The Boring Infrastructure of Enterprise AI: What Lockheed Martin, Banco do Brasil and Five Other Giants Actually Built

Key Takeaways

• Seven recent Networks Journal case studies, spanning banking, defence, insurance, automotive engineering, infrastructure and professional services, show a consistent pattern: large organisations are spending most of their AI effort on governance, data consolidation and institutional memory, not on customer-facing novelty.
• Lockheed Martin’s AI Factory only became possible once the company had unified 46 separate data systems into one platform with IBM, consolidation that came before any model was trained.
• Banco do Brasil and KPMG have both concluded, independently and in different sectors, that trust in an AI system’s output is downstream of trust in the underlying data, and have built governance frameworks accordingly.
• Humana’s decade-long experience with conversational AI, begun in 2016, shows that even a genuinely successful deployment needs continuous monitoring and retraining rather than a single launch.
• Amey’s Gennie platform and JLR’s engineering roadmap with Aiimi both point to the same insight: AI’s most valuable near-term use inside a large enterprise may simply be making decades of scattered human expertise findable again.

There is a version of the corporate AI story that gets told at conferences, and it involves a chatbot with a friendly name, a slide showing a hockey-stick curve, and a chief executive using the word “transformation” as though it were a verb one could simply perform. Then there is the version that shows up in the unglamorous middle of an IT budget, where the actual work happens: reconciling forty-six data systems that have never agreed with one another, writing down who is allowed to approve a machine’s decision and on what basis, and building, patiently, the plumbing that has to exist before any of the exciting parts can safely begin. Having read through seven recent case studies published on this site, drawn from Nationwide Building Society, Lockheed Martin, Banco do Brasil, Humana, Jaguar Land Rover, KPMG and Amey, it is the second story that keeps recurring, told in different accents by companies with almost nothing else in common.

The unglamorous prerequisite: data before intelligence

Start with Lockheed Martin, because its case is the starkest illustration of a rule that seems to hold everywhere else too. The aerospace and defence company had, by its own account, valuable data locked away across 46 disparate systems: multiple data lakes, separate analytics tools, disconnected business intelligence platforms, each one presumably built with good reason at the time and each one now a wall between the company and its own information. Stephan Gerali, the company’s Chief Architect and Senior Fellow, put it plainly: “Data is the foundational element by which we pull everything together across the enterprise.” Working with IBM, Lockheed Martin didn’t begin its AI programme by picking a flashy use case. It began by replacing those 46 systems with one connected, accessible environment, built on IBM watsonx.data. Only once that foundation existed did the more visible layer arrive: the Lockheed Martin AI Factory, a secure environment in which 10,000 engineers now build and deploy AI solutions, accelerated by IBM’s Granite open models, with agentic frameworks handling workflows as mundane and essential as HR queries in natural language.

The results, where Lockheed Martin has published them, are notable for their modesty and their honesty. An early generative AI question-answering pilot moved from 45.45 percent response accuracy to 56.8 percent with prompt engineering, and to 66.0 percent after further query optimisation. Those are not the numbers of a magic trick. They are the numbers of an engineering discipline, improving in increments, the way engineering usually does. “We’re getting closer each day to a data-driven culture,” Gerali said, and the emphasis belongs on each day, not on some single unveiling.

Trust is not a feeling, it is a system

If Lockheed Martin’s lesson is about plumbing, Banco do Brasil’s and KPMG’s are about something adjacent and, for a UK readership working under GDPR, the Online Safety Act and an increasingly assertive Financial Conduct Authority, rather more urgent: governance. Banco do Brasil is over two hundred years old and serves, by its own description, a central role in Brazilians’ daily lives; it is also a public institution, which means every AI decision it makes carries a different weight than one made by a private firm answerable chiefly to shareholders. Rafael Rovani, the bank’s Head of Artificial Intelligence, framed the stakes with a clarity that would not sound out of place at a UK regulator’s roundtable: “For us, AI doesn’t replace humans, it unleashes people’s potential by automating tasks. But that’s only possible with the right structure, culture and, most importantly, governance.” The bank brought in EY to design a Generative AI Lifecycle and a Continuous AI Monitoring framework, defining who is responsible for what from model design through to deployment, including benchmarking and vulnerability assessments for the foundation models underneath. IBM then operationalised that strategy through its watsonx.governance toolkit, giving the bank real-time monitoring, proactive alerts and the kind of explainability that a regulator, or a journalist, might reasonably demand an answer to.

KPMG arrived at a strikingly similar conclusion from a completely different direction. As one of the Big Four professional services firms, its problem was not a public mandate but a mountain of client-sensitive information, sitting across a fragmented data estate that made it hard for consultants to access insights without either breaching a confidentiality obligation or simply giving up and doing the work manually. Chris Allen, KPMG UK’s Chief Data Officer, distilled the entire logic of enterprise AI governance into a single, quotable sentence: “To have trust in AI outputs you must first have trust in your underlying data.” KPMG’s answer, a three-year engagement with data specialist Aiimi deploying its Workplace AI platform, is built to classify, refine and govern the firm’s data landscape so that trust can be established before it is required, rather than assumed and later regretted. It is a small thing, on the page, that a professional services giant and a two-century-old state bank in Brazil independently arrived at the same architecture of belief. It suggests the conclusion is not a fashion. It is closer to a physical law of the domain.

