Why AI gives manufacturers confident, wrong answers
Three quote revisions, three prices, and the AI picked the newest one. It was the wrong one. The problem is missing context, not a weak model, and the fix is a checklist any buyer can use.
Jose Acevedo · CRO and co-founder, Indax
· 4 min read
Summary. Generic AI tools answer from whichever document they find first. In a plant, the right answer usually depends on several documents that disagree. Reliable AI needs context about manufacturing objects, links between sources and a source for every answer.
Three revisions, one wrong answer
A customer asks for pricing on a part. Three quote revisions exist:
- Rev A: $42,500
- Rev B: $44,100
- Rev C: $43,800
There are also two drawing revisions. A generic AI tool takes the most recent document and answers with full confidence. The customer receives the wrong quote. The model did not fail; it had no way to know which document was correct.
The problem is context
Deloitte's 2025 Manufacturing Industry Outlook cites a survey in which nearly 70% of manufacturers name data quality, contextualization and validation as a significant obstacle to AI. The obstacle is not the model or its cost.
One order lives in five places:
- The PO is in the ERP.
- The real ship date is in a supplier email.
- The spec change is in a Teams thread.
- The drawing is a PDF on SharePoint.
- The history is in a senior buyer's head.
A general assistant sees one slice of this, often a single person's inbox. It can help one person write an email. It cannot tell the company what is true.
Where generic retrieval breaks
We tested this on a manufacturing version of the EnterpriseRAG benchmark, comparing a context layer built for manufacturing with standard retrieval:
| Question type | Standard retrieval | Manufacturing context layer |
|---|---|---|
| The ERP and an email disagree | About 40% | 93% |
| The answer spans a PO, an email and a spreadsheet | 0% | 74% |
The questions that matter in a plant are the second kind. "Which POs are at risk?" is never answered by one document. The ERP says on time, the supplier's email says 10 days late, and the answer is the gap between them.
A checklist for AI on plant data
- It understands manufacturing objects. A PO, a quote, a work order, a promise date and a supplier are different things, not just text.
- It links sources. It matches each supplier email to the right PO line in the ERP as the email arrives.
- It resolves contradictions. A supplier's email from yesterday outranks an ERP date entered three weeks ago.
- Every answer has a source. One click opens the email or record behind it.
- It works for the whole company. It respects permissions and keeps an audit trail across every inbox and system.
The model is now the easy part. The context underneath it is the work, and it is what Indax builds.
