Boeing Cuts App Build Time to 48 Hours with Enterprise AI
Ai4
Boeing scaled internal Generative AI to 38,000 daily active users by deploying proprietary models within a private cloud to eliminate data leaks and hallucinations.
## Background
Published by Ai4 on YouTube on August 28, 2025, this video captures a presentation delivered by Abhi Seth, Chief Enterprise AI & Data Officer at Boeing. According to the page description, the session focuses on exploring the inside of Boeing's GenAI strategy from the cockpit perspective.
## TL;DR
Abhi Seth details Boeing's enterprise-wide Generative AI implementation, addressing how the 108-year-old aerospace giant balances rapid technology adoption with strict security, zero-trust microservices, and hallucination controls. The presentation covers key milestones like rolling out Boeing Conversational AI to 170,000 employees, establishing a GenAI Academy, and reducing custom app deployment time to under 48 hours. Seth highlights tangible high-impact applications across engineering, shop-floor manufacturing non-conformance reports, legacy drawing computer vision enhancements, and automated safety risk classification.
## Atmosphere & style
The session carries a highly professional, pragmatic, and authoritative tone tailored for an enterprise technology audience. Seth presents with measured confidence, balancing the vast legacy and safety-critical nature of aerospace engineering with forward-looking technical innovation. The atmosphere feels grounded and structured, avoiding speculative AI hype in favor of concrete metrics, architectural diagrams, and real-world operational impact.
## Key takeaways
- [00:04](https://www.youtube.com/watch?v=5C5tSNcKaSs&t=4s) Boeing leverages a 108-year aerospace legacy to deploy enterprise Generative AI across 170,000 global employees and 62,000 engineers.
- [00:32](https://www.youtube.com/watch?v=5C5tSNcKaSs&t=32s) Aerospace constraints require zero tolerance for hallucinations, mandatory source attribution, and bidirectional linking for engineering specifications.
- [00:58](https://www.youtube.com/watch?v=5C5tSNcKaSs&t=58s) To prevent data leakage, Boeing built a Zero Trust microservices architecture integrating enterprise IAM and hosting LLMs in a private cloud.
- [00:58](https://www.youtube.com/watch?v=5C5tSNcKaSs&t=58s) Boeing deployed its internal Conversational AI platform to 170,000 employees in 12 weeks, with adoption reaching 38,000 daily active users across 53 countries.
- [10:03](https://www.youtube.com/watch?v=5C5tSNcKaSs&t=603s) Standardized architecture templates allow Boeing functional teams to build and deploy custom GenAI applications in under 48 hours.
- [12:06](https://www.youtube.com/watch?v=5C5tSNcKaSs&t=726s) RAG implementation unlocked 7 million engineering documents across 28 share drives, providing instant paragraph-level source citations.
- [17:13](https://www.youtube.com/watch?v=5C5tSNcKaSs&t=1033s) GenAI is prohibited directly on aircraft flight control systems, but is heavily used on the factory floor to classify safety and quality risk reports into 14 categories.
- [18:42](https://www.youtube.com/watch?v=5C5tSNcKaSs&t=1122s) Data classification mechanisms enforce ITAR and clearance rules at a granular document level without blocking global workforce collaboration.
- [20:23](https://www.youtube.com/watch?v=5C5tSNcKaSs&t=1223s) To eliminate shadow IT risks and secure corporate IP, external public LLM access was disabled in favor of internal tools updated within 6 hours of new model releases.
- [23:08](https://www.youtube.com/watch?v=5C5tSNcKaSs&t=1388s) Computer vision tools automatically enhance faint legacy engineering drawings directly inside mechanics' software, handling 25,000 requests per day.
- [25:12](https://www.youtube.com/watch?v=5C5tSNcKaSs&t=1512s) While rapid RAG builds take 48 hours, Boeing conducts 2 to 3 weeks of SME Q&A validation before releasing applications to production.
- [27:36](https://www.youtube.com/watch?v=5C5tSNcKaSs&t=1656s) Modular RAG architecture separates vector databases from foundation models, enabling seamless LLM upgrades without disrupting underlying enterprise data.
## Detailed walkthrough
### [00:04](https://www.youtube.com/watch?v=5C5tSNcKaSs&t=4s) Boeing's 108-Year Legacy and the Scale of Enterprise AI
Abhi Seth, Chief Enterprise AI & Data Officer at Boeing, opens the session by reflecting on the company's 108-year history defining the future of aviation, spanning commercial aerospace, defense systems, and space travel. To set the backdrop for Boeing's Generative AI journey over the past two years, Seth outlines the company's core mission—connecting, protecting, and exploring the world sustainably—across its three primary business divisions. Commercial Airplanes has been redefining flight since the first Boeing Model 1 (B&W) seaplane flew in 1916. Defense, Space & Security provides national security deterrence, highlighted by their work on the F-15EX Eagle II / F-47 sixth-generation fighter jet capabilities. Global Services acts as a customer service champion, maintaining fleet safety and minimizing operational downtime for both commercial airlines and military customers.
Seth emphasizes that while AI hype shifted toward agentic workflows in 2024, Boeing's involvement in advanced technology spans decades. The aerospace giant has relied on physics-based modeling, virtual reality, and simulation for many years, alongside traditional AI and machine learning investments. However, implementing Generative AI across a massive global workforce—comprising 170,000 employees in 65 countries, including 62,000 engineers—presents a distinct governance challenge that requires balancing rapid technological adoption with strict responsibility, ethics, transparency, and model robustness.
- Every 2 seconds, a Boeing commercial aircraft takes off or lands somewhere across the globe.
- Boeing aircraft have transported approximately half of the world's population over time.
- Total global headcount spans 170,000 employees, featuring a technical base of 62,000 engineers.
[00:32](https://www.youtube.com/watch?v=5C5tSNcKaSs&t=32s) Core Challenges: Precision, Security, and Traceability in Aerospace
When approaching Generative AI in the highly regulated aerospace domain, a "good enough" or plausible-sounding answer is completely unacceptable. Seth highlights the critical risks associated with early enterprise adoption of LLMs, particularly data leaks of proprietary company secrets and model hallucinations. Engineering designs and technical documentation require absolute accuracy and traceability to authoritative sources. In aerospace workflows, AI must not fabricate answers or cite unverified information, making strict controls and source attribution mandatory before any AI output can be trusted by employees.
To navigate these constraints without throttling innovation, Boeing established a strategy centered on elevating employee AI literacy while establishing stringent guardrails. Seth notes that simply banning AI tools was never a viable option because employees would inevitably seek out public alternatives. The core mission became providing a secure, internal ecosystem where workforce members across all 65 countries could safely leverage frontier models while fully protecting Boeing's 108 years of proprietary domain knowledge.
- Primary risks addressed: Accidental leaks of proprietary data on public LLM platforms, unverified AI hallucinations, and loss of document lineage.
- Strict requirement for authoritative sources: Engineering and process documentation must maintain direct bidirectional linking to underlying specifications.
[00:58](https://www.youtube.com/watch?v=5C5tSNcKaSs&t=58s) The Technical Foundation: Zero Trust, Hallucination Control, and RAG Metadata
To secure proprietary data across all enterprise interactions, Boeing built a microservices architecture operating on strict Zero Trust principles. Security mechanisms are embedded directly inside every individual application layer—from the front-end user interfaces to back-end execution logic and data access repositories. This setup integrates dynamically with Boeing's enterprise Identity and Access Management (IAM) infrastructure to enforce fine-grained access permissions based on employee credentials and geography.
