AI PPAP Review: How Omnex’s O-BOT Automates 19-Document PPAP Validation

A PPAP submission is nineteen documents deep — and most companies still hand that review to whoever has the time, not necessarily the experience.

That’s the quiet risk in a lot of supplier quality programs. Reviewing a Production Part Approval Process package well means cross-checking a 2D drawing against a process flow diagram, a PFMEA, a control plan, a dimensional layout report, and a process capability study — and catching it when any two of those disagree with each other. It takes real judgment. Too often, it’s given to the newest person on the team, simply because the most experienced engineers don’t have nineteen documents’ worth of hours to spare.

This is part two of our webinar series on digitalizing product development and the shop floor. Part one covered the digital thread from DFMEA to the shop floor. This time, Chad Kymal and Antony John walk through the other half of that thread: turning PPAP review into the foundation of shop floor digitalization, using O-BOT, the AI agent built into AQuA Pro.

Why PPAP Reviews Are Such a Bottleneck

Ask any supplier quality team what slows PPAP down, and the answers are remarkably consistent:

• Your most experienced people should be doing this — and usually aren’t. Reviewing all 19 PPAP documents well takes real expertise, but that review often lands on whoever has bandwidth, not whoever has the background.
• Tracking APQP and PPAP projects is its own job. Visibility into where a project actually stands is hard to get, especially across a large organization.
• The data lives everywhere except one place. In large companies, different pieces of a PPAP package are produced in different systems and different plants, and someone has to pull it all together before a review can even start.
• Collecting and evaluating it all is slow. Even once the data is in one place, working through it document by document is a heavy time investment.
• There’s a shortage of people who can do this well. Technically skilled reviewers are hard to find and harder to scale across every supplier and every program.

The logic behind Omnex’s approach is simple: if AI can review a PPAP package, it can review a software PPAP package too. The AI takes in unstructured or structured data — a scanned drawing, a spreadsheet, a PDF — runs it through O-BOT, and returns an approve-or-reject-style analysis, the same way an experienced reviewer would.

How AI Actually Reviews a PPAP Package

It starts with the drawing. A typical 2D print carries a lot of critical and significant characteristics, bubbled dimension numbers, and material specifications — and O-BOT reads all of it directly off the drawing that was actually submitted in the PPAP. For every characteristic, it identifies the characteristic type, the nominal value, the plus/minus tolerance, whether it’s flagged as a key characteristic, the drawing revision number, and which sheet it appears on.

Omnex has been building this specific capability for four years. The addition of large language models — the current implementation uses Gemini, though the approach is LLM-agnostic — gave it a real jump in graphical extraction: pulling text boxes, bubble numbers, and GD&T symbols straight out of the drawing, not just the text around it. On top of that extraction sits a library of roughly 300 configurable rules that can be turned on or off. One example: any note on a drawing is expected to carry a bubble number, and if it doesn’t, O-BOT flags it.

Then it checks everything against everything else

Once the drawing is read, O-BOT compares it against the rest of the package, layer by layer:

• Drawing vs. process flow diagram: confirms the same part number and revision number appear in both, and checks that significant and critical characteristics for that supplier aren’t missing a bubble number.
• The FMEA: checks the header for missing information, incorrect RPN calculations, missing failure modes, and incomplete actions. For the top 20% of RPN values, it specifically flags a missing responsibility, missing target date, or missing recommended action — in red, since a high RPN with no action attached is the highest-risk gap in the document. It can also catch a corrective action that should really be a preventive action (or vice versa), a weak failure mode or cause statement, or “column switching,” where someone has typed a cause into the failure-mode field. Omnex trained this against hundreds of real FMEAs, using both machine learning and LLMs.
• The control plan: flags a missing control plan number, missing revision date, missing part name, missing supplier plant, a missing bubble number, or special characteristics that don’t match the drawing.
• The dimensional report: checks for at least five samples (the rule requires five), missing specifications, a missing design record level, and a missing note on where the inspection was performed.
• The initial process study / process capability study: flags a missing gauge, missing characteristics, blank PPK validations, and a PPK below the expected 1.67. If a six-pack report from a common statistical package is uploaded, O-BOT also checks whether the distribution is normal.

