You probably met AI in a chat window. You typed a question, got a decent paragraph back, and filed it under "fancy search box." That was fair. Chat apps now search the web and read files you upload, but they still do nothing for the stack of certificate requests in your quality inbox, because they cannot open your orders.
A chatbot is something you talk to, back and forth. A new hire is someone with a role, a purpose, and a job to do whether or not you are talking to them. The AI we build for manufacturers is the second kind, and the easiest way to understand it is to onboard it like one.
Onboard it like any new hire
Picture the checklist for a new coordinator's first week. They need an email address. They need a computer. They need logins to the systems they will use. They need training on how your shop does the job, a clear idea of what they are responsible for, and a manager who checks their work until they have proven themselves.
An AI employee needs exactly the same list. That is how we set one up.
A few of these deserve a plain explanation.
The computer is a virtual machine: a computer that runs as software on a server instead of sitting on a desk. The AI works there the way a person works at a desk. It opens files, runs programs, and keeps its work in its own folders. Working through commands and files on a computer is native to how today's AI already operates, so this gives it a place to do real work without touching anyone else's machine.
The logins are granted one system at a time, and reading comes before writing. For a certificate job, it might read the shared quality inbox and the order history, and write only to a drafts folder.
The training is your standard operating procedures plus 50 to 200 real past examples of the job done well: the customer's email, the certificate that went out, and a note on anything unusual. The model underneath does not learn on its own. Those examples go into files it reads every time it works, and we add new ones as it meets new cases.
The inbox: read, sort, never send
There are two ways to give an AI employee an inbox. It can have its own address that customers or colleagues write to, or it can connect to a person's existing inbox and read along. Either way, our rule at this stage of AI is simple: it can read, and it cannot send. Every reply it drafts waits for a person.
Reading is where the AI earns its keep. In one of our builds, the AI sorted and categorized 500 emails in under 8 seconds, because it works on many at once instead of one after another. No person can do that, and no person should have to.
The point is that your people stop being glued to their email. Requests that land after hours or over the weekend get read, sorted, matched to their orders, and drafted, so Monday morning starts with a queue of first passes instead of a pile of unread mail.
How it works on the floor
At one manufacturer we work with, customers email requests for Certificates of Conformance and Certificates of Origin to a shared team inbox. Before, someone opened each email, worked out which order and parts it referred to, pulled the details, and typed up the certificate. Each request was a small job, and together they ate hours of someone's week.
Now an agent sits in that inbox. It reads each request, matches it to the order, and drafts the certificate. A person on the team checks the draft and sends it. The team stopped typing certificates and kept every decision about them.
A second client, a U.S. building-products manufacturer, is running a pilot where one agent follows each order from the mill certificates through receiving, production, the yard, and dispatch, and passes each order's record into the next quote. Most of the time it keeps the record straight. When something needs a decision, such as a mismatch, a delay, or a question only a person can answer, it calls in the right person.
AI does the first pass. People verify, decide, and sign.
It will make mistakes, so we build the checks in
Every owner asks this, and the honest answer is yes. A model writes by predicting what should come next. When the order record is in front of it, the prediction matches the record. When the record is missing, it still predicts, and a made-up heat number looks exactly as confident as a real one.
People make mistakes too, and your operation already has ways to catch them. We build the same thing for the AI, in three parts.
First, we map the decisions. Before anything goes live, we sit with your team and write down, step by step, where the work goes and where a person has to verify it. That map decides what the AI does alone (sorting, matching, drafting) and where it must stop and ask.
Second, it flags what it does not know. If a heat number on the mill cert does not match the receiving log, or a request does not match any open order, the AI does not guess. It stops, flags it, and brings in the right person.
Third, it keeps an activity feed. When someone checks its work, they can see every step it took, in order, with the time and the source it used. Nothing is hidden inside the software.
Before it touches live work, it also has to pass a test: 50 past requests where you already know the right answer, graded field by field against a bar you wrote down first, with zero critical errors allowed. Then it drafts on live work for two weeks while nothing it writes is sent. Only after that does it go live, with a person approving every output.
Reviewing AI work is a skill, and most manufacturers have not trained anyone for it yet.
The highlighted row is the one that matters for a new hire. The person who reviews the drafts needs to know what a wrong one looks like, and an hour spent walking them through the AI's actual mistakes from its test is worth more than a general awareness course.
A new hire starts work without being asked
A chatbot only works when someone types to it. A new hire has a role, a purpose, standard operating procedures, and expectations, and those are exactly what we give the AI. That changes when it works. It does not wait to be asked. When a request comes in, it starts. When something is waiting on you, it nudges you before you go looking.
