The Manufacturing Brain: Seven Lessons on Keeping What Your People Know
Your plant makes two products. Only one leaves on a truck.
When the quote has to wait for Jim
Friday, 4:40 p.m. An RFQ lands: three hundred parts, a customer you last served in 2019, drawings that reference numbers from the ERP you retired a decade ago. The one person who can price it with confidence is Jim, and Jim is up north until Thursday. The quote waits. Sometimes the customer waits too. Often they don’t.
Every manufacturer has a version of this story, because every plant makes two things each day: product and know-how. The product gets shipped, invoiced and counted. The know-how, the judgment behind the price, the reason it’s this material and not that one, the fix that got the line moving in March, mostly evaporates. It stays in one head, or in an email thread nobody can find, or in an AI chat window that closes at five o’clock.
A useful name for the fix is a manufacturing brain: a company-owned knowledge layer built for the realities of a plant, where the knowledge lives in drawings, travelers, quotes and a few irreplaceable heads. This article is not about a product. It describes a way of working that any manufacturer can adopt, with its own team, its existing systems and whatever outside help it chooses.
The lessons of building a manufacturing brain come from a real plant. Allor Manufacturing in Brighton, part of WebsterRenold’s Industrial Motion Group, has started building one, beginning with quoting. What follows is what that work has taught so far: what held up, what surprised the team and what we would tell any plant considering the same path.
Two ways valuable knowledge leaves the business
The first door is retirement. Roughly one in four American manufacturing workers is now 55 or older, according to US Bureau of Labor Statistics data. The Manufacturing Institute projects that US manufacturing will need as many as 3.8 million new employees by 2033, and that 1.9 million of those jobs could go unfilled. Michigan feels this more than most states as manufacturing is its largest sector and its manufacturing workforce skews older than the state average. Behind every retirement sits an 18-to-24-month ramp for a replacement, if one can be hired at all. In our own operating group, we have seen a quarter of a business’s working knowledge sit in one veteran’s head.
The second door is rented AI. Your people are already using AI to solve real problems: working through quotations, interpreting specifications, comparing drawings, investigating failures. Along the way they add context, correct incomplete answers and find better ways to work. In most companies, almost none of that learning stays with the business. It disappears when the session ends, stays with one power user or has to be rebuilt the next time the question comes up. Where it does accumulate, it often accumulates inside tools the company does not control like switch tools and you start from zero.
Neither door shows up as a line on the P&L. Owners feel both anyway: quotes that take a week, new hires who take two years, a plant that struggles if three specific people call in sick. And, when it is time to sell or borrow against the business, undocumented know-how adds nothing to the valuation. A business that depends on who shows up is worth less than a business that knows what it knows.
Where Allor Started
Allor did not begin with a grand AI strategy. It began where the pain was sharpest: quoting. Years of quoting knowledge were spread across a legacy ERP, the current ERP, drawings, file shares, prior quotations and the experience of a few key employees. Pricing a returning customer’s RFQ could mean days of digging. The first goal was simple: let an estimator see prior parts, comparable work, relevant customer history, and the source behind each answer, in one place.
Seven lessons have come out of that first step. None of them is specific to quoting, and none depends on a particular software product.
Lesson 1: Start with One Expensive Workflow, Not a Platform
The temptation is to start with the big idea: connect everything, then look for uses. The better sequence is the reverse. Pick one workflow that costs real money every week, where the knowledge is scattered and a few people are bottlenecks. Quoting, troubleshooting, specification review, onboarding and technical customer response are all practical front doors.
The first workflow has to deliver measurable value on its own. The knowledge layer is what it leaves behind. Each later workflow builds on the connections already made, so the second use case should cost meaningfully less than the first. If it does not, the foundation was not built to be reused.
Lesson 2: The Most Valuable Knowledge Lives Between Systems
At Allor, the hardest questions did not sit inside any single system. They sat in the gap between the old ERP and the new one: which retired part number became which current one, which past job was truly comparable, which price was still valid. That gap is where experienced people earn their keep, and where generic AI tools fail.
