MPI, a 54-year manufacturer of investment-casting (wax-room) equipment with about 80 employees, deployed Annora as an AI layer on top of its existing ERP, MES and knowledge base. Employees with decades on the floor began using it on day one, with no formal training, turning more than 400 combined years of know-how, once scattered across systems and people, into one place anyone can ask.
- Company
- MPI. Manufacturer of wax-injection equipment for the investment-casting industry, in business 54 years, ~80 employees, selling globally.
- Before Annora
- 400+ combined years of know-how scattered across the ERP, spreadsheets and staff memory. Getting an answer meant knowing exactly who to ask.
- What it cost
- Time lost interrupting three or four veterans for routine answers, and the constant risk of that knowledge retiring out the door for good.
- With Annora
- An AI layer added on top of the existing knowledge base, ERP and MES. No rip-out, no training. Anyone asks one place and gets a sourced answer; veterans are freed to mentor.
- Since then
- That layer became the foundation. MPI's quoting system already runs on it, a customer service system is in progress, and automated procurement and field service reporting build on the same base next.
For 54 years, the answer always existed. Somewhere, in someone
MPI's most valuable asset, the judgment of its long-tenured staff, lived in three disconnected places: the ERP, spreadsheets, and people's heads.
MPI builds the equipment for the wax room, where investment-cast metal parts begin, and sells it globally. With about 80 employees carrying more than 400 combined years of experience, the company never lacked for answers. The trouble was finding them: getting a reliable one meant knowing exactly who to interrupt, walking the floor to track that person down, and hoping they weren't mid-pour.
“We were over 400 years of experience with multiple people… how do we make sure we don't lose it?”
- 54 years in business; ~80 employees carrying 400+ years of combined know-how.
- Answers split across ERP, spreadsheets and individual memory.
- Every quote or spec meant tracking down the one person who knew.
- The real risk: that knowledge walking out the door at retirement.
So MPI gave that knowledge one place to live
Annora went on as a layer over the tools MPI already ran, so adoption needed no rollout project and no formal training.
Grounded in MPI's knowledge base, ERP and MES, Annora added capability without ripping anything out. And without a new device on the desk. People started using it on day one, including employees who had been on the floor for decades.
“The system is so easy to use that getting better with it just requires playing with it. Once they realize they're not going to break anything, it's kind of no holds barred.”
As Aaron puts it, the tool is not going to cost anyone their job. It's going to support their job, make their lives easier, and give them the information they want faster. Some of the people he least expected to dive in did exactly that.
Now the answer comes back before the walk to find it would have
Anyone at MPI can now ask a single place a plain question and get an answer pulled from across the business. Quotes, drawings, memos and orders.
The change shows up in the small moments that used to cost the most time, and it compounds: the more the team relies on Annora, the more value it returns.
“I can't tell you how many times I'd interrupt three or four people to find the right template. Now I just ask the system and it's right there. In less time than it would've taken me to find the thing.”
The best of both worlds.
“It gives us a lot more time to train our young team and lean on our knowledge masters.” MPI's veterans are not replaced. They are amplified, and freed to spend their time where it matters most.
The same layer now runs MPI's quoting. And that was just the start
What MPI adopted was never one tool. It's the foundation. The quoting system already runs on it, and more is being built on the same base.
The knowledge layer holds MPI's SOPs, its product information, and the records that make the business run. That is exactly the context an AI agent needs to get an answer right. So once the layer was in place, the next system didn't start from scratch. The quoting system (CPQ) went live on top of it, configuring and pricing a wax-injection cell from the same knowledge the floor already trusted. A customer service system is being built on it now, with automated procurement and field service reporting next. Each one builds on the same base instead of standing up its own.
- Quoting system. Live now. Configures and prices a wax-injection cell from the same knowledge the floor already trusts.
- Customer service. In progress. Answering on the same layer.
- Automated procurement. Next. Ordering on the same layer.
- Field service reporting. Next. Reporting on the same layer.
Everything sits on top of the ERP, so data gets entered once
Each system reads from and writes to the tools MPI already runs. Enter a fact once and every system has it.
This is the part most software gets wrong. When the same order has to be keyed into the ERP, then a spreadsheet, then a quoting tool, two things happen. Error rates climb, because every re-entry is another chance to fumble a number. And the workload doubles, because someone is doing the same job twice. MPI's layer removes both. It sits on top of the ERP and the other systems and works with them natively, so a fact lives in one place and shows up everywhere it's needed. The AI agents that do the work draw their context from that same layer: the SOPs, the product data, the essential company records. That is why each new system can be trusted to get the details right.
Aaron had one hard requirement. The IP never leaves the building
Aaron's requirement was blunt. Whatever MPI ran had to be a black box that protected the company's IP, and its customers' IP too.
MPI builds equipment for military and aerospace work. The designs it touches, its own and its customers', are not allowed to leak, and they are certainly not allowed to train someone else's model. So in the video interview Aaron drew a hard line. He needed a black box. Nothing about MPI's know-how could go out the door. That requirement is what shaped how the layer was built.
So we built it as a black box. The layer runs on open-source models, privately hosted inside MPI's own security perimeter. Your data never leaves your environment. Nothing trains on it, and nothing about your designs is exposed to anyone outside your walls. That protects MPI's IP. And it protects the IP of MPI's customers, which for a supplier to military and aerospace programs is not a nice-to-have. It's the whole deal.
“We needed a black-box AI system that protects our IP and our customers' data from public AI systems.”
Ask us anything
What problem did MPI solve with Annora?
MPI's institutional knowledge, more than 400 combined years of experience, was scattered across its ERP, spreadsheets and the memory of long-tenured staff. Getting an answer meant knowing exactly who to ask, and that knowledge risked being lost at retirement. Annora consolidated it into one place anyone can query.
How much training did MPI's Annora rollout require?
None. Annora was added as a layer on top of MPI's existing knowledge base, ERP and MES, with no rip-and-replace. Employees, including veterans with decades on the floor, began using it on day one with no formal training.
What else has MPI built on the Annora intelligence layer?
The layer is a foundation, not a single tool. MPI's quoting system already runs on it, configuring and pricing wax-injection cells from the same knowledge the floor trusts. A customer service system is in progress on the same layer, with automated procurement and field service reporting next, each built on the same base rather than standing up its own system. Because everything sits on top of MPI's ERP and existing systems, a fact is entered once and shows up everywhere it's needed, so error rates stay low and nobody does the same job twice.
How does Annora protect MPI's IP and its customers' IP?
MPI supplies military and aerospace programs, so protecting design IP was a hard requirement. The layer is built as a black box: open-source models, privately hosted inside MPI's own security perimeter, with data that never leaves MPI's environment. Nothing trains on MPI's data, and nothing about its designs is exposed outside its walls. That protects both MPI's own IP and the IP of MPI's customers.