Every manufacturer has an org chart. It might be a clean diagram in the HR folder or a whiteboard in the GM's office with names crossed out and rewritten. Either way, everyone knows who does what and who they answer to.
That chart tells a new hire whose approval they need, and when something goes wrong on a shipment, it tells you whose desk to walk to.
Now add software that drafts certificates and reads the purchasing inbox, with no owner and no written limits. When it gets something wrong, nobody knows whose desk to walk to. We recommend drawing a second org chart before that happens.
Going by the Census figure above, most companies your size have not started. Our view from client work is that when AI stalls, it usually lacks an owner before it lacks capability. The second org chart gives it one.
What an AI org chart is
Your company already has a human org chart. It now needs an AI org chart next to it.
Each AI employee on that chart gets the same things you would give a person: a role, a manager (a named person), permissions, and limits on what it can do without asking. The chart exists so the people on the first chart spend less time retyping and more time on the calls only they can make.
An AI employee is an agent: software built on an AI model (the part a company such as Anthropic, OpenAI, or Google trains on huge amounts of text, code, and images) that can take actions inside your systems, like reading an email or drafting a document, instead of only answering questions in a chat window. You onboard it like a hire, with its own computer, its own files, and access to the places where your work happens.
The roles depend on what you make. A circuit board assembler might add a role that drafts travelers for new builds, a flow line maker one that compiles pressure test records for each shipment, and a sign company one that drafts proof approvals for customers to sign.
Six things to write down for every AI role
When you add a person to the org chart, you answer these questions without thinking. For an AI employee, answer them on purpose.
- The role is one job stated in a sentence, such as "Drafts CoCs and CoOs from requests in the quality inbox." "Helps with quality" is too vague, and a vague role produces vague work from people and software alike.
- The manager is a named human, which rules out a department, IT, or the vendor. The manager reviews the AI's work, answers its escalations (the cases it hands back because they fall outside its job), and decides when it has earned more responsibility. If nobody will put their name on it, the role isn't ready.
- The permissions say which systems it can see and whether it can only read or can also write. The certificate role might read the quality inbox, the order history, and the parts list, and write only drafts. It has no business in payroll or the pricing file.
- What it can do alone covers tasks where nothing leaves the building and nothing gets committed, such as sorting incoming requests, filing attachments, and pulling an order number from an email.
- What needs approval is anything that reaches a customer, commits the company, or certifies a product. Certificates, quotes, and shipping releases need a human signature, and we recommend keeping that signature for good.
- How you measure it is a one-page weekly scorecard the manager fills in without a spreadsheet project: drafts produced, accepted as-is, corrected, rejected, and minutes from request to sent. Add a count of critical errors the reviewer caught, meaning wrong values in fields where a mistake causes harm, such as a heat number.
Example: a Certificate Desk that reports to the quality manager
A manufacturer we work with receives CoC and CoO requests in a shared team inbox. On the second org chart, that work becomes a role called the Certificate Desk, which reports to the quality manager.
The Certificate Desk reads each request as it arrives, matches it to the order, and drafts the certificate. A role like this depends on groundwork. At another client, we turned every paper and spreadsheet form into a structured digital form, where each value sits in its own labeled field instead of in handwriting on a page. That gave the company full visibility and traceability across orders: any order can be followed back to its material and forward to its shipment. That kind of record makes auto-generated compliance paperwork possible, because drafting a certificate becomes a lookup instead of a research project.
A person on the quality team reviews every draft before it goes out. The team no longer spends hours typing certificates, and the person who signs each one still decides whether it is right.
A second example shows a wider role. In a pilot with a U.S. building-products manufacturer, one agent follows each order from the mill certificates through receiving, production, the yard, and dispatch, and on into the next quote. On the second org chart, that is an Order Tracker reporting to the operations lead. Its scope covers the whole life of an order, but its authority is narrow. It keeps the record straight and calls in a person only when a decision is needed. Copy that split for your own roles, so an AI employee can see a lot and still decide very little.
What to tell a team that worries their job is next
Put an AI org chart on the wall and someone on your team will look at it and wonder if their box is next.
Take that seriously. Your crew reads the same headlines as everyone else. Pew has asked Americans the same question since 2021, and the share who are more concerned than excited about AI in daily life jumped in 2023 and has stayed at about half since.
