MIT’s research puts it bluntly: 95% of enterprise AI pilots fail to deliver measurable business impact. PwC’s leadership survey finds more than half of companies see zero ROI from AI investment.
Those numbers get quoted a lot. What gets quoted less is why - and in batch manufacturing and lab operations, the reasons are structural. They have almost nothing to do with model quality.
The pilot dies in legal review, not evaluation
Cloud AI is architecturally disqualified from regulated manufacturing before anyone types a query.
OEM supplier quality agreements carry explicit data confidentiality clauses. Data residency requirements sit on top of them. So the moment a pilot proposes sending batch records, formulations or lab results to a third-party cloud, it stops being a technical decision and becomes a contractual breach.
This is why so many AI initiatives in this industry die quietly in IT or Legal. The architecture was wrong before the first demo.
Retrieval was never the hard part
The second failure is subtler. Most enterprise AI tools - Microsoft Copilot, Glean, SAP Joule - solve document retrieval within individual applications. That is roughly 20% of the problem.
The actual problem looks like this:
A SAP batch record → a LIMS test failure → a QMS CAPA → an Outlook email → a SharePoint document from another plant two years ago → a verbal agreement made during an OEM site visit.
No amount of in-app search connects that chain. It requires synthesis across systems simultaneously, which is a different architecture, not a better index.
Lab reports break standard RAG
There is a third failure that is specific to lab-driven operations.
An analytical lab report is not prose and it is not a clean table - it is both, interleaved. Test results for multiple samples and batches, across multiple test methods, mixed between tables and narrative text.
Standard RAG can find a document that mentions a parameter. It cannot reliably link that parameter to the test method that produced it, the sample it came from, or the batch that sample belongs to. And it cannot filter by batch number or parameter value.
So the analyst does what they always did: open the PDF, find the table, re-key the value, repeat.
The fix is not a bigger context window. It is a parsing agent that maps the relationships explicitly - Parameter → Test Method → Sample → Batch - and stores them as structure.
The knowledge that was never written down
The last one is the least technical and the most expensive.
Your specialists hold intelligence no system captures: why one plant needs lower spray pressure, which seasonal HVAC cycle causes silicon contamination, which deviation was approved verbally during a changeover, who to call at 9pm. When they leave, it leaves.
Hiring does not fix it - the talent does not exist, ramp is 12 to 24 months, and parallel hires just create parallel silos.
What actually has to be true
For AI to work here, four things have to hold at once:
- It runs inside your boundary. Not “encrypted in transit to our cloud” - inside.
- It synthesizes across systems, rather than searching within one.
- It understands batch and sample-level granularity, because that is the shape of the data.
- It captures the undocumented context, not just the documents.
Miss any one and you get a pilot that demos well and dies in production.
Rather than take our word for it: run the exercise. A dataset, seven queries, your own AI tools. No form, no gate.