To eliminate hallucinations and prevent data leakage outside the enterprise perimeter, Boeing brings Large Language Models directly into its private Virtual Private Cloud (VPC). The engineering team applies hyperparameter tuning—such as strictly capping the model's temperature—and enforces prompt-level system instructions mandating that the model explicitly state when it lacks the information rather than inventing an answer. For factual accuracy, Boeing enhanced its Retrieval-Augmented Generation (RAG) pipeline to ingest detailed document metadata alongside text content, preserving crucial context such as author identity, document classification, and official authorization status.
- In July 2023, Boeing published and deployed enterprise-wide Generative AI usage guidelines to 170,000 employees.
- The company launched the "GenAI Academy"—a free internal educational platform covering effective prompt engineering, safety considerations, and enterprise use cases.
- In November 2023, Boeing built and launched "Boeing Conversational AI" in 12 weeks, scaling it to all 170,000 employees by January 2024.
- Industry partners such as Microsoft confirmed that Boeing was among the very first global enterprises to deploy a secure GenAI application at this scale.
[10:03](https://www.youtube.com/watch?v=5C5tSNcKaSs&t=603s) Enterprise Impact: Adoption Metrics and 48-Hour App Deployment
After nearly 20 months of active enterprise deployment, Boeing Conversational AI has become a core component of daily operations. The platform counts 38,000 active daily users across 53 countries. Internal user surveys reveal that the tool saves individual team members between 1 to 2 hours per day. Furthermore, Boeing's GenAI Academy has trained over 8,000 employees, cultivating a specialized network of 2,600 "super users" who actively build custom AI workflows on the underlying platform.
Through standardized architecture patterns and reusable microservices templates, Boeing reduced the development cycle for new GenAI applications to under 48 hours. This rapid deployment enables functional groups—such as Finance, HR, and Supply Chain—to replace unwieldy, hundreds-of-pages-long procedure manuals and informal "personal notes" with specialized AI co-pilots. These co-pilots deliver zero-hallucination, compliance-verified answers directly to employees during daily tasks.
- Active daily footprint: 38,000 users spanning 53 countries, saving 1–2 hours per user daily.
- Community growth: 8,000 GenAI Academy graduates and 2,600 active super-user developers.
- Rapid prototyping: New custom enterprise GenAI use cases deploy from concept to functional build in under 48 hours.
[12:06](https://www.youtube.com/watch?v=5C5tSNcKaSs&t=726s) Unlocking Engineering Data, Mechanics' Quality Tags, and Support Automation
Seth outlines several high-impact enterprise use cases, starting with Product Development. Engineering teams previously faced massive friction locating design standards and historical procedures dispersed across 7 million documents stored in 28 separate network share drives. Finding relevant specifications previously took days. Boeing's custom RAG implementation unlocked these repositories with full source traceability, allowing engineers to ask a question, receive a precise summary, and click directly through to the exact document, paragraph, and page source.
In manufacturing and assembly, the platform assists shop-floor mechanics and quality inspectors with Non-Conformance Reports (NCRs). When physical quality issues or non-conformance tags arise during airplane assembly, mechanics must resolve and sign off on the issue before the plane moves down the production line. GenAI applications aggregate and index tens of thousands of historical NCRs, allowing mechanics to search past solutions, draft new NCR records automatically, and identify proven resolution strategies instantly.
```
+-----------------------------------------------------------------+
| Boeing Knowledge Ecosystem |
+-----------------------------------------------------------------+
| |
v v
+---------------------------+ +---------------------------+
| Product Development RAG | | Shop-Floor Mechanics |
| 7M documents / 28 drives | | Non-Conformance Records |
| Instant paragraph citation| | Automated resolution match|
+---------------------------+ +---------------------------+
| |
+-------------------+-------------------+
|
v
+-----------------------------------------------------------------+
| Enterprise Functional Support |
| +----------------------------------+ +---------------------+ |
| | Finance Policy & Closing Co-Pilot| | IT Help Desk Bot | |
| | 1,000+ active users | | Instant resolution | |
+-----------------------------------------------------------------+
```
Beyond engineering, Seth details an unexpected success story with the Finance Co-pilot. Created to help teams navigate complex finance policies—such as determining whether expenditures should be capitalized or expensed during annual planning—the tool was initially launched and left unattended. Six months later, analytics revealed over 1,000 active finance users leveraging the tool, resulting in faster month-end financial closings with significantly fewer human errors. Similarly, an automated IT Help Desk chatbot replaced lengthy phone wait times and high per-call third-party service costs by providing real-time technical troubleshooting based on internal knowledge base articles.
- Knowledge repository scale: Ingested 7 million engineering and design documents across 28 share drives to provide instant paragraph-level source links.
- Model agility: When OpenAI released GPT-5, Boeing integrated the model into Boeing Conversational AI within 6 hours.
- Enterprise App Portfolio: Over 70 custom, domain-specific GenAI applications leveraging internal Boeing data are currently deployed across the business.
[16:27](https://www.youtube.com/watch?v=5C5tSNcKaSs&t=987s) Demonstrating ROI and the Value of Generative AI
Abhi Seth highlights the tangible business outcomes achieved through Boeing's Generative AI implementations. Call volumes and costs have decreased while employee satisfaction has risen, serving as proof that Generative AI delivers concrete value beyond industry hype. Traditional machine learning provided significant benefits in the past, but the speed of GenAI adoption is unprecedented. When applied securely through a responsible, reusable architectural pattern, GenAI yields massive value.
[17:13](https://www.youtube.com/watch?v=5C5tSNcKaSs&t=1033s) Safety Risk Management and Quality Control
When asked whether Generative AI is deployed directly in flight control systems or on aircraft, Seth clarifies that GenAI is not used on airplanes due to potential model errors. Instead, Boeing utilizes GenAI on the shop floor to process human inputs across various safety and quality reporting systems. Employees submit risk reports into multiple disjointed databases, making manual analysis difficult. GenAI ingests these inputs in real time, classifies them according to Boeing's 14 structured risk categories, and assigns risk levels (low, medium, or high). This allows quality teams to identify systemic risks efficiently, supporting Boeing's safety mission.
- **Aircraft Systems:** No Generative AI is used directly on plane control or flight safety systems.
- **Shop Floor Application:** Real-time ingestion and classification of quality/safety risks reported by employees across multiple systems.
- **Risk Taxonomy:** Automated sorting across **14 distinct risk categories** and severity levels (low, medium, high).
---
[18:42](https://www.youtube.com/watch?v=5C5tSNcKaSs&t=1122s) Data Governance vs. AI Expansion: Managing Security & ITAR
An audience member raises the conflict between information governance (which favors data retention limits for compliance and litigation) and GenAI (which relies on large datasets). Seth explains that Boeing's primary challenge is data classification rather than data minimization. Boeing manages complex defense designs, commercial airplane designs, and derivative aircraft that mix both domain types. Overseeing access controls across diverse employee nationalities and international sites requires granular categorization.
