
Testing an AI system on your own real work and checking its answers against what a person got right. On your 50-case test set, the team scores each run field by field.
Field-level accuracy
The share of single fields the AI got right on the test set, counted field by field. Take 50 cert packets with 12 fields each, for 600 fields in all. If the AI gets 588 right, it scores 98%.
Critical error
A wrong value in a field where a mistake can cause harm, such as a heat number or a country of origin. Your team names those fields in writing before testing starts. One critical error on the test set means the AI does not go live, however high the overall accuracy score.
Go-live threshold
The written bar the AI must clear before it starts live work. It sets the field-level accuracy the AI must reach and allows zero critical errors. An example bar is 98% of fields correct on the test set, with no wrong heat numbers.
Shadow mode
The AI drafts on live work for two weeks, and nobody sends those drafts. A person compares the drafts with what your people sent. The customer service team works as usual during those weeks while the AI drafts the same orders on the side.
Assisted mode
The AI drafts live work, and a person approves every output. It drafts an order entry from a PO, and the customer service rep checks that entry against the PO and saves it.
Verification
A person checks the AI's work against the source documents and corrects any errors before signing off. The reviewer opens the MTR next to the drafted cert and checks each heat number, or opens a 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, you should be able to point to where the data goes, who can see it, and who signs before it leaves the building.

Permissions
The rules that say what an AI can see and do on its own login. A new hire does not get the keys to everything on day one, and an AI should not either. An order-entry agent can read POs and draft orders, but it cannot change prices.
Guardrails
Limits built into an AI system that stop it from doing things it should not. One example stops the AI from sending an email to a customer without approval. They do the job a machine guard does.
Human in the loop
A named reviewer checks the AI's work and approves it at an approval step. The work cannot move ahead until that person says yes. The AI drafts the quote, and the estimator decides the price. For cert packets, the quality lead signs off before the packet goes to the customer.
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.
Prompt injection
Instructions that someone hides inside a document or email the AI will read, to make it do something its owner never asked for. Take a supplier email with a buried line that tells the AI to send your price list to an outside address. Limited permissions and an approval step block that move, because the AI's login cannot reach the price list and a person sees every outgoing message before anyone sends it.
Data privacy
The rules that say who can see your data and what a vendor may do with it. Get it in writing from every vendor that they will not use your drawings, prices, and job records to train their models or anyone else's.
Cloud
A provider runs software on its computers, and you use that software over the internet. Most AI models run in the cloud. Your text travels to the model provider's servers and back.
On-premises
Software that runs on computers in your own building, such as a server in your office, where your team runs it. You get more control and more maintenance work.
Open-weight model
A model you can download and run on your own computers, because the company that built it makes its trained settings public. Vendors call those settings weights. A shop with ITAR drawings can run one in-house, so the drawings never leave the building.

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.

Token
A small chunk of text, often part of a word, that an AI reads and writes. Vendors measure pricing and context windows in tokens, so a long RFQ package costs more to process than a one-line email.
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 you wait for the AI to answer. A minute is fine when the AI drafts a quote. It matters more when a person waits for an answer, like a customer on the phone.
Fine-tuning
Further training for a model, in which a specific set of examples teaches it to behave a certain way. Our view is that most companies do not need it. Examples and instructions in files, which the model reads each time it works, can do the job, and the files move with you when you switch models.
Digital twin
A live digital model of a machine, a line, or a whole facility that mirrors what the real one is doing. A screen can show a press line's cycle times as they happen. Our view is that you do not need one to start with AI, because most first AI jobs are paperwork.

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, and you still have no way to know the AI works.