In six months, a U.S. defense electronics manufacturer and its AI build partner went all-in together and shipped RTP.OI — a private, owned AI operating system for the finance desk, the factory floor, the workforce, and the quote desk. Humans in the loop. Controlled data in government cloud. No client data training public models.
Regal builds mission-critical electronics for U.S. defense primes. Quality and on-time delivery are why customers choose them. But like most manufacturers, the operation ran on tribal knowledge — one planner's morning ritual, one finance VP's mailbox folders, one quote meeting — and every ungoverned prompt into a consumer AI tool was a compliance incident waiting to happen.
A three-person AP team drowning in a shared mailbox, an ERP, and a separate accounting system. Service invoices that never touched a receiving report simply disappeared — and vendors chased past-due payments nobody knew existed.
10–15 min of manual search-and-match per invoice · 3–6 hrs per supplier statementRoughly 40% of work is new product introduction, so the promised dock date means nothing. The real start line is clear-to-build — the moment a kit is actually runnable — and one planner was watching it every hour of every day, by hand, across a year of backlog.
Clear-to-build monitored manually, hourlyBid/no-bid was a meeting, not a system. The bill-of-materials process alone could take seven weeks against a six-week quoted manufacturing lead time — burning the proposal team and the supply base on quotes that never converted.
~7-week BOM process vs. 6-week committed lead timeITAR and CUI obligations meant public AI tools were off-limits — but employees were already using them anyway. The company needed AI it could own, inside a boundary it could defend.
Ungoverned consumer AI use = the real riskWhat you are building for us is the complete elimination of tribal knowledge. Get that knowledge out, and then you have a platform that is process-oriented, that is predictable — and that's what buyers want. They want someone that is predictable.
RTP.OI — Regal's "operational intelligence" platform — sits on the company's own systems: ERP, MES, accounting, the AP mailbox, the quoting stack, and Active Directory. One door in, role-based permissions, and one rule set by leadership on day one: human in the loop. The system surfaces. A person decides. A person acts. The system learns.
Reads the AP mailbox, extracts 25 fields per invoice at 89–97% confidence, matches POs and receipts, handles tariff and freight lines, flags duplicates, and routes exceptions to a human queue. Non-invoice mail — vendor statements, banking-change requests, inquiries — becomes a task queue, never an automation.
Human in the loop: a clerk reviews every match; no auto-vouchering at launch; banking changes are phone-verified by a person, always.
Proof: tens of hours per week returned to a three-person finance team. The mailbox stopped being the ERP.
Material readiness, blockers, kit staging, max-quantity-buildable, split kits, customer holds with dollar values, and predicted dock dates — rebuilt around the planner's actual morning ritual. The signature feature came straight from the floor: "Will it tell me the maximum I can build right now with what's on the shelf?" It does.
Human in the loop: the planner releases kits; predicted dates are never customer commitments until manufacturing confirms.
Proof: in daily use; built to support 70+ manufacturing employees.
Training, certifications, equipment documentation, and employee-interview capture — so the next hire learns from the last expert instead of shadowing them for a year. The first module built for every single employee, and the one the VP of Operations calls half the value of the entire platform, because the company is scaling faster than it can clone experts.
Proof: launched to all 112 employees with a company memo, lunch-room learning sessions, and required trainings before module access.
Ask the company a question in English — shortages, on-time delivery, backlog, cost variances — and get an answer grounded in 26 live data tools, inside the security boundary. Tested against a 500-question evaluation harness before broad release; the client's own program managers then contributed 47 more questions, and 10+ employees volunteered as testers within a day.
Proof: accuracy improved from 67% to 91.5% across the eval set before employees ever touched it.
Bid/no-bid as a system instead of a meeting: technical-data-package intake, AI-assisted BOM extraction with human review, quoting integration, and win/loss analysis across three years of quote history. The objective isn't replacing spreadsheets — it's better bid selection, profitability, and win rates.
Target: cut 8–12 weeks of unquoted lead time out of the sales cycle.
"Successful AI adoption is 50% vendor, 50% client. What we're proving is what the two halves look like when they're both ALL-IN." — Scott Lantiegne, VP of Operations, writing to his team at the five-month mark. That line named the methodology nBrain now runs on every engagement.