```
+-----------------------------------------------------------------+
| Boeing Design Documents |
| (Commercial, Defense, & Mixed Derivative Designs) |
+-----------------------------------------------------------------+
|
v
+-----------------------------------------------------------------+
| Granular AI Data Classification |
| (Identifies ITAR, IP, & Clearance Constraints) |
+-----------------------------------------------------------------+
|
v
+-----------------------------------------------------------------+
| Role-Based Enterprise Access |
| (Delivers appropriate access based on citizenship & role) |
+-----------------------------------------------------------------+
```
- **Granular Classification:** AI tags documents at a fine level to ensure ITAR and proprietary defense data are restricted based on clearance and citizenship without broadly blocking general access.
- **Global Compliance:** Enables Boeing's expanding global workforce to access permitted documents securely.
---
[20:23](https://www.youtube.com/watch?v=5C5tSNcKaSs&t=1223s) Enterprise Adoption Metrics and Internal Security Controls
Discussing user adoption figures, Seth addresses a question regarding statistics showing **170,000 onboarded users** compared to **38,000 active users**. He notes that achieving 38,000 active users among non-technical employees within two years represents substantial adoption. A large portion of Boeing's factory workforce builds airplanes directly and does not use personal computers. A notable metric is that over **2,000 employees use the internal GenAI platform on weekends**.
To prevent data leaks, Boeing disabled external access to public GPTs once its secure internal platform matured. The platform operates under a target response window of **within 6 hours of a new LLM release**, ensuring employees have rapid access to state-of-the-art models within a secure environment permitted to process Boeing proprietary and ITAR-regulated data.
- **User Base:** 170,000 onboarded users, 38,000 active enterprise users, and 2,000 weekend active users.
- **Time Savings:** Employees report drafting customer emails in 10 minutes, down from 2 hours.
- **Shadow IT Mitigation:** Access to external LLMs was blocked in favor of the internal secure portal.
- **Model Deployment Speed:** New LLM releases are brought into the secure environment within 6 hours.
---
[23:08](https://www.youtube.com/watch?v=5C5tSNcKaSs&t=1388s) Computer Vision on the Factory Floor: Legacy Drawing Optimization
Seth shares a practical application of AI in manufacturing. During quality stand-downs, mechanics reported issues reading faint engineering drawings from legacy systems containing tens of millions of documents spanning multiple aircraft generations. Previously, faint drawings were sent to an external vendor (Xerox) to be manually darkened, scanned, and re-uploaded—a process that was slow and expensive.
Boeing developed a computer vision tool that identifies faint regions across thousands of drawings and automatically darkens them. This capability was integrated directly into the decades-old drawing software via a single button for mechanics.
- **Legacy Database Size:** Tens of millions of engineering drawings across multiple generations of aircraft.
- **Usage:** The automated drawing enhancement button is used **25,000 times per day** by factory mechanics.
---
[25:12](https://www.youtube.com/watch?v=5C5tSNcKaSs&t=1512s) RAG Evaluation, Subject Matter Expert Validation, and Air-Gapped Environments
When asked how Boeing validates 48-hour Retrieval-Augmented Generation (RAG) deployments, Seth emphasizes that build speed must be paired with rigorous validation. AI adoption in manufacturing relies on change management and user trust. After assembling a RAG solution in 48 hours, the team spends two to three weeks working alongside subject matter experts (SMEs).
```
+-----------------------------------------------------------------+
| Build Rapid RAG (48 Hours) |
+-----------------------------------------------------------------+
|
v
+-----------------------------------------------------------------+
| SME Q&A Testing (2-3 Weeks) |
| (Engineers generate and validate question-answer sets) |
+-----------------------------------------------------------------+
|
v
+-----------------------------------------------------------------+
| Model Refinement & Fine-Tuning |
| (Iterative adjustment based on SME feedback) |
+-----------------------------------------------------------------+
|
v
+-----------------------------------------------------------------+
| Production Deployment |
+-----------------------------------------------------------------+
```
Regarding air-gapped classified networks, Seth notes that the primary GenAI system is deployed on unclassified networks. Air-gapped classified environments require localized infrastructure and dedicated local LLMs due to isolated network requirements.
- **Validation Workflow:** Iterative question-and-answer testing with SMEs over 2-3 weeks prior to production release.
- **Classified Networks:** Air-gapped environments rely on local LLMs and local infrastructure.
---
[27:36](https://www.youtube.com/watch?v=5C5tSNcKaSs&t=1656s) Upgrading Underlying LLMs in Modular RAG Architectures
An audience member asks how knowledge is transferred when upgrading to newer foundation models. Seth clarifies that Boeing does not train foundational LLMs from scratch. Instead, new models are integrated into the existing secure enterprise interface as they are released.
Because Boeing's applications rely on a modular RAG architecture, vector databases, search pipelines, and security controls reside within the RAG layer rather than the LLM itself. When upgrading from models like GPT-4 or GPT-4.5 to future releases, the underlying vector database remains intact, allowing the newer model to deliver more articulate answers without requiring modifications to the core data infrastructure.
## Notable quotes
> [00:04](https://www.youtube.com/watch?v=5C5tSNcKaSs&t=4s) Every 2 seconds, a Boeing commercial aircraft takes off or lands somewhere across the globe. — Abhi Seth
> [00:58](https://www.youtube.com/watch?v=5C5tSNcKaSs&t=58s) In July 2023, Boeing published and deployed enterprise-wide Generative AI usage guidelines to 170,000 employees.
> [00:58](https://www.youtube.com/watch?v=5C5tSNcKaSs&t=58s) In November 2023, Boeing built and launched "Boeing Conversational AI" in 12 weeks, scaling it to all 170,000 employees by January 2024.
> [10:03](https://www.youtube.com/watch?v=5C5tSNcKaSs&t=603s) Active daily footprint: 38,000 users spanning 53 countries, saving 1–2 hours per user daily.
> [10:03](https://www.youtube.com/watch?v=5C5tSNcKaSs&t=603s) Rapid prototyping: New custom enterprise GenAI use cases deploy from concept to functional build in under 48 hours.
> [12:06](https://www.youtube.com/watch?v=5C5tSNcKaSs&t=726s) When OpenAI released GPT-5, Boeing integrated the model into Boeing Conversational AI within 6 hours.
> [12:06](https://www.youtube.com/watch?v=5C5tSNcKaSs&t=726s) Over 70 custom, domain-specific GenAI applications leveraging internal Boeing data are currently deployed across the business.
> [17:13](https://www.youtube.com/watch?v=5C5tSNcKaSs&t=1033s) No Generative AI is used directly on plane control or flight safety systems. — Abhi Seth
> [17:13](https://www.youtube.com/watch?v=5C5tSNcKaSs&t=1033s) Automated sorting across 14 distinct risk categories and severity levels (low, medium, high).
> [20:23](https://www.youtube.com/watch?v=5C5tSNcKaSs&t=1223s) Access to external LLMs was blocked in favor of the internal secure portal.
> [23:08](https://www.youtube.com/watch?v=5C5tSNcKaSs&t=1388s) The automated drawing enhancement button is used 25,000 times per day by factory mechanics.
