AI and data for mid-market discrete manufacturing
Your AI pilot didn’t fail. Your definitions did.
Mid-market manufacturers lose AI projects to a problem nobody budgets for: two plants, two definitions of on-time, three ways to count scrap. We fix the definitional layer first. Then we build.
A real life example · $80M manufacturer
Every Wednesday, a meeting about ship dates leaked $34,000 a year.
Nobody called it a definitions problem. They called it Wednesday.
The promise date lived in one system. The date the customer was actually told lived in another, updated by hand. When the two diverged, which was constantly, the fix was a recurring meeting.
| Wednesday alignment meeting | 12 hrs/wk |
|---|---|
| Thursday spreadsheet updates | 4 hrs/wk |
| Customer service re-keying | not counted |
| Every week | $34,000/yr |
That is one meeting. The same six people spend six to eight hours a week on reporting of the same shape, which is $76,000 to $102,000 a year. This was the first slice anyone had bothered to price.
And that is the cheap part. Between Wednesdays, somebody's customer was told a date that was no longer true.
This is what a definitional failure looks like from the outside. Inside the building it's an argument between finance and operations. Outside, it's a broken promise. The customer doesn't know or care which system it came from.
The Problem
Where manufacturing AI actually breaks
Three failures show up in almost every stalled program, and none of them are about the model you picked.
The data is there. The agreement isn’t.
Your ERP has the numbers. Finance, operations, and the plant floor each read them differently. An agent inherits every one of those disagreements and reports them as fact.
Pilots prove nothing.
A pilot that runs on a clean, hand-picked dataset tells you the tool works. It does not tell you the tool works on your data. That gap is where budgets die.
The vendor leaves. The problem stays.
Most engagements end with a system your team cannot maintain. You bought a dependency, not a capability.
Value Engineering
Starting over is the most expensive fix
Four ways to respond when the numbers stop agreeing. Three of them spend money on the wrong layer.
| What it costs you | 01Fix how work gets doneProcess change, no system work | 02Add tools or modulesMore software, same foundation | 03Start overNew system, same definitions | 04Settle the definitions firstStandard work before AI work |
|---|---|---|---|---|
| What you keepThe systems you already paid for | Everything. No system changes at all. | Your stack, plus one more license to maintain. | Very little. Replacing is the point. | The ERP, the MES, the historian. All of it. |
| What you throw awayConfiguration, knowledge, or the disagreement itself | Nothing. That is the problem. | Nothing. The disagreement gets copied forward. | Years of configuration, and the knowledge inside it. | The disagreement. Nothing else has to go. |
| Time to first usable resultSomething an operator can act on | Fast, then it drifts back. | However long the implementation takes. | Quarters, at best. | Weeks, for one value stream. |
| Cost in three yearsWhat you pay for a second time | Paid again. Nothing was standardized. | Another module on the same foundation. | The same fight, on newer software. | Extending the model, not rebuilding it. |
01
Fix how work gets done
Process change, no system work
- What you keep
- Everything. No system changes at all.
- What you throw away
- Nothing. That is the problem.
- Time to first usable result
- Fast, then it drifts back.
- Cost in three years
- Paid again. Nothing was standardized.
02
Add tools or modules
More software, same foundation
- What you keep
- Your stack, plus one more license to maintain.
- What you throw away
- Nothing. The disagreement gets copied forward.
- Time to first usable result
- However long the implementation takes.
- Cost in three years
- Another module on the same foundation.
03
Start over
New system, same definitions
- What you keep
- Very little. Replacing is the point.
- What you throw away
- Years of configuration, and the knowledge inside it.
- Time to first usable result
- Quarters, at best.
- Cost in three years
- The same fight, on newer software.
04
Settle the definitions first
Standard work before AI work
- What you keep
- The ERP, the MES, the historian. All of it.
- What you throw away
- The disagreement. Nothing else has to go.
- Time to first usable result
- Weeks, for one value stream.
- Cost in three years
- Extending the model, not rebuilding it.
You do not have a software problem. Your ERP already holds the numbers, and replacing it moves the disagreement rather than settling it. The cheapest thing on this page is the one nobody budgets for.
