Watch your best estimator for an hour. She opens an RFQ package, reads the drawings, and then spends most of the morning typing part numbers into the quote sheet and material grades into the costing spreadsheet, filling in quantities, finishes, and tolerances one cell at a time.
The part of her job that makes you money is the ten minutes where she looks at a weldment and says, "This will fight us in the fixture. Add two hours." The typing is what wears her out, and many of the skilled people in your office spend their days the same way.
What "verify, don't retype" means
When AI does the first pass, it does the reading and retyping. It pulls the values out of the documents, fills in the draft, and marks anything it is unsure about. Then a person checks the draft, fixes what is wrong, makes the calls that need experience, and signs off.
People stop building documents from scratch and spend their time checking and deciding. We call it verify, don't retype. The pile of copying shrinks, and the part of each job that needed a person in the first place gets bigger.
How three jobs change
The estimator
Before: she builds every quote from a blank sheet. She reads the RFQ, pulls the drawings, types the line items, looks up material, estimates hours, checks the whole thing, and sends it.
After: the RFQ comes in with the first pass done. The software has pulled the line items from the drawings, matched material to your price list, and attached past quotes for similar parts. Anything it could not read clearly, like a smudged callout or a note that contradicts the title block, sits flagged at the top.
She reviews the draft and fixes the two line items it got wrong. She adds the two hours for the fixture, because the software does not know your fixtures the way she does. Then she sets the margin and sends it.
The quality manager
Before: she checks every compliance packet page by page, comparing each heat number on each MTR against the receiving log and each grade against the PO. Most of it matches. She checks it anyway, because the one that does not match comes back as a customer complaint.
After: the software builds the packet and checks every value against its source document. It marks matching values as matched, with a link to where each came from, and puts mismatches at the top of her list: this heat number on the MTR does not appear on the receiving form, this grade differs from the PO.
She reviews the flagged mismatches first, decides what each one means, and spot-checks the rest before she signs. Her signature means what it always did, and she spent her attention where the risk was.
The coordinator
Before: she lives in the shared inbox, where cert requests, PO changes, drawing revisions, and "where's my order" emails arrive mixed together. Most of her day goes to reading and sorting before any real work starts.
After: the software reads, sorts, and routes incoming requests. It matches a routine cert request to the job and drafts a reply, and it summarizes a change order and sends it to the right person. The coordinator gets what does not fit: the customer who wants partial shipment on a job that is not split, or the rev change on a part that is already cut.
Those exceptions are where you win or lose customers, and they were always the part of the job that needed her.
What gets harder: reviewing well takes training
Checking well is a skill. Reviewing a draft is different from building one. When you build from scratch, you think about every line because you have to. When you review, it is tempting to skim, especially when the draft is right nine times in a row. The tenth draft is the one that matters, and it will look exactly as confident as the other nine, because the software sounds just as sure when it is wrong.
Teach your reviewers to work in the same order every time:
- Flags first. Anything the software marked as unsure or mismatched gets looked at before anything else.
- Critical fields next. These are the values where a mistake causes real harm. On a cert packet that is usually the PO, the part number and revision, the heat number, and the country of origin. On a flow line pressure test record, it is the test pressure and hold time. Check each one against its source, even when it is not flagged.
- A random spot check. Pick a few unflagged values a day and check them anyway.
- A one-line reason for every fix, such as "wrong rev, customer sent B on the 12th." Those lines show you which mistakes keep coming back.
Then watch for rubber-stamping. A reviewer who approves 40 drafts in ten minutes without opening the attachments has turned the safety step into decoration. Track minutes per review alongside the fixes, and a number that drops close to zero will usually show it first.
Few manufacturers train for it yet. In NAM's second-quarter 2026 survey, about a quarter of manufacturers trained frontline workers on data, analytics, or reading AI outputs, which is the reviewing skill this whole setup depends on. Put your reviewers through that training and you are ahead of roughly three in four manufacturers.
Source: NAM Q2 2026 Manufacturers' Outlook Survey, question 19. Respondents could pick more than one answer. 7.62% were uncertain.
Accountability also stays with people. Your quality manager signs the CoC and your estimator commits to the price. If a bad packet ships, your customer will not accept "the AI did it," and you should not accept it either.
That is the right design, but it means people have to trust the tools enough to use them and distrust them enough to check. The review order above is how you hold that balance on a busy day.
What gets better for your people
The hours come back as the typing, cross-checking, and sorting that burned people out shrinks.