No ideation workshops. Working sessions at the actual desks of the people who do the work — the finance VP walking real invoices live, the planner walking her morning clear-to-build ritual, the proposal lead walking bid/no-bid. AP went first because it had a desperate owner and non-controlled data: an invoice is an invoice. That's why production arrived in week nine, not month nine.
Data classification came before model selection: controlled data lives in Azure Government cloud, only non-controlled work touches commercial models, and client data never trains public models. Then the company named the platform, branded it internally, and held an all-hands — employees even named the chatbot — to kill the "AI is here to replace you" stigma before anything shipped.
Every module was built first as an interactive product against realistic synthetic data — the operator's actual workflow, rebuilt on a screen. The operator gave the green light. Only then did wiring to live systems begin. When the planner reviewed clear-to-build, six of her UX asks were designed or shipped the same day. Meeting transcripts became build artifacts: decisions went straight into the platform's memory.
The client-side program lead went from a 10–20% allocation to full-time — the single decision that made adoption stick. Metrics and evaluation harnesses were designed at project initiation, not bolted on: a 500-question test set for the agent, a per-module metrics sheet reviewed line-by-line by the VP of Operations, and one executive KPI everyone believes — hours saved.
Sequenced go-lives: AP in May, clear-to-build and permissions in August, workforce and the agent in September, sales behind them. Then the whole company got read-only access — explore freely, break nothing — and ten-plus employees volunteered as testers within a day. Because the client owns the platform outright, the same brain can be copied onto the next plant or the next acquisition.
In a regulated business, trust is the product. Two doctrines ran the entire build — and both were set by the client's own leadership, not by us.
Every number below comes out of the engagement record — meeting transcripts, go-live announcements, and the evaluation harness. Where something is a target rather than a banked result, it says so.
| Metric | Result |
|---|---|
| Time to first production module | 9 weeks — AP Automation live May 4, 2026, inside a six-month fixed-price engagement |
| Modules delivered | 5 business modules + shared platform layer (unified login, role-based permissions, knowledge base, evaluation harness) |
| Invoice extraction | 25 fields per invoice at 89–97% confidence, with PO/receipt matching, tariff handling, and duplicate detection |
| Finance capacity returned | Tens of hours per week across a three-person AP team |
| Agent accuracy | 67% → 91.5% across a 500-question evaluation harness before broad employee release |
| Company-wide access | All 112 employees on the platform as of September 8, 2026; 70+ manufacturing users on clear-to-build daily |
| Operating cadence | 37 working sessions in the final three months alone; weekly syncs plus monthly scope reviews with the CEO in the room, including on-site |
| Beyond contract | 7+ capabilities delivered above the original contracted scope at the client's request |
| Sales-cycle compressionTarget | 8–12 weeks of unquoted lead time targeted for removal from the RFQ process |
Anchor. Kickoff. Working sessions at the finance desk and the factory floor. The CEO sets the doctrine: after "AI," the next words are "human in the loop."
Lock. Split-cloud architecture agreed — government cloud for controlled data, commercial for the rest. The platform gets its name: RTP.OI. The finance VP green-lights the AP build after reviewing her own workflow on screen.
First production. AP Automation goes live — week nine of the engagement. The planner's clear-to-build review produces six UX asks; most ship the same day.
Install. The client-side program lead goes full-time. Monthly scope review held on-site, CEO in the room.
Waves. Clear-to-Build and role-based permissions live August 17. Workforce Intelligence follows. The agent's eval harness climbs past 91% accuracy.
Multiply. Every one of the 112 employees gets platform access. The all-hands announcement: "This platform was built by Regal people, for Regal people." Sales/RFQ enters final build. A company-specific small language model goes on the roadmap.
Five months ago, RTP.OI was a scope document and a vision. Today we walked our CEO through it line by line — and the story it tells is bigger than any single module. Build costs are trending down while capability trends up. This isn't an ending — it's the foundation.
The CEO has since repositioned the entire company around the platform — in his words, from a defense contract manufacturer to "an AI-enabled platform for mission-critical defense electronics." He rewrote the company's own LinkedIn positioning himself, on a weekend, a month before the build finished. That's what ownership looks like.
When you own AI built on your data, your competitors can never replicate it. Regal owns theirs — the platform, the data, the prompts, the agents, the workflows. nBrain designs, builds, and operates Private, Owned AI Platforms for companies that want the same.
Talk to nBrain about your build