## Mentioned
Abhi Seth, Boeing, Boeing Model 1 (B&W), F-15EX Eagle II, F-47, Commercial Airplanes, Defense, Space & Security, Global Services, Boeing Conversational AI, GenAI Academy, Microsoft, OpenAI, GPT-4, GPT-4.5, GPT-5, Xerox, ITAR
## Video description
Abhi Seth, Chief Enterprise AI & Data Officer, Boeing
Boeing Chief Enterprise AI & Data Officer Abhi Seth details how the aerospace company deployed an internal Generative AI ecosystem across its 170,000-employee workforce, prioritizing proprietary data protection, traceability, and modular architectures over public tools or unvetted automation.
Security Architecture and Zero-Trust Guardrails (0:58)
Operating in a strict aerospace regulatory environment, Boeing rejects plausible-sounding model hallucinations and unverified data generation. Because an outright ban would drive staff to external tools, the enterprise brought Large Language Models directly into a private Virtual Private Cloud (VPC) governed by a Zero Trust microservices architecture. Security layers integrate dynamically with Boeing's Identity and Access Management infrastructure to control data queries by employee role, geography, and clearance.
To ensure technical accuracy, Boeing limits model temperature and mandates that models explicitly acknowledge missing data rather than fabricating responses. Its Retrieval-Augmented Generation (RAG) pipelines ingest detailed document metadata—such as authorship, authorization status, and classification tiers—enforcing bidirectional linking directly to underlying specifications before outputs reach users.
Enterprise Adoption and Rapid Application Deployment (10:03)
Boeing Conversational AI was built in 12 weeks and scaled across the company between late 2023 and early 2024. The internal platform records 38,000 active daily users across 53 countries, with an internal survey reporting daily time savings of 1 to 2 hours per user. Outside regular workdays, over 2,000 employees access the system on weekends. To prevent proprietary leaks, Boeing disabled external access to public consumer LLMs once internal tools matured, while committing to integrate newly released foundation models into the secure perimeter within six hours.
Adoption is supported by internal education. Over 8,000 staff have completed the company's GenAI Academy, creating a network of 2,600 super-users who develop custom solutions. Standardized microservice templates have reduced prototyping cycles for custom enterprise applications to under 48 hours, enabling functional groups like Finance and IT to deploy more than 70 domain-specific tools that replace dense procedure manuals with interactive co-pilots.
Engineering, Manufacturing, and Quality Operations (12:06)
In product engineering, RAG workflows index 7 million design and standard documents scattered across 28 share drives, replacing multi-day searches with instant paragraph-level citations. On assembly lines, the system searches tens of thousands of historical Non-Conformance Reports (NCRs), allowing mechanics to reference past resolutions and generate new records quickly when assembly defects arise.
Boeing also applies AI to quality triage and legacy systems. Factory safety and quality reports from disconnected databases are parsed in real time and categorized into 14 distinct risk classifications to surface systemic manufacturing issues. Additionally, an internal computer vision tool darkens faint legacy engineering drawings across an archive of tens of millions of records, a single-button utility integrated into decades-old software that factory mechanics use 25,000 times daily.
ITAR Compliance and Modular Foundation Upgrades (18:42)
Because Boeing handles mixed commercial, defense, and derivative aircraft designs, managing data requires fine-grained AI classification rather than data minimization. Automated document tagging identifies International Traffic in Arms Regulations (ITAR) constraints, proprietary trade secrets, and security clearance requirements, ensuring compliant access for an international workforce without blanket restrictions.
To decouple proprietary enterprise knowledge from underlying foundational models, Boeing maintains its vector databases, search logic, and access guardrails within the modular RAG layer rather than training models from scratch. Foundation model upgrades improve conversational articulation without altering the underlying vector store. While primary deployments run on unclassified corporate networks, classified environments remain air-gapped, requiring local LLMs on isolated infrastructure.
Seth stresses that Generative AI is strictly excluded from flight-control and onboard safety systems due to error risks. While rapid prototyping enables 48-hour application assemblies, Boeing mandates two to three weeks of iterative question-and-answer validation by subject matter experts before releasing tools into production workflows.
[0:04] Abhi Seth: Every 2 seconds, a Boeing commercial aircraft takes off or lands somewhere across the globe.
[0:58] In July 2023, Boeing published and deployed enterprise-wide Generative AI usage guidelines to 170,000 employees.
[0:58] In November 2023, Boeing built and launched "Boeing Conversational AI" in 12 weeks, scaling it to all 170,000 employees by January 2024.
[10:03] Active daily footprint: 38,000 users spanning 53 countries, saving 1–2 hours per user daily.
[10:03] Rapid prototyping: New custom enterprise GenAI use cases deploy from concept to functional build in under 48 hours.
[12:06] When OpenAI released GPT-5, Boeing integrated the model into Boeing Conversational AI within 6 hours.
[12:06] Over 70 custom, domain-specific GenAI applications leveraging internal Boeing data are currently deployed across the business.
[17:13] Abhi Seth: No Generative AI is used directly on plane control or flight safety systems.
[17:13] Automated sorting across 14 distinct risk categories and severity levels (low, medium, high).
[20:23] Access to external LLMs was blocked in favor of the internal secure portal.
[23:08] The automated drawing enhancement button is used 25,000 times per day by factory mechanics.
Original transcript
[0:04] Good, good afternoon everybody. Um,
[0:06] looks like my mic is working so we are
[0:08] set to to talk. I think it's my
[0:09] privilege here. Um, honestly, thank you
[0:12] for taking time and joining this
[0:13] session. It's my privilege to represent
[0:14] the Boeing company, a name that has been
[0:17] defining the future of aviation for the
[0:21] last 108 years. um all the way from
[0:25] spacecraft to you know defense systems
[0:28] of course and commercial airspace that
[0:29] you are all very very familiar with.
[0:31] Today I'll be talking about you know the
[0:33] journey we have taken in the last couple
[0:35] years um in generative AI.
[0:39] Um first let me give you a brief
[0:40] overview of u of Boeing. You know the
[0:43] whole idea about Boeing our mission is
[0:44] really connecting protecting um and
[0:47] exploring the world and doing it in a
[0:49] way so that we can make our planet a
[0:51] sustainable place uh to live. You know
[0:54] we have three key businesses commercial
[0:56] airplanes which is really takes people
[0:58] you know across the planet from one
[1:00] place to another shortens connectivity
[1:02] uh for folks. Uh our first plane uh the
[1:06] Boeing SE plane flew in 1916 and we've
[1:08] been redefining you know the aviation
[1:10] future ever since. We have uh defense
[1:13] and space and security which is really
[1:15] all about protecting our country um and
[1:17] providing deterrence. Um
[1:20] really excited and honored to be
[1:22] building the F-47 uh the most advanced
[1:24] fighter jet ever made uh which is the
[1:27] sixth generation fighter jet. So there
[1:28] is uh a lot of activity happening in
[1:30] that space. Really excited about that.
[1:32] And then the last business I really want
[1:34] to talk about is the global services.
