Articles / Explore / Smart Manufacturing
Can AI safely configure a custom product when sales, engineering and production describe it differently?
No. It can produce an answer faster, but it may be using the wrong version of the product rules. Before adding AI to quoting or configuration, a manufacturer needs one governed definition of valid options, constraints and versions that every downstream system can trace.
What changed in the discussion
On September 15, 2026, the Manufacturing Leadership Council published an analysis by Nils Olsson arguing that configurable products often exist as separate engineering, sales and manufacturing definitions. The article describes the resulting reconciliation work and says generative AI can make configuration easier only when its proposals are checked against governed rules.
That is a useful warning, but its percentages should be treated carefully. Olsson is Tacton’s chief strategy officer, and the page does not link a survey instrument, sample or methodology. This article therefore uses the report to identify the operational problem, not as an independently verified market benchmark.
The problem is version conflict, not just bad data
Consider a fictional industrial-pump maker. Engineering updates rule R-42 after learning that a chemical-duty seal cannot be paired with motor M7. Sales still quotes from R-39. The factory’s routing system holds R-40.
An AI quoting assistant reads the sales catalogue and recommends the incompatible seal-and-motor combination. Its wording may be confident and its retrieval may be accurate: it faithfully found an obsolete rule. The failure comes from conflicting product definitions, not necessarily from model hallucination.
NIST’s Digital Thread for Smart Manufacturing project described the underlying issue years earlier. Its project record says gaps and one-way information flows prevent enterprise-wide use of product information. NIST’s proposed response included common information elements, standards, trusted information in context and feedback from inspection to design. The project concluded in 2018; its page was updated on March 26, 2025.
The Digital Twin Consortium offers a complementary definition: a digital thread is bidirectional, dependable and trustworthy, connecting information across lifecycle stages. That means a useful thread is more than a shared folder. It preserves relationships, versions and feedback.
One rulebook does not require one database
A manufacturer can keep specialized engineering, quoting and production systems. The requirement is narrower: one authority must own each configuration rule, give it a version and expose an unambiguous translation to the other systems.
For every AI-generated proposal, a reviewer should be able to answer three questions:
- Which rule version approved this option combination?
- Did the quote, engineering definition and build record use the same option identifiers?
- Where does a factory exception return for engineering review?
If those answers are unavailable, adding a conversational interface hides the disagreement rather than resolving it.
A 20-order check before automation
Start with a small baseline. Select 20 recently completed configurable orders. For each order, compare the option identifiers and rule version in the accepted quote, engineering definition and final build record.
In the fictional example below, 15 orders match across all three records and five contain at least one mismatch. The alignment rate is 15 divided by 20, or 75%.
That does not mean 25% of products failed, nor does it predict return on investment. A mismatch may be an approved exception, a translation error or stale documentation. It means five orders need classification before AI is allowed to reuse their logic.
What this changes for AI decisions
First, AI readiness becomes traceability work. Teams should not ask only whether the model can understand a buyer’s request. They should ask whether every recommendation resolves to a current, approved rule.
Second, feedback matters as much as distribution. Engineering rules must reach quoting and production, while exceptions, inspection results and field findings must travel back with context. A one-way export becomes stale by design.
Third, the first metric is agreement, not speed. Quote time may improve while unbuildable combinations rise. A cross-system alignment rate gives leaders a baseline that the next pilot must protect.
Three takeaways
- AI should reference a governed product rule and its version, not whichever departmental copy is easiest to retrieve.
- A digital thread must carry feedback as well as instructions.
- Measure cross-system agreement before claiming that faster configuration creates value.
One next action
Audit 20 completed configurable orders this week. Record the quote, engineering and build rule versions; classify every mismatch as an approved exception, translation defect or stale record. That creates a measurable baseline for any AI configuration pilot.
Limitation
The recent MLC article is vendor-authored and does not expose its survey methodology on the page. NIST’s cited project is foundational rather than a current deployment study, and the pump example and 20-order audit are fictional. A shared rulebook improves traceability; it does not by itself prove integration quality, safety or financial return.
Related reading
Continue with Industry AI Decision’s analysis of the systems synchronization behind a flow factory and the evidence gates for a factory robot pilot.
Sources
- Nils Olsson, “Scale and Accuracy with the Buyer-Centric Smart Factory,” Manufacturing Leadership Council, published September 15, 2026.
- NIST, “Digital Thread for Smart Manufacturing,” project concluded 2018; page updated March 26, 2025.
- Digital Twin Consortium, “What is the Digital Thread?,” accessed September 23, 2026.
Leave a Reply