If your crew looks like the country, about half of them are worried before you say a word, so silence will not read as neutral.
Look at what goes on the second chart: drafting certificates, sorting an inbox, tracking where an order sits. These are pieces of jobs, usually the pieces people like least. Nobody on your quality team got into manufacturing to retype heat numbers into a template.
The first chart keeps the judgment calls, like deciding whether a certificate is right, handling the customer who is upset about a late load, and spotting that a mill cert doesn't match what came off the truck. The second chart can't do that work, and every AI role on it reports to a person who does.
The drawing shows this on its own. Every line on the AI chart runs to a human, and no AI box sits above a person. Show your team that picture, name the manager of each role, and tell them what you want them to do with the hours they get back. People worry less when they can see where they stand.
If you add AI without telling anyone and hope nobody asks, rumors will fill the gap.
What this means for how you manage AI
Your first AI hire is a management decision. You have to decide who owns it, what it touches, and what it's allowed to send, which are questions you answer every time you hire a person.
Your records set the ceiling. A Certificate Desk only works if the forms behind it are digital and traceable, and an AI role can only be as good as the information it's allowed to see. One client imported more than 100,000 parts so its AI would know what the company makes and stocks.
Manufacturers say the same about their own records. In a 2025 NAM and Manufacturing Leadership Council survey, more respondents reported data and skills problems than reported using AI at all.
Both gaps land on the second org chart. The data gap decides what an AI role can see, and the skills gap decides whether its human manager can check the work.
Trust comes in steps. Start every role with a human approving everything that leaves, and watch the numbers. When a manager has seen weeks of clean drafts, they can propose loosening one specific permission, the same way you would treat a new estimator.
Put that on the calendar. The manager walks the scorecard in a 30-minute review each week, and leadership looks at every AI role in a 45-minute review each month. A permission changes only at the monthly review, with weeks of scorecards on the table.
Accountability stays with people. When something goes wrong, you walk to the desk of the manager on the first chart whose name is on the line.
Every AI role reports to a person. If nobody will put their name on it, it isn't ready.
What to do this week
You can do all of this with a pencil.
- Put your current org chart on the left side of a page. On the right, sketch one to three AI roles you would want, and draw a line from each to the named person it would report to.
- Pick the most promising role and write its job description using the six lines: role, manager, permissions, what it does alone, what needs approval, and how you'll measure it. The Certificate Desk above works as a template.
- List every system that role would need to read, and find out whether that information is digital, on paper, or in someone's head. The gaps are your real starting work.
- Show the draft to the manager you named and ask which parts of this work they would gladly hand off and what they would need to see before trusting a draft.
- Walk one person who does the work today through the chart, including the line that runs back to a human. Listen to what worries them and write it down.
Questions people ask
Who should an AI employee report to?
An AI employee should report to a named person. A department, IT, or the vendor cannot fill that role. That manager reviews its work, answers its escalations, and decides when it has earned more responsibility. For example, an AI role that drafts certificates would report to the quality manager.
What should an AI employee be allowed to do without approval?
An AI employee can handle tasks alone when nothing leaves the building and nothing gets committed, like sorting incoming requests, filing attachments, and pulling an order number from an email. Anything that reaches a customer, commits the company, or certifies a product needs a human signature. That includes certificates, quotes, and shipping releases.
How do you measure whether an AI employee is doing its job?
The manager keeps a one-page weekly scorecard: drafts produced, drafts accepted as-is, drafts corrected, drafts rejected, minutes from request to sent, and critical errors the reviewer caught. The manager walks it in a 30-minute review each week.
Terms in this piece
- Agent
- software built on an AI model that takes actions on your behalf, like reading an email or drafting a document, instead of only answering questions.
- Model
- the AI "brain" a company such as Anthropic, OpenAI, or Google trains on huge amounts of text, code, and images, such as GPT or Claude. It knows nothing about your company until you show it.
- AI employee
- an agent set up like a new hire, with its own computer, access to specific systems, a defined job, and a human manager.
- AI org chart
- a chart of your AI employees drawn next to your human org chart, showing each one's role and the named person it reports to.
- Permissions
- the specific systems an AI employee is allowed to see or change, and whether it can only read or can also write.
- Escalation
- the moment an AI employee stops and hands a decision to a person because it falls outside its scope.