In a live demo of this traceability check, the AI reviewed a print in any form, identified its GD&T symbols and special characteristics, then read the process flow diagram — even as a PDF — and flagged real discrepancies: two characteristics marked in the print weren’t marked as special in the process flow diagram, even though they should have been, and several balloon numbers present in the print were missing from the process flow diagram entirely. The same cross-check ran against the PFMEA, catching a characteristic that was marked special on the print but never tagged special in the PFMEA. All of this works across PDF, Word, Excel, scanned or printed documents, and different balloon styles — squares, handwritten marks, whatever a given supplier actually used — because the underlying LLM can interpret them.

Bringing Suppliers Into the Review, Before They Submit

The second half of the demo showed how this extends past your own team and into your supplier base. A supplier quality engineer starts by initiating a PPAP project — a fully configurable “project charter” — filling in the product information and the current PPAP level, which automatically establishes every level of the PPAP from the template. An email notification then goes out to the supplier.

The supplier logs into a branded landing page and uploads the drawing first. O-BOT reviews it immediately, extracting the balloon print elements and interpreting GD&T symbols and notes as structured, reusable data rather than an image. From there, the supplier uploads each remaining document one at a time — and once everything is in, they can trigger an AI review of the entire PPAP packet themselves, before it ever reaches your team. A Part Submission Warrant (PSW) form can be filled in directly online as part of the same flow.

The point of building it this way is to shift ownership onto the supplier: they find and fix their own findings before submission, instead of your team discovering them after the fact. Once every PPAP element has been processed this way, the package is what Chad Kymal calls “data rich” — every document has been converted into structured data and reviewed by AI, not just filed away as a PDF.

Some OEMs are also opening up their own customer portals to accept inspection data directly through connectors. Omnex’s NetSync connector platform can be configured to move data into a customer portal — for inspection data, and potentially for the PPAP submission itself.

What This Actually Changes: 450x More Efficient Review

The benefits break into two groups. Inside your own company, your PPAPs become fully digitalized — you know everything about your own product and your suppliers’ product, and if you’re already using Omnex software for problem-solving, you also know what’s actually failed in the field, which gives you a genuinely large body of data to analyze going forward.

Across your supply base, the review itself becomes 100% coverage instead of a sample, and dramatically more efficient — Omnex calculates it at roughly 450 times more efficient than a manual review. Companies that have adopted it are seeing ROI in a matter of months, not years.

Licensing, Data Segregation, and Security: What Buyers Ask First

Four questions came up in the live Q&A that most quality leaders ask before they’ll bring an AI reviewer anywhere near their supply base:

How do you manage licensing when suppliers are involved?

Suppliers come into a portal that’s completely branded for your organization, with their own logins authenticated through multi-factor authentication (MFA). Supplier licenses are priced separately from — and relatively inexpensively compared to — internal user licenses.

How is data segregated between different suppliers?

Each supplier’s data is configured just for that supplier; by default, one supplier cannot see another supplier’s data. The rules and models themselves are generic, but the evaluation of those rules against a given supplier’s data stays confined to that supplier — the AI does not transfer learning from one supplier’s data to another’s. That control operates both at the supplier access level and inside the AI itself.

Can the rules be customized, including by division?

Roughly 300+ rules ship pre-programmed, built on AIAG standards and 40 years of Omnex’s own best practices as a knowledge bank. Every rule is configurable — you can turn individual rules on or off, and where a rule needs to be adjusted to fit a specific division’s requirements, that customization can be done.

How is proprietary drawing data protected in the cloud?

O-BOT supports both a cloud implementation (on your own cloud, or whatever cloud solution you choose) and an on-premises deployment inside your own four walls. Whatever protection you already apply to your drawings or supplier PPAPs today carries over to this system. The underlying models are never exposed to general-public AI models — each deployment runs on a virtual private cloud dedicated to that company and its suppliers alone.

Looking ahead, Omnex is extending this same agentic approach to automatic DFMEA and PFMEA generation — where an engineer can simply tell the AI to add or change a failure mode and watch it happen — plus 8Ds and audits. We covered the FMEA side of that roadmap in our companion article on agentic AI for FMEA.

Frequently Asked Questions