[1:35] And this is all about being service
[1:37] champions, going out in the market,
[1:39] helping our customers solve problems,
[1:41] making sure the planes we are providing
[1:43] uh continue to fly safely um and
[1:46] minimizing the downtime for our
[1:47] customers both on the airline side um as
[1:50] well as the defense side. Boeing is a
[1:53] company that has been investing in AI
[1:55] for several decades. Um some fun facts
[1:58] about Boeing. If you think about it,
[1:59] every two seconds a Boeing jet, you
[2:01] know, um takes off or lands um on the
[2:05] planet and we have basically flown
[2:07] around half the global population um put
[2:10] them on an airplane. In terms of AI, I
[2:13] think uh we have been doing physics
[2:15] based modeling, VR, um simulation, you
[2:18] know, for many decades and investment in
[2:20] AI and ML has been going on for the last
[2:22] couple decades. when generative AI
[2:24] really came came along in 2022 2023 time
[2:28] frame we saw a huge opportunity like
[2:30] everybody else to go tap in the
[2:32] potential right I think we are reading
[2:33] about articles um that hey is genetive
[2:36] AI is living to the hype everybody has
[2:38] already moved on to agentic AI as a
[2:41] conversation this year right how are we
[2:44] doing right and I think I just want to
[2:45] make this more of a conversational
[2:47] um chat um and leave more time for the
[2:50] end hopefully for questions and we have
[2:52] taken this approach coach on how can we
[2:54] you know it's not really just about
[2:55] technology you know we have done a lot
[2:57] of work in responsible AI in the last
[2:59] couple years and how do we really make
[3:00] sure we are intentional about using it
[3:03] where do we use it um and then how do we
[3:06] drive responsibility of uh you know and
[3:09] ethical and bias and transparency and
[3:11] robustness in our models no matter where
[3:13] we are using it we have around 170,000
[3:16] employees in 65 countries so governing
[3:18] AI um and 62,000 of them is are
[3:21] engineers right So um you can imagine
[3:25] it's a lot of work uh to govern AI
[3:28] practices
[3:29] thinking about um how do we approach
[3:32] right I think aerospace is an
[3:34] interesting industry right um when you
[3:38] think about approaching AI and
[3:39] especially generative AI you know a good
[3:41] enough answer in aerospace is really not
[3:44] an acceptable answer right we we are a
[3:46] very highly regulated industry we have
[3:48] to make sure any answer that AI is
[3:50] providing is accurate Right, we need to
[3:53] make sure there is traceability to
[3:55] swords. You know, the first thing we
[3:56] really thought about like as we wanted
[3:58] to go about doing gen AI, it's really
[4:00] not about um AI as a technology and as
[4:03] it keeps on evolving is like what do you
[4:04] do with it to solve the problems you
[4:06] have at hand within the company. Um one
[4:10] of the first things we really wanted to
[4:11] make sure that we have 18 years of uh
[4:13] domain knowledge and a lot of Boeing
[4:15] proprietary information. How do we keep
[4:16] our data safe? Right? We saw a lot of
[4:18] news articles around accidentally
[4:21] proprietary company data getting leaked
[4:23] when people were using genai
[4:24] applications early on. How do we make
[4:27] sure our data is secure? How do we limit
[4:29] hallucinations? Right? I think a good
[4:32] enough answer or an incorrect answer
[4:33] that sounds about right isn't something
[4:36] we can actually do. Um traceability, how
[4:39] do we really we always talk about
[4:40] authoritative source, right? especially
[4:42] when you get into engineering design you
[4:45] know any technical documentation our
[4:47] processes there is a lot of robust
[4:49] control um because of the industry we
[4:51] operate in so if we are going to apply
[4:54] generative AI uh we have to make sure we
[4:56] are able to link to authorized source of
[4:59] information uh for people and then the
[5:02] last one really is really balancing
[5:04] innovation and risk right so don't use
[5:05] it is not a good enough answer because
[5:08] everybody's going to use it so how do
[5:09] you really use it in a safe secure way
[5:12] So you're not throttling innovation
[5:14] while you manage you know the risk of
[5:16] using these capability and for all of
[5:18] that I think you know what it all h you
[5:20] know relies on is elevating our employee
[5:22] AI literacy right so how do you elevate
[5:24] the AI literacy around genai emerging
[5:26] capabilities like genai gentic AI all
[5:29] the you know we we just heard
[5:31] um open AI open source some models there
[5:34] is all kinds of open source models out
[5:35] there how do we make sure people are
[5:37] using it um the right way right so we
[5:39] took this um
[5:42] approach and these challenges in mind
[5:44] and thought about it very deliberately
[5:46] and how do we enable access to all of
[5:48] our 170 170,000 employees across 63 65
[5:52] countries um with genai um we created
[5:58] and I I'll just answer a few questions
[5:59] around I'm sure a lot of you have
[6:01] questions around how did we do that um
[6:03] so securing our data we created a
[6:04] microservices zerorust architecture um
[6:07] so that we can actually don't put
[6:09] security around the components but
[6:11] within each component all the way from
[6:13] the front end, back end and the data
[6:15] access component layer um as as well as
[6:19] really integration of that microservices
[6:21] architecture with our um identity and
[6:24] access management capabilities that we
[6:26] have, right? So that when you know as as
[6:28] any of the global employees are trying
[6:30] to go find something or access or put
[6:33] Boeing data in, we are securing it at
[6:36] all um at all levels within within our
[6:39] application architecture limiting
[6:41] hallucinations. A lot of the work really
[6:43] is done in that case is bringing um LLMs
[6:47] within our own VPC as we secure the data
[6:49] from going out and really fine-tuning
[6:52] LLM so that um we do some hyperparameter
[6:56] tuning you know limiting the temperature
[6:57] of the model so that we can actually ask
[6:59] it to say hey do not answer the question
[7:01] do not make up an answer only provide
[7:03] the answer if based on the information
[7:05] that is actually provided to you. uh
[7:08] traceability. This has kind of been um a
[7:10] little bit of a longer journey and it's
[7:13] really about adding more metadata to the
[7:16] knowledge that we are inserting into our
[7:19] um retrieval augmented uh generation
[7:21] capabilities right our rag architecture
[7:23] we were able to not just ingest the
[7:25] documents but also metadata around the
[7:27] documents which sources are we bringing
[7:28] it in what is the is it an authorized
[7:31] source who is the author etc etc um so
[7:35] those are kind of things we we had to do
[7:37] and very quickly. Um, so we talked about
[7:40] it in July. You know, one of the first
[7:42] things we did was really create Genai
[7:45] guidelines and then launch those
[7:46] generative AI guidelines across the
[7:48] across the company to 170,000 users. As
[7:51] you go use Genai, here are the things
[7:52] you really need to consider. Please do
[7:54] not put any company proprietary
[7:56] information on public Genai websites for
[7:59] example, right? And what to do, what not
[8:01] to do, right? I think we opened a geni
[8:03] academy and which is basically a course
[8:05] where we said hey you know come be a
[8:07] geni advocate um this is a simple course
[8:11] where people can enroll for free um and
[8:14] up their geni um literacy right so how
[8:17] can you do effective prompting how to
[8:19] use it what can it do for you um so a
[8:23] lot of that and then I talked about the
[8:24] zero trust architecture as we created
[8:26] this capability called Boeing
[8:28] conversational AI that we launched right
[8:30] so we launched Boeing conversational AI
[8:31] in November of 2023
[8:33] very quickly in around 12 weeks
[8:37] and we were able to actually you know
[8:39] this was Boeing's first secure genai
[8:41] application and in January of 24 we
[8:43] scaled it to 170,000 employees and we
[8:46] were wondering it's like should we do
[8:47] that right and I think we talked to
[8:49] Microsoft we talked to you know some
[8:50] other consulting companies who actually
[8:52] have exposure across uh across industry
[8:55] and they were mentioning Boeing is one
[8:57] of the first such large scale genai
[8:59] application
[9:01] uh to be scaled back in 2024 January um
[9:06] which was been an interesting thing. We
[9:07] thought about should we slow roll it or
[9:10] or open it to everybody and we decided
[9:11] to just open it to everybody and it
[9:13] scaled really really well. Um we asked
[9:16] people to come take geni training and
[9:19] get access to the conversational AI app.