A knowledge layer that closes the gap typically has three parts, each with a plain-language job:
- A connected map of facts. Parts, customers, machines, quotes, materials and people, linked to each other, with every fact carrying its source document and the dates it was valid. A price that expired in 2024 is remembered as exactly that.
- Readable reference pages. A page per part family, customer or machine, generated from the map. Every page shows an owner, a review date and its sources, so staleness has a name attached.
- A validated memory. Corrections, lessons and answers to questions the system once missed, each approved by a named person before it becomes company truth.
The design principle is capture once, serve everywhere. The connections built to fix quoting also serve customer response, onboarding and troubleshooting.
Lesson 3: Capture How Work Really Happens, Not How It was Designed
Manufacturers already have process maps, approval matrices, ISO documentation and ERP work instructions. These describe how work was designed to happen at one point in time. Day-to-day execution is messier: people move between systems, resolve exceptions, call the person who knows the customer and apply rules nobody wrote down.
At Allor, the most valuable capture came from watching where estimators actually went for answers; how experienced people connect old and new part numbers, judge which jobs are comparable, apply pricing experience, and decide when engineering needs to be involved. That operating reality, contributed by the people who do the work, is what made the tool useful.
Static documentation can support an audit. A living view of actual work can improve the business.
Lesson 4: Nothing Becomes Company Truth Without a Named Person
The learning loop should be deliberately simple. Someone asks a question inside their normal work, in the quoting screen or the RFQ inbox rather than a separate chat window. The system answers and cites its sources. When it misses, or a user corrects it, that moment becomes a candidate for review. Once a week, a curator and one rotating expert spend about twenty minutes on the handful of items worth human attention, ideally inside the collaboration tool people already use. Each item gets one of three decisions: add it to the company’s knowledge, keep it private to the person who contributed it or discard it.
One rule matters more than any algorithm. System data such as ERP records, documents and drawings can flow in with its source stamped on it. But what people say in conversations should never become company knowledge automatically. A decision by a named person sits in between. That single rule answers "who said that?" answers "what if it learns something wrong?" and keeps the knowledge base from filling with noise.
Lesson 5: Put the Experts’ Names on the Knowledge
The people who know the most are often the most wary, and with reason: nobody wants to document themselves out of a job. Two practices change that conversation. First, make credit visible. Every validated fact carries the name of the expert who approved it, and a short weekly summary ("Jim validated seven facts this week") goes to leadership. For a veteran, that is not replacement; it is a legacy and a way to mentor the next generation at scale.
Second, protect experts’ time. Keep participation opt-in, build interviews around real jobs rather than "tell me everything you know" and cap expert sessions in writing, for example a handful of 90-minute sessions in the first phase. A knowledge project must never become a tax on your best people.
The system doesn’t watch people work. People hand it what is worth keeping, and their name stays on it.
Lesson 6: Trust is Earned One Answer at a Time
One confidently wrong quote can kill trust in the whole effort, so the handling of wrong answers is not a detail; it is the design. Four practices make the difference:
- Every answer cites its sources, with a click through to the page or drawing.
- When the system is not sure, it says so. "I don’t know, ask Dave in engineering" is a good answer. It routes the question to the named owner and logs the gap for the next weekly review.
- Corrections stick, with the name of the person who made them.
- Accuracy is visible. A fixed set of fifty or so questions with known right answers is scored every week, and anyone in the company can see the trend. That line is the honest measure of whether the system is earning its keep.
Lesson 7: Ownership has to be Designed In, Not Promised
The whole point is that the knowledge belongs to the company, so ownership cannot be a slogan. It should be both architectural and contractual, whether you build in-house or with outside help:
- The system runs in the company’s own cloud account, under its own keys, with the sign-in employees already use.
- Role-based access is in place from day one, so a sales login never retrieves cost builds.
- The AI model provider does not train on company data, and that is in writing.
- Everything (the map of facts, the reference pages, the documents) can be exported at any time in open formats.
- Components are standard and proven, and the AI model can be swapped, so any competent team could take over and the knowledge outlives any one tool.
A system that is replaceable by design is one you can trust with your company’s knowledge.
Measuring What It is Worth
Nobody should pay for documentation. A manufacturing brain is worth building for one reason: leverage on skilled labor, the scarcest resource in Michigan manufacturing. That means agreeing on a baseline before any work starts. The measures below are a practical starting set; the right-hand column describes the direction to look for and your own pilot should put numbers against it.