People also have more attention left for judgment. An estimator who is not exhausted from data entry looks harder at the weird part, and a quality manager who is not checking three hundred matching values catches the one that does not match.
We also think the work becomes easier to learn. A new estimator who reviews drafts with every value linked to its source sees where each number came from, which a blank sheet and a stack of drawings never shows. The experienced people can spend their time teaching judgment instead of teaching data entry.
Roles shift toward the harder problems. The receiving clerk becomes the person who resolves receiving exceptions, and the coordinator becomes the person who handles the customer's hardest questions.
When your people ask whether AI will replace them
Your people will ask whether this is how you replace them, a few of them out loud and most of them to each other in the parking lot.
They hear the same news everyone else does. In Pew Research Center's June 2026 survey, 52% of US adults said AI makes them more concerned than excited, up from 37% in 2021, and 71% expect it to mean fewer US jobs over the next 20 years. Assume most of your floor already leans that way.
Answer it before they have to ask, and answer it honestly: the software does the first pass, and people verify, decide, and sign. We build the system that way on purpose. A person has to approve each item before it goes out the door, and the person who approves it is accountable for it.
Then say what you are going to do with the hours that come back. Look at your own backlog first. In the same NAM survey, 46.95% of manufacturers named attracting and retaining workers as a challenge. If you are one of them, the hours go to the work you have no one for today: quoting the RFQs you currently pass on, catching problems before they ship, and calling back the customer with the question nobody else can answer.
Do not promise what you cannot keep. If you are not sure how roles will shift in two years, say so. People can handle "I don't know yet, and I'll tell you when I do." They cannot handle finding out later that you knew.
Words to use when you tell your team
A few lines you can use:
- "The software does the typing. You do the thinking."
- "Nothing goes out without your name on it, same as today."
- "If it gets something wrong, you're the one who catches it. That's the job."
- "We're doing this so you stop drowning in data entry, not so we can do without you."
- "Tell me what's annoying about it. You're the one using it."
Avoid consultant words like "efficiency," "optimization," and "headcount." They land badly even when you do not mean them the way people hear them.
Measure reviewers by what they catch
The harder decision is what you ask your people to do all day, and the software is the easier one. If you roll out AI and keep measuring people on how many pages they typed, you will get skimmed reviews and resentment. Measure what they caught instead: the drafts they fixed, the one-line reasons, and the critical errors that stopped at their desk.
Your most experienced people become more valuable, because they know what a wrong answer looks like.
What to do this week
- Shadow one estimator, one quality person, and one coordinator for an hour each, and write down how much of that hour goes to typing, copying, or sorting.
- Ask each of them, "If the typing were done for you, what would you spend the time on?" Write down the answers.
- Pick the one task where a mistake costs the most, and list its critical fields. That list is where review training starts.
- Draft your own version of the message above and change it until it sounds like you. Hold it until you know what you are rolling out and when.
- For quotes, packets, and customer replies, write down which person is accountable for approving each one.
Questions people ask
What does human in the loop mean for AI in manufacturing?
It means a person reviews and approves the AI's work before it counts. The software drafts quotes, cert packets, and replies, and a person checks them against the source documents and signs off. Nothing goes out the door until that person says yes.
How does AI change an estimator's job?
With AI doing the first pass, an RFQ arrives with line items pulled from the drawings, material matched to the price list, and similar past quotes attached. The estimator reviews the draft, fixes what is wrong, adds the judgment calls the software cannot make, sets the margin, and sends it. The estimator types far less and keeps every judgment call.
How should I tell my employees about AI?
Tell them before they have to ask: the software does the first pass, and people verify, decide, and sign. Say what you plan to do with the hours that come back, and do not promise what you cannot keep. Avoid words like efficiency, optimization, and headcount, which land badly on the floor.
Terms in this piece
- First pass
- the initial draft the software produces by reading documents and filling in values, before any person reviews it.
- Verification
- a person checking the software's draft against the source documents and correcting what is wrong.
- Flag
- a mark the software puts on a value it is unsure about or that does not match another document, so a reviewer checks it first.
- Critical field
- a value where a mistake causes real harm, such as the heat number or the part revision on a cert. The reviewer checks it every time, flagged or not.
- Approval step
- the point where a person must say yes before anything leaves the building. The software cannot skip it.
- Human in the loop
- a setup where a person reviews and approves the software's work before it counts.
- AI
- software that can read, write, and make first-pass decisions from documents and instructions, instead of only following fixed rules.