[9:21] We provided developers API access. So
[9:24] all of that you know if you want to be a
[9:26] you know if you want to use a ji
[9:28] capability as a chat GPT chatbot right
[9:30] you can use that but if you want to
[9:32] build your own applications you know you
[9:33] can use the secure API
[9:36] um we then enabled capabilities over
[9:38] time in the next six months around you
[9:40] can upload your documents so anybody you
[9:42] know you need to know you know how to
[9:43] code you don't know how to code you can
[9:45] upload document you can do Q&A with the
[9:47] document real time and if you build your
[9:49] rag template you can bring it in and we
[9:51] can actually help you productionize that
[9:53] rack template. very quickly.
[9:55] So really reusable patterns on how can
[9:58] we scale Gen AI, right? Um where are we
[10:03] today? And this is kind of my last
[10:04] slide. If you think about it, you know,
[10:06] we have Genai applications that has been
[10:08] used for
[10:11] now 18 or I would say a little bit more
[10:14] than close to 20 months, right? We got
[10:16] 38,000 active users across 53 countries
[10:20] um that use it daily.
[10:22] And we did a survey of our users is like
[10:24] what how are you using it? We are
[10:26] constantly thinking about like how do
[10:28] you use it? What value do you get out of
[10:29] it? We get you know videos out to
[10:32] educate the rest of the rest of the
[10:34] group. Um and what we are finding out is
[10:37] basically they're we the teammates are
[10:40] saving around one to two hours per day
[10:42] um for each member right we've got 8,000
[10:45] people trained in our um genai academy
[10:49] and 2600 super users um who are
[10:52] basically building applications on our
[10:54] genai platform globally
[10:58] with the reusable pattern work that we
[11:00] have done um if now there is a new
[11:02] unique genai use case that we have to
[11:04] actually go build we are able to build
[11:05] that use case in less than 48 hours. So
[11:07] a lot of the focus is really I think
[11:09] speed at which you can roll some of this
[11:10] capability out and scale it out quickly.
[11:13] So for example we have a finance team or
[11:16] we may have an HR team or we may have a
[11:18] supply chain team who may come and say
[11:19] hey we got 800 procedures and it takes a
[11:23] long time for our teammates to go out
[11:25] and read through the those procedures to
[11:27] make sure to do their job they're
[11:29] compliant or what's the right way to do
[11:30] to do the job. A lot of times some of
[11:33] these are you know being being an old
[11:36] company and a highly regulated company
[11:37] right I think people have their own
[11:39] personal notes on how to do the job well
[11:41] right and I think it's it's not scalable
[11:44] um and as we onboard new people you know
[11:47] nobody wants to go read hundreds of
[11:49] pages and and learn how to do the job
[11:51] right and I think so these kind of
[11:52] capabilities as we build these co-pilots
[11:55] we are able to actually zero the
[11:56] hallucination we are going to be able to
[11:58] test these things and then deploy these
[12:00] co-pilots for program management for
[12:03] supply chain. Um a lot of work we have
[12:06] done is really engineering right we
[12:08] talked about 62,000 engineers right uh
[12:11] one of the things I wanted to talk about
[12:12] is um
[12:15] our product development teams we have
[12:18] around 7 million documents in 28 or
[12:22] what's 28 share drives that were
[12:25] existing that people have to go find to
[12:27] figure out how do we you know do a
[12:29] design procedure because there is a lot
[12:31] of um knowledge that has been created
[12:33] over decades that's documented which is
[12:35] Boeing standard engineering practices or
[12:38] design practices and it's not it's kind
[12:41] of unwieldy for people to go find that
[12:43] information. It took them days to go
[12:45] find it. Um and we were able to now
[12:47] unlock that capability with a lot more
[12:50] structure and traceability. So you can
[12:52] ask a question, it'll give you the
[12:54] answer. It'll also give you the source
[12:55] and you click a button, it'll actually
[12:57] take you to the source to that document
[12:59] um and that paragraph. So you can
[13:01] actually go read that, right? you don't
[13:03] have to go find and mine that
[13:04] information in so many different places
[13:06] and some of that has been very very
[13:09] wellreceived. Um we have you know
[13:12] several around 70 plus genai apps that
[13:14] are custom apps using Boeing data that
[13:16] are deployed. So the Boeing
[13:18] conversational AI think about it as a
[13:20] ecosystem right so it's really a general
[13:22] chat GPT now GPT5 was released last week
[13:25] in 6 hours we were able to bring that in
[13:27] uh into Boeing conversational AI so you
[13:29] have access to it we want to bring in
[13:31] anthropic model so it's kind of a secure
[13:33] interface to use all kinds of LLM
[13:36] interaction right and then you are also
[13:38] able to use the API access to do your
[13:40] personal development and um use the
[13:44] build your own rag to actually once you
[13:46] build a build your own rag grid then we
[13:48] have a pattern in which you can actually
[13:49] put it to production using our security
[13:51] and compliance practices and um in one
[13:54] or two days. So we have 70 plus apps
[13:56] that are running and a lot of the work
[13:58] we have seen is really productivity
[14:00] coming out of uh you know several
[14:02] different knowledge worker areas. uh a
[14:04] lot of it is in engineering like for
[14:06] design practices for regulatory
[14:08] information
[14:10] um a lot of information around for our
[14:12] mechanics for example you know I think
[14:14] as you are trying to go assemble a plane
[14:16] um you need to know like hey what is
[14:19] what is the information so we call it
[14:21] non-conformance for example if you find
[14:23] an issue if it's non-conforming you have
[14:24] to create a non-conformance tag um in
[14:27] aerospace
[14:29] and then you some quality person is
[14:30] going to check your non-conformance tag
[14:32] and actually qualify buy it and sign it
[14:34] off and then you need to go address the
[14:36] non-conformance before you can the move
[14:38] the plane forward right you have to fix
[14:40] the quality issue now those issues
[14:42] happen thousands and we have resolved
[14:45] thousands but every time we look at it
[14:47] like it's a new one right so we are able
[14:49] to now use genai to search through
[14:51] thousands of old NCRs is like hey this
[14:54] is the problem here is how we have
[14:55] solved it here is a most likely solution
[14:58] how can we actually automatically use
[15:00] jai to create an NCR and resolve an NCR
[15:03] So those are examples that are kind of
[15:04] helping our mechanics, our engineers,
[15:06] our supply chain professionals. You
[15:08] know, finance co-pilot is an interesting
[15:10] one. Um we didn't think there was much
[15:12] of a need. But people said, "Hey,
[15:15] finance policies are complex, right? We
[15:17] need to figure out what can we
[15:18] capitalize, what can we expense, right?