The Approach, Based in Lean
Standard Work Before AI Work
Lean manufacturing solved this problem forty years ago. The same rule applies to data, and almost nobody has applied it.
On the floor
You cannot improve a process you have not standardized.
In the data
You cannot automate a metric your organization has not defined.
How We Work
Six steps, in order
Each one produces something the next one needs. Nothing here is a workshop for its own sake.
We start by building the semantic model: what each metric means, who owns it, how it is calculated, and where the exceptions live. Scoped by value stream, so the decisions route to someone who can actually settle them. The output is a ratified roadmap, not a slide deck.
Roughly 80% of what we deliver is conventional business intelligence. That is not a hedge. BI earns its keep on day one and quietly builds the semantic model every agent will need later.
Assess
01
Find
We compare your spreadsheets, SQL, and reports against each other and show where the same metric resolves differently.
02
Locate
Each divergence pins to a step in the value stream. Usually a handoff, because that is where ownership changes and nobody owns the boundary.
03
Price
One divergence traced to a real customer consequence, with a number attached.
Architect
04
Ratify
A working session where every metric gets one owner and one written definition. Two definitions can survive if they serve different decisions. What cannot survive is two definitions with one name.
Adopt
05
Hold
The value stream map and the definitions live together, versioned, so the standard outlives the project and the person who wrote it.
Own
06
Monitor
We re-run the comparison against live systems and flag when something drifts from the ratified definition.
Compounding Value
Each value stream makes the last one worth more
Value streams are not the same size or shape, so this is not a story about the next one being cheaper. It is about what becomes answerable once they share definitions.
One value stream
What becomes answerable
On-time delivery means one thing everywhere. Ask two people and get the same number.
Two value streams
What becomes answerable
Was the date we promised ever achievable? Neither stream can answer that alone. It only exists in the overlap.
Three value streams
What becomes answerable
Did we miss the date because the material was late, and are we scheduling work we have no material for?
Two value streams have one overlap. Three have three. Four have six. The model is what makes those overlaps readable, and it is the part you only build once.
The Closed Loop
A standard is only real if you can see when reality departs from it.
Six months after ratification, someone builds a new report with a fifth version of on-time delivery. Under the old model, you find out when a customer does. Under this one, detection fires the week it happens.
Engagements
Three ways to start
Pick the one that matches how far along you already are.
Proof
The same failure, in three industries
Nobody could agree what “done” meant
A company-wide SAP S/4HANA program had no reliable read on its own status. The work started in Jira data hygiene, because the reporting was only as good as the definitions underneath it. A three-phase roadmap of Visibility, Transparency, and Reliability ran over nine months with the internal PMO, and the engagement extended.
A new team with no definition of demand
Leadership stood up an internal consulting group without knowing what it would be asked to do. We built a demand-based operating model: lightweight discovery to name the first services, a prioritization framework to sequence them, and the skill sets each one needed.
Two companies, one set of product data
A post-acquisition assessment covering people, process, and technology. The eCommerce and PIM work landed six months faster than the original plan, because the product data model got settled before the platform did.
Free download
The Definition Audit
Twelve questions to ask before you fund an AI pilot. Each one is a question a COO can walk into a room and ask, with a short note on what a bad answer means. Score it yourself.
Get the Definition Audit→ISO/IEC 42001
This is not a methodology. It's a control.
ISO/IEC 42001 asks for three things. The loop above is all three, running continuously.
- Clause 8Control over AI system inputs and data qualityRatify
- Clause 9Monitoring and measurementMonitor
- Clause 10Corrective action when something deviatesMonitor
Increasingly your customers are asking about it in supplier questionnaires before you've decided to care.
Questions
What mid-market manufacturers ask first
Short answers. If one of them disqualifies us, that saves us both a meeting.
Maybe not. The question is whether your organization agrees on what the numbers in it mean. If two plants report on-time delivery differently, the warehouse is storing the disagreement, not resolving it.
First Rep is one hour. An Assess engagement scoped to a single value stream produces a ratified roadmap in weeks, not quarters.
We build. The assessment exists to make the build worth doing.
Discrete manufacturers between roughly $75M and $500M in revenue. Below that, the fixed cost of a semantic model is hard to justify. Above it, you probably have an internal team.
That is the normal starting condition. It is also why the definitional work comes first.