Measure |
How to baseline it |
What improvement looks like |
Quote turnaround |
Median days from RFQ to quote over the last 90 days |
Routine RFQs answered the same day, each answer citing its source |
Estimating capacity |
Quotes per estimator per week, and the backlog trend |
The same volume handled with fewer hours; backlog shrinking |
New-hire ramp |
Months until a new hire works without shadowing |
A shorter ramp, with the knowledge layer as an on-call mentor |
Expert interruptions |
Questions per week reaching your top two or three experts |
Routine questions answered from knowledge they validated; only complex ones reach them |
Hiring bar |
Experience level required for open roles |
The option to hire one level down and close the gap faster |
Employees spend less time searching and more time doing the work well. Managers see fewer repeated interruptions and faster ramp-ups. Owners gain capacity, lower key-person risk, faster customer response and a business that is demonstrably less dependent on who shows up, which matters the day a lender, an insurer or a buyer looks at the company.
Best Practice: A 90-day Pilot
Ninety days is usually enough to learn whether the approach works in your plant. A simple structure:
- Weeks 1 to 3: baseline and connect. Sign off on the baseline measures above. Connect an ERP export and the few hundred most-used documents. Hold a small number of time-capped sessions with the experts.
- Weeks 4 to 8: go live on real work. The first tool is used on live RFQs or tickets, the reference pages open to the team and the weekly review starts.
- Weeks 9 to 12: score and decide. Measure against the baseline and the fixed question set, put the economics on one page, and decide whether to expand or stop.
Either way, whatever was built should stay with the company.
Questions to Ask Before You Build
Whether the work is done by your own IT team, a software vendor or a consulting partner, these questions separate a durable capability from a good demo:
- Can you explain, in plain English, how the system finds an answer and which parts we own?
- Where does every answer’s source live and what exactly happens when the system is not sure?
- Who approves what the system learns each week and can we see that review step?
- If we change providers or tools, what stays with us, in what format and is that in writing?
- What will it cost to run after the build? A useful sanity check: running costs should come in below the cost of one junior hire and the system should clearly outperform one. If not, ask why.
A Note for Companies that Grow by Acquisition
Acquirers often begin integration by planning ERP, CRM and financial-reporting conversions: expensive projects that can absorb leadership attention and disrupt the acquired business for months. When existing systems are adequate, there is another sequence worth considering: build a knowledge layer above the systems first. Each operating company keeps the tools suited to its products and customers, while leadership gains a shared view across businesses that were never designed to operate alike. System conversions can then happen where the business case truly warrants them. For a private equity owner, there is a second effect: a business that can answer its own questions without three specific employees in the room is easier to defend in diligence.
Local intelligence where the work happens. Shared visibility where leadership needs it.
Why Now
Manufacturers already hold the knowledge needed to answer customers faster, develop employees sooner and depend less on a few individuals. Most of it is fragmented, informal or rebuilt from scratch each time. Three things make this the moment to act. AI models have become good and affordable enough that the model is no longer the constraint; context is. A knowledge layer the company owns can serve whichever tools come next. And the retirement wave is not waiting: the know-how walking toward the parking lot this Friday does not come back.
The tools will change. The knowledge your people create using them should stay with the company and become more valuable over time.
About the Authors
Fabien Cros is Chief Data & AI Officer at Ducker Carlisle in Troy. He previously served as Data & AI Country Lead for Manufacturing at Google Cloud and spent seven years with Accenture in Detroit, focused on the industrial and automotive sectors. He may be reached at fcros@duckercarlisle.com.
Shane Marcheson, President of WebsterRenold’s Industrial Motion Group, which includes Allor Manufacturing in Brighton.
Ducker Carlisle is an MMA Premium Associate Member and has been an MMA member company since October 2026. Visit online: duckercarlisle.com
Sources: US Bureau of Labor Statistics workforce data; The Manufacturing Institute workforce outlook (2024); State of Michigan employment and output figures (BLS/BEA, 2025) and the 2026 Michigan Manufacturing Industry Workforce Analysis.