[15:20] As we planned for the next year, and
[15:22] some of you are from finance, so you can
[15:23] understand. I see some smiles in the
[15:26] room. Uh it's it's not easy." and and we
[15:29] were able to actually create this
[15:31] co-pilot and then you know get some
[15:33] validation and launch it and then we
[15:35] forgot about it. Um just okay it's done
[15:38] and then 6 months later we go back and
[15:40] it has got probably one of the most
[15:41] hits. It's like thousand people using
[15:43] finance copilot really. Uh and then we
[15:46] find out that finance community is super
[15:47] active and they were able to now get the
[15:49] responses back in real time. The month
[15:51] closing is faster. It's with a lot less
[15:53] errors. It's been an interesting journey
[15:55] for for that one. And I think the IT
[15:58] help desk is the last one um that I want
[16:00] to highlight which is really I think you
[16:02] know when there is a IT problem all of
[16:04] us have IT help desk where we phone call
[16:06] and there is a 30-minute wait time and
[16:08] every time somebody calls we have to
[16:10] actually go
[16:12] um you know pay a third party company x
[16:15] amount of dollars to go for each call.
[16:17] How do we you know to get it resolved?
[16:19] Now we have a chatbot that takes care of
[16:21] it um and gives you answers real time.
[16:24] If there is a knowledge article, it'll
[16:25] help you uh why something is not
[16:27] working. And our call volume has got
[16:29] dropped. Our cost has come down and our
[16:31] employee satisfaction has kind of gone
[16:33] up. Um
[16:36] so lots of value from genai. I really do
[16:39] not want people publishing geni does not
[16:41] have value kind of papers or has geni
[16:44] lived up to the hype. I feel like it is.
[16:47] Yes, traditional ML has done a lot of
[16:49] good work for us, but Genai has come at
[16:51] a speed, you know, we all know it's
[16:53] unprecedented and if we can apply it in
[16:55] a secure way um in a reusable pattern,
[17:00] you know, architecture which is secured
[17:01] and responsible, I think a ton of value
[17:04] can come out of JAI. So with that, thank
[17:08] you um for listening and I'll open it up
[17:11] for any questions.
[17:13] Is J AI used in flight safety related
[17:18] for example the plan or the control
[17:21] system of the plan because we all know J
[17:23] AI may have error and that's what we all
[17:27] concerned about. We do not use Nai on
[17:30] the plane but we have used Gen AI in
[17:34] understanding our safety risk right I
[17:36] think on the shop floor and that's kind
[17:38] of where if you think about uh an
[17:40] interesting example that we are kind of
[17:42] working on we have for example quality
[17:45] risk or safety risk that our teammates
[17:47] are identifying they are publishing it
[17:49] into so many different systems right um
[17:52] it's hard for somebody to go and read
[17:54] each one of these things and then come
[17:56] back with an answer to understand okay
[17:58] how do we to resolve it. Now we are
[17:59] using a genai model to ingest human
[18:01] input from so many different systems in
[18:03] real time and Boeing has a very
[18:05] structured process like 14 different
[18:07] categories of risk that we have to go
[18:09] categorize what kinds of risk is it a
[18:11] product safety risk is it you know
[18:13] hazard is it something and then what is
[18:15] the classification of risk like low
[18:17] medium high right so ji is able to now
[18:20] classify all of that and then bring that
[18:23] in and we have validated the model right
[18:25] and I think it's now able to present
[18:27] that to the quality team. So now instead
[18:29] of having somebody go into so many
[18:30] systems and read manually, we are able
[18:33] to understand systemic risk that kind of
[18:35] exists um and it can it can boil that up
[18:39] so we can be a safer company.
[18:42] We help very large organizations uh
[18:46] manage all of their unstructured data to
[18:48] drive geni. The biggest conflict of
[18:51] interest that they've experienced is
[18:54] that one side is information governance
[18:57] which requires support for litigation,
[19:00] compliance, records keeping, privacy
[19:04] and that's in conflict with Genai which
[19:07] likes to keep more data versus
[19:09] governance which likes to keep less
[19:11] data. Uh obviously the risk is viewed
[19:14] differently from both sides. How do you
[19:16] reconcile these conflicts?
[19:19] the the view we have taken is I think
[19:21] applying genai is really I think
[19:24] we don't really we are not thinking
[19:26] about honestly less data we are thinking
[19:28] about our challenge is data
[19:29] classification right I think we have a
[19:32] lot of as I said defense designs and we
[19:35] have a lot of commercial designs and
[19:36] then we have a lot of designs that kind
[19:38] of blend across defense and commercial
[19:39] because they are derivative designs of
[19:41] airplanes right and if we actually mark
[19:45] the procedures and design uh documents
[19:47] at a very high level then you know we
[19:51] can't really unlock that to a lot of the
[19:52] population at Boeing right we talk about
[19:55] only US people can actually access it
[19:57] for example and as we grow globally
[19:58] that's a challenge or we need to really
[20:00] limit that information to certain you
[20:03] know citizen types for example so we are
[20:05] using AI to more do that data
[20:07] classification at a granular level so
[20:09] when we are applying and ingesting it
[20:11] into Gen AI or other AI capabilities we
[20:14] can actually make sure the right
[20:15] information goes to the right person
[20:17] right and it's not unne unnecessarily
[20:18] blocking it.
[20:21] >> Thank you for the question.
[20:23] >> Yeah, I had a quick question regarding
[20:25] this graphic that you have here. Uh I
[20:27] noticed that you listed that you have
[20:28] 170,000 users on boarded. Um but it
[20:31] seems like 38,000 active users. Uh, I
[20:34] was curious like what in your time at
[20:36] Boeing do you think led tound and I
[20:40] guess 30 132,000 users that either like
[20:44] uh well I don't know like what are your
[20:46] thoughts?
[20:48] >> To me I don't think two years ago if
[20:50] somebody said in 2 years we'll have
[20:52] 38,000 active users we'll get here. Uh
[20:54] okay that's a lot of people using it
[20:56] right I think and these are people who
[20:58] are not who just hear about AI they are
[21:00] not really their job is not AI um the
[21:04] other thing I would say is really not
[21:05] every Boeing employee has a computer uh
[21:08] personal computer because a lot of our
[21:09] workforce is really building airplanes
[21:11] and not using jai right so think about
[21:14] it from that perspective so a lot of our
[21:16] knowledge workers are definitely using
[21:17] it we see interestingly the fun
[21:20] statistic that is not listed here is
[21:21] 2,000 people are using it on the weekend
[21:23] Um so that was that was more interesting
[21:26] to me than the 38,000 but you know the
[21:28] population keeps on growing is kind of
[21:30] the thinking right and if you go in the
[21:33] first year of launching and you go give
[21:35] a talk like that within Boeing and say
[21:37] hey we have built this please use it
[21:39] people will be like oh interesting can
[21:41] you give me the link right and and then
[21:43] 6 months nine months later when you go
[21:45] and give a talk like that they're like
[21:46] oh I've been using it and I'm responding
[21:48] to my customer emails it took me two
[21:50] hours now I can do it in 10 minutes or
[21:52] I'm using it for you know writing
[21:55] employee communication or writing job
[21:57] description right I think and all of
[21:58] that can happen in a very secure way
[22:00] right we have built this application
[22:02] with security so you can not just put
[22:05] brewing proprietary information but also
[22:07] put ITAR information right which is um
[22:09] something which is uh which is you know
[22:11] more of a challenge
[22:14] okay
[22:14] >> so uh sorry I have a question following
[22:16] up on your last point of security from a
[22:19] I guess cyber security standpoint is
[22:20] there concerns or pain points um you
[22:22] currently are dealing with or dealt with
[22:25] along the way up to now.