This works in every department that runs on reading, writing, and processing files: quality, sales, purchasing, customer service, accounting.
The question that decides success: is the job clear?
Owners often ask, "Can AI do accounting?" or "Can AI do quoting?" In our experience, that is rarely the question that decides anything. Today's AI can do most of what a person does at a computer. It can work a web browser, read and write files, and analyze a spreadsheet far larger than anyone wants to scroll through. Voice is coming too, although we do not deploy AI on phone calls today.
It still has limits, and the limits are worth knowing. On OSWorld 2.0, a public test of long desktop workflows that take a person about an hour and a half, the best score was 20.6% when the test launched in July 2026, and Anthropic reported 41.7% for its Fable 5.1 model in September. Long, unclear jobs are still where AI struggles most. Short, clearly defined jobs with a person checking the result are where it shines.
So the question that decides success is: are we clear on what it needs to do? Which requests, which documents, which decisions, and who checks. If a company cannot get to that clarity, we will not deploy an AI system, because an AI employee with a vague job does vague work.
Clarity first, then software, then AI
This surprises people, given how much we talk about AI: in manufacturing, a lot of problems are solved first with plain software: a digital form instead of a paper traveler, one record for each order instead of five spreadsheets, a clean parts list. AI is not the answer by itself. Clarity is.
So we do not throw AI at a problem. We start with clarity about the job, build the software and systems it needs, and then let the AI sit on top and use those systems as its tools, the way a new hire uses the ERP and the shared drive. That is what makes it act like an employee instead of a chat window.
There are two differences from a human hire. It is always on, including nights and weekends. And it can work on thousands of files and data points at the same time, where a person has to go one at a time.
The fears, named and answered
"My people will think they're being replaced."
Some will, at least at first. Say it out loud instead of hoping it goes away. AI takes the work nobody wanted first: retyping data into certificates, copying numbers between systems, hunting for the right PDF. The person who did that work still owns the result. They review it, catch problems, and handle the customer who calls with a real question. US manufacturers had 580,000 open jobs in July 2026 (preliminary), so for most manufacturers the hours that come back go to work that is already waiting for someone.
"I'm not technical."
You don't need to be. The hardest part of an AI employee is clarity, and clarity is a management job: which requests come in most often, which documents get typed by hand, who checks what. You know your operation better than any engineer. The technical setup can be handed to someone else. Deciding what the job is and who reviews it cannot.
What to do this week
None of these require buying anything.
- Pick one office role and write the onboarding checklist you would use for a new hire in that seat: inbox, computer, logins, training, job description, manager. That is the checklist for its AI employee.
- Open that role's inbox and tally last week's requests by type. Find the one that shows up most and gets answered the same way each time.
- For that request, write down each step and mark where a person has to verify something. That is the start of your decision map.
- Write down what should happen when the answer is not clear, for example a heat number that does not match. Those are the moments the AI should stop and flag.
- Name the person who would review the drafts, and ask them which parts of the job they would happily hand over.
Questions people ask
What is an AI employee?
An AI employee is an AI agent set up like a new hire: an inbox, a virtual machine to work on, logins to specific systems, training on your SOPs and past examples, one defined job, and a named person who reviews its work. It reads and drafts inside your systems, and a person approves anything that leaves the building.
Can an AI employee send emails on its own?
In the systems we build today, no. It can read an inbox, sort requests, and draft replies, but a person sends every email. That keeps a human signature on everything that reaches a customer.
What happens when the AI is not sure?
It flags the item instead of guessing, for example when a heat number on a mill cert does not match the receiving log, and brings in the right person. Every step it takes is written to an activity feed that anyone checking its work can read.
Terms in this piece
- Model
- the AI "brain" a company such as Anthropic, OpenAI, or Google trains on huge amounts of text, code, and images. It predicts, and knows nothing about your business until you show it.
- Agent
- a model put to work. Software that takes actions, like reading an email and drafting a document, instead of only answering questions.
- Chatbot
- a chat window where you type a question and get an answer back. It does not act inside your systems.
- Virtual machine
- a computer that runs as software on a server. The AI employee works on one the way a person works at a desk.
- Decision map
- a written, step-by-step map of a job that marks where the AI may act and where a person must verify.
- Activity feed
- the step-by-step record of everything the AI did, with times and sources, that a reviewer can read.
- Test set
- 50 past cases with known answers, set aside and used to grade the AI before it touches live work.
- 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.