The oldest use case still works: making a phone call cheaper and better

Humana offers a useful corrective to the idea that any of this is new. The health insurer began working with IBM Watson in 2016, long before conversational AI carried the cultural weight it does now, because it had a genuinely unglamorous problem: more than a million calls a month from healthcare providers asking routine, well-defined questions about benefits and eligibility, most of which were nonetheless being routed to expensive, outsourced human agents because callers didn’t trust or didn’t bother with the automated system. The eventual solution, the Provider Services Conversational Voice Agent, stitched together several Watson applications, seven language models and two acoustic models tuned to different kinds of query, replacing what might once have been a seven-page fax with a specific, targeted spoken answer. Sara Hines, Humana’s Director of Provider Experience and Connectivity, described her own reaction in a single unguarded sentence: “I’m excited to keep exploring the infinite possibilities of artificial intelligence.” It is worth noting, in fairness to the company, that its own account resists overclaiming: Humana is explicit that the system is not something to deploy and abandon, that it requires ongoing monitoring, training and supervision as provider queries and health plans continue to shift underneath it. Nearly a decade in, the work of keeping it accurate has not stopped. That, more than any single statistic, is the most honest thing a case study can say about enterprise AI: it is a maintained system, not a finished product.

What AI is actually for, most of the time: remembering things

The most quietly persuasive theme across all seven case studies, though, is that a great deal of enterprise AI spend is not being directed at prediction or generation in any exotic sense. It is being directed at memory. Amey’s bid writers sat on decades of asset management expertise that existed only in scattered documents and the recollection of a handful of senior staff, so that a comprehensive bid response took upwards of four days and its quality depended, uncomfortably, on which particular colleague happened to be reachable that week. FOIL’s answer, a generative AI knowledge platform called Gennie, wired directly into Amey’s Microsoft environment, cut that four-day process to thirty minutes, a 95 percent reduction, with 80 percent of tender content now generated through the system and, in Amey’s own account, the entirety of that institutional knowledge retained rather than walking out the door whenever a senior employee eventually does. “We have this vast, deep experience of how we manage assets for our clients,” said Rob Curley, Amey’s Project Director. “What if you could bring together all of that knowledge and make it searchable and instantaneous?” It is a founder’s question asked inside a company old enough to have stopped asking it, which may be exactly why it worked.

Jaguar Land Rover’s engagement with Aiimi runs on the same current, at a different scale. The company’s Head of Data Driven Engineering and AI, Dave Hird, needed to build a business case for an entirely new engineering operating model, which meant first working out which of the countless plausible data and AI projects were actually worth pursuing, measured against impact, feasibility, cost and the company’s own strategic mission. Aiimi’s contribution was as much triage as technology: a prioritised roadmap, and an extension of enterprise search and document classification across JLR’s product engineering documentation, so that some fraction of what 16,000 engineers already collectively knew became something any one of them could actually find. “They helped us show where data and AI could make a real difference to engineering,” Hird said, “aligning that to business value and shaping an AI roadmap that builds momentum quickly.” Nobody in that sentence is describing a chatbot. They are describing a company learning, at considerable expense, how to stop forgetting what it already knows.

The scaffolding nobody photographs

Nationwide Building Society’s case study ties the whole set together, because it is explicitly about scaffolding rather than any single application. Serving more than 16 million members as the UK’s second-largest mortgage provider and the world’s largest building society, Nationwide’s leadership understood, correctly, that the interesting part of a generative AI programme is not the first pilot but the structure that lets a second, third and fortieth pilot happen safely afterwards. Its AI Centre of Expertise, built with IBM Consulting and powered by Microsoft Azure OpenAI Service, exists to run the AI lifecycle, evaluate proposed use cases through a cross-functional council, and build solutions against a set of technical patterns agreed in advance rather than improvised under deadline. “It is where all the important functions with the Society come together, like Legal, Procurement, Technology, Security, Partner ecosystem, to provide one place where we can discuss and approve AI initiatives,” said Nitin Kulkarni, CIO of Data Platform at the centre. Suresh Viswanathan, the society’s Chief Operating Officer, allowed himself the plainest sentiment in any of these seven case studies: “We are very excited to embark on our gen AI journey, like the rest of the world we have been captivated by the technology and the possibilities.” It is a very human admission, from a very large institution, and it sits comfortably beside all the governance machinery precisely because the two are not in tension. The excitement is what justifies building the scaffolding properly in the first place.

What this actually tells a UK reader

Put the seven cases side by side and a picture assembles itself that has almost nothing to do with the AI coverage most executives are fed. None of these organisations bought a single product and declared victory. Each one did the less photogenic work first: consolidating data, as Lockheed Martin did; building governance before scaling deployment, as Banco do Brasil and KPMG both did, independently, in different regulatory universes; accepting that a working system still needs tending, as Humana has for nearly a decade; and recognising that the highest-value near-term use of these tools may simply be making a company’s own accumulated expertise legible to the people who need it, as Amey and Jaguar Land Rover both discovered. For a UK business audience operating under some of the most exacting data protection and financial regulation in the world, that is not a discouraging conclusion. It is closer to good news. The organisations getting real value from AI right now are not the ones with the flashiest demo. They are the ones treating it as infrastructure: unglamorous, load-bearing, and, if you build it properly the first time, blessedly boring to maintain.