[22:26] >> I think one of our concerns has been
[22:28] even after we released the guidelines is
[22:30] how many people are using external geni
[22:32] capability, right? I think uh and that
[22:34] continue to happen. So once we had this
[22:37] capability mature and scalable, we ended
[22:39] up shutting down access to all the
[22:41] external GPTs and geni capabilities. If
[22:44] you want to use it for work, this is the
[22:46] place. And we want to make sure this is
[22:48] you know that's the reason why the
[22:49] six-hour availability because as soon as
[22:51] you shut everything down people are
[22:52] saying hey this is not good enough I'm
[22:54] using this LLM that is better right so
[22:56] you want to make sure the latest
[22:57] technology is available and it doesn't
[22:59] take 6 months to get after release right
[23:02] now it takes six hours to get after
[23:04] release into into conversational AI
[23:07] >> uh from building the airplane
[23:08] perspective in a manufacturing supply
[23:11] chain um side of it what are some of the
[23:14] challenges maybe you can share um or
[23:16] things that you think that have been
[23:18] done by AI. Uh it could be Gen AI or
[23:21] other AI platforms.
[23:22] >> I'll give you a fun example. Um so we
[23:26] did a lot of quality standowns last year
[23:28] um with our our mechanic population
[23:31] who's building the airplane and we said
[23:32] hey tell us about some of the bigger
[23:34] challenges you have and we have these
[23:36] you know the drawing system for Boeing
[23:38] right. uh there is one drawing system um
[23:41] which has several tens of millions of
[23:43] drawings over so many generations of
[23:45] airplanes. It's an old system. A lot of
[23:48] our drawings as they are trying to
[23:49] assemble the plane, they're not able to
[23:51] read properly because they were kind of
[23:54] faint in certain regions. Okay. Now what
[23:57] were we doing? We would actually
[23:58] somebody say hey I can't read this
[23:59] drawing and I'm putting in plane
[24:01] together. We will go and take the
[24:02] drawing send it to Xerox. They will
[24:04] darken the whole drawing and then we'll
[24:06] bring it in. we'll scan it and we'll
[24:08] upload it. And you can imagine the time
[24:10] and money it costs to do that and you
[24:12] you fixed one drawing and then somebody
[24:14] else comes with another drawing, right?
[24:16] And we got millions of them. So as we
[24:18] heard it and it it is a blocker, right?
[24:21] So we ended up thinking about building a
[24:23] computer vision application outside and
[24:26] then we tested it and it's now able to
[24:28] actually figure out out of thousands of
[24:30] drawings if you input it can go through
[24:32] the whole set figure out which ones and
[24:34] which areas are faint and it can
[24:36] automatically darken it. Great. But it
[24:38] doesn't really help if you have a
[24:40] separate application, right? So we ended
[24:42] up embedding that capability into our
[24:44] drawing application which was decades
[24:46] old. And now when a mechanic is actually
[24:48] looking at it, if they can't read it,
[24:50] they push a button and it'll actually um
[24:53] make the drawing readable, right? Um so
[24:56] that's a huge help, right? It's a small
[24:58] I would say it's a in innovative
[25:00] application, not something you would
[25:02] kind of think of if you don't have the
[25:03] problem, but it's used 25,000 times a
[25:06] day today. So that's helping right as we
[25:09] grow the build rate.
[25:12] >> Hey, thanks for sharing these insights.
[25:14] And I had a question on the 48 hours
[25:16] build and deploy um new geni racks right
[25:20] it's an interesting data point uh
[25:22] building is easy but um evals and
[25:26] validation in my experience has been the
[25:28] biggest challenge can you talk something
[25:30] about how you were able to um expedite
[25:32] that make it rapid and deploy it
[25:34] >> so a lot of on the genai side I think
[25:37] you know you're absolutely right
[25:38] validation is val without validation you
[25:41] can't really go too far because you The
[25:44] whole if you think about what is AI
[25:46] right I think the whole idea about doing
[25:48] AI in an industrial company is not to do
[25:50] AI it's about driving change management
[25:52] solving problems and making sure that we
[25:53] are able to build planes our job is to
[25:55] build planes not to do AI how can AI
[25:58] help build faster safer better aerospace
[26:00] right um so validation is really
[26:03] important a lot of the work we have done
[26:04] in Gen AI is really um you know you you
[26:07] build a rag in 48 hours and then we
[26:09] spend two three weeks withmemes on doing
[26:12] a whole bunch of question an answer,
[26:14] right? So, we will actually produce
[26:15] questions. They will give us questions.
[26:17] We'll produce answers. We'll give it to
[26:18] them. They will validate those answers
[26:21] and questions and say, are they correct
[26:22] or do they need to change and we'll
[26:24] bring them in and fine-tune the model
[26:25] and run that again and then we'll give
[26:27] it to them. So, we have an iterative
[26:29] approach to actually re uh improving the
[26:33] accuracy of the model and once they once
[26:35] they approve it, then we go uh to
[26:38] production because it's really all about
[26:39] change management and trust building. If
[26:41] the if this is giving you the wrong
[26:43] answer or any AI is giving you the wrong
[26:45] answer, people lose trust and they they
[26:46] don't come back and use that, right? So
[26:50] that's kind of how we think about it.
[26:52] Hello. Were you able to deploy this for
[26:56] both the unclassified network as well as
[26:59] possibly airgapped classified network
[27:01] where Boeing employees are? And if not,
[27:04] are do you have any plans to take this
[27:06] deployment into these airgapped areas?
[27:09] we haven't deployed it into air gaped
[27:11] areas. Um so there we are really looking
[27:14] more I would say you know local uh
[27:18] infrastructure and local LLMs right I
[27:20] think the air gap space is a totally
[27:22] different space it's a whole new ball
[27:24] game yes they those employees have
[27:26] access to this but they can't really use
[27:28] the problems that they are solving in
[27:30] air gap on year right I think they have
[27:33] to be very again training comes into
[27:35] play
[27:36] >> thank you great speech today my name is
[27:38] Michael um you mentioned that you are
[27:42] constantly developing new models and
[27:45] trying to implement them. What is the
[27:46] process like of u you already taught
[27:50] your past models? How do you transfer
[27:53] that knowledge that you taught them the
[27:55] first model to the next one? Um how much
[27:58] time, energy, everything? How do you
[28:00] make it efficient?
[28:01] >> So sorry I I would like to just correct
[28:03] what I said earlier. We are not building
[28:05] LLMs. We are just getting as the new LLM
[28:07] gets released we are able to bring it
[28:09] into the interface into the
[28:10] conversational AI secured interface and
[28:13] have make it accessible to the
[28:15] employees. Now to your question as you
[28:17] build 70 of these applications you know
[28:20] we we are basically they will keep
[28:23] running for example on GPD45 as we get
[28:26] GPD5 you know I think the team is going
[28:27] to go and update that uh but it's really
[28:30] a simple rag architecture and a lot of
[28:32] the security is really built into the
[28:33] rack side versus the LLM side. So as the
[28:36] LLM gets better with the same vector
[28:38] database and the lag, you know, rag
[28:40] infrastructure, it's able to actually
[28:42] pick the answers and be more eloquent.
