A customer emails your inside sales rep: "Do you have the 4-inch version of the bracket we bought last spring, and can you ship 200 by the end of the month?"
Your best rep answers that in five minutes because they know the part and the customer, remember the order, and know who to ask about stock. A sharp new hire with no training cannot answer it at all, since they have never seen your parts list.
AI starts out in the new hire's position. It can read, write, and reason well, but it knows nothing about your company until you show it.
What the AI can see decides what it can do
AI is only as useful as what it can see when it works.
We call that context: everything the AI has in front of it at the moment it does a task, such as the email it was asked to answer, the part record it looked up, and the customer's last three orders. If something is not in the context, the AI does not know it. It will usually guess, and the guess sounds exactly as sure as a right answer. Give it a part number it has never seen, and it may describe a part you do not make.
The first question to ask of any AI project is what the AI will be looking at.
Example: a client that loaded 100,000 parts first
One of our clients works with more than 100,000 parts. At that scale, nobody holds the whole list in their head, and no AI can guess it.
Before we asked the AI to do anything useful, we imported every one of those parts into the system, so it knows what the company makes, what it stocks, and what it has shipped. Now when a request comes in, the AI looks it up instead of guessing.
Everything after the import depended on it. Without the parts list, the AI would be the untrained new hire from the opening. With it, the AI can answer the bracket question the way your best rep would, with the part, the last order, and what is on the shelf. A person then confirms the ship date and replies.
A smart new hire with no parts list is still a new hire with no parts list.
The parts list is the first piece. For most of the jobs custom manufacturers want AI to help with, four sources matter most:
- The parts master lists what you make and buy (brackets, circuit board assemblies, sign cabinets), with numbers, revisions, materials, and descriptions that match what customers call the parts.
- The customer list records who your customers are, what they buy, their spec and cert requirements, and any special terms.
- Job history shows what you built before, for whom, in what quantity, and with what routing and problems.
- Approved costs cover material, labor rates, outside processing, and the margins your estimators have signed off on.
Give an AI quoting assistant those four, and a new RFQ lands next to the three most similar jobs you have run, with approved costs attached, for the estimator to review. Take them away, and the same AI produces a confident-sounding number built on nothing. As a job moves, the same record also collects supplier certs and the emails about it, which is what lets the next draft start from everything you know.
Why the data matters more than the model
A model is the AI engine a company like OpenAI, Anthropic, or Google trains and sells access to. People spend a lot of meeting time arguing about which one to use.
Our view is that for your work, that argument matters less than it seems. The leading models all read an email, find the relevant record, and draft a response well, and the scores below show how close they sit. Whether the model can see your job history decides whether you get a useful draft.
Artificial Analysis, an independent testing firm, runs the same set of tests on every model and rolls them into one score. The top model scores 58, OpenAI's best scores 53, and Meta's best, 13th on the list, scores 48, so the leaders from three labs sit within 10 points of each other.
Source: Artificial Analysis leaderboard. Each model was tested at its most thorough setting, where it takes longer to work through a problem. The order changes every few months as labs ship new models.
None of those tests measure whether a model knows your part numbers, and that is what decides the quality of a draft in your shop.
Estimators work the same way. A brilliant estimator with no access to past jobs or approved costs is guessing, while an average estimator with ten years of job history at hand is quoting.
This also means the model is replaceable and your data is not. By our count of lab announcements, the eight biggest AI labs, including Anthropic, OpenAI, and Google, made 52 notable model releases in the 52 weeks to September 25, 2026. Your parts master, your customer requirements, and your job history are what make any model useful in your shop, and they keep their value no matter which one you use next year.
How clean your data needs to be
Most owners tell us early on that their data is a mess, and the NAM numbers at the top say most of their peers would agree. They have duplicate part numbers, descriptions typed three different ways, costs in a spreadsheet that one person updates, and customer spec requirements sitting in someone's email folder.
Your data does not need to be perfect. It needs to be clean enough for the job you are asking the AI to do, which means:
- The parts list lives in one system or one file that everyone agrees is the real one, even if it has flaws.
- Each part number means one thing. Someone has marked duplicates and obsolete numbers, even if nobody deleted them.
- The fields the task needs are filled in. A CoO needs country of origin and a quote needs material and routing, but you do not need every field on every record.
- A named person owns the list and fixes errors when someone finds them.
Standardizing every description, reconciling every old job, and retiring every spreadsheet are good goals, but you do not need them to start.
Starting also speeds up the cleanup. When the AI drafts from your records, it shows you where they are wrong or empty. A cert draft that says "country of origin missing" points to one record, and once you fix it, every future draft that uses it is right. We prefer that to a six-month cleanup project, because each fix is tied to a document someone needed that day, so the records that matter get fixed first.
Data work comes first, and it pays off without AI
The first step in an AI project is usually data work: getting the parts list, the customer list, the job history, and the approved costs into a place where software can read them.
That work pays you back even without AI. New hires can find answers in the list instead of asking around, and fewer answers depend on the one person who has been there for decades. About 1 in 4 manufacturing workers is 55 or older, so that person is probably closer to retiring than you would like.
It also changes the vendor conversation. When someone pitches you an AI tool, ask what it will be looking at and how your data gets there. If the answer is vague, the tool will be too.
What to do this week
- Find the parts list people go to when they need an answer, which may differ from the one they are supposed to use.
- Find out who adds new parts, who fixes errors, and who would notice if the list were wrong. If the answer is nobody, write that down.
- Pull 20 recent orders and check whether each part number, description, and cost matches what is in that list.
- Ask your estimators, buyers, and coordinators which spreadsheets they keep on their own because the system is not good enough, and list them.
- Pick one job for AI to do first, such as cert drafting or quote starters, and list the fields that job needs. That list is your definition of clean enough.
Questions people ask
Does AI need access to my company's data to be useful?
Yes. An AI model knows nothing about your company until you show it your parts, customers, job history, and costs. Without that context, it guesses, and the guess sounds as sure as a right answer.
Which matters more for manufacturing AI, the model or the data?
The data matters more. The leading AI models are all good at reading an email and drafting a response, so the difference between them on a quote or a certificate is small. The difference between a model that can see your job history and one that cannot is the difference between a useful draft and a useless one.
How clean does my data need to be before starting an AI project?
It needs to be clean enough for the task. That means one agreed place to look, part numbers that mean one thing, the fields the task needs filled in, and a named person who fixes errors. Letting the AI draft from your records also shows exactly where they are wrong or empty, which speeds up cleanup.
Terms in this piece
- Context
- everything the AI has in front of it when it does a task, such as the email, the records it looked up, and the job history. If it is not in the context, the AI does not know it.
- Model
- the AI engine a company like OpenAI, Anthropic, or Google trains and sells access to. It reads, writes, and reasons but knows nothing about your company on its own.
- Parts master
- the single list of every part you make or buy, with numbers, revisions, materials, and descriptions.
- Import
- loading records from one system or file into another so software can read them.
- Clean enough
- data that is in one agreed place, has the fields a specific task needs, and has an owner who fixes errors.
- Quote starter
- a first-pass quote the AI drafts from similar past jobs and approved costs, which an estimator then reviews and decides on.
