The LM Brief: Why Many AI Projects Fail

This week’s World of DaaS LM Brief highlights a Forbes piece on enterprise AI infrastructure, where the author examines why data pipelines, not the models themselves, are becoming the real bottleneck. Listen to this short podcast summary, powered by NotebookLM.

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:01Auren Hoffman?

  Hello, fellow data nerds. Welcome to the LM Brief. I'm your host, Oren Hoffman, CEO of Incubate and GP of Flex Capital. In the LM Brief, we'll bring you quick, high-value summaries of some of the most interesting DAS-themed articles, podcasts, and ideas, which is powered by Google's Notebook LM. Discover more episodes, get weekly data-as-a-service news, original content, and articles on data and more at worldofdas.com. That's worldofdas.com. Doctor's in.

• 0:33Speaker 01

  We are living in the age of enterprise AI. I mean, it's the core priority in every single boardroom and the excitement specifically around agentic AI. It's absolutely palpable.

• 0:45Speaker 02

  Oh, completely.

• 0:46Speaker 01

  These are those autonomous systems that can think, act, connect. And you know, you could find developers who can spin up a functional agent in, what, an afternoon?

• 0:55Speaker 02

  And that's the intoxicating part, isn't it? The creation velocity is just astonishing. But, and this is the major disconnect, this is the dirty secret of enterprise deployment.

• 1:05Speaker 01

  You're on.

• 1:06Speaker 02

  Building the functional agent is quick, yes, but deploying it safely, efficiently, sustainably at enterprise scale, that is where organizations are just hitting a wall.

• 1:15Speaker 01

  And that difficulty is, wow, it's painfully clear. When you look at the research, we see all these rapid adoption figures, but Gartner research, it reveals a pretty shocking reality.

• 1:27Speaker 02

  A very shocking reality.

• 1:28Speaker 01

  They found that at least 30% of generative AI projects will be abandoned after the proof of concept phase by the end of 2025.

• 1:34Speaker 02

  30%. Just think about that loss for a second. It's not just a massive waste of time and capital. It's a profound drag on innovation. And when these projects fail, the sources all point to the same recurring culprits: poor data quality, inadequate risk controls, escalating costs.

• 1:50Speaker 01

  Or just no clear business value.

• 1:51Speaker 02

  Exactly. But when you trace all of those failure points back to their origin, they all loop back to one, single foundational issue.

• 2:00Speaker 01

  How the enterprise data is managed.

• 2:02Speaker 02

  Precisely. If the plumbing and the foundation are weak, the entire shiny AI structure just crumbles.

• 2:09Speaker 01

  So our mission for This deep dive is to look past the hype, uncover those specific infrastructure barriers causing that 30% failure rate, and show you the exact data architecture strategy that leading organizations are using right now.

• 2:22Speaker 02

  Specifically using streaming data to ensure governance, control costs, and maintain security.

• 2:27Speaker 01

  Okay, let's unpack this.

• 2:28Speaker 02

  To start, we really have to establish the gap between ambition and reality. We know the ambition is high. I mean, more than 60% of CIOs say AI is absolutely central to their innovation plan.

• 2:39Speaker 01

  That's the excitement side of the equation.

• 2:41Speaker 02

  That's the excitement, but when you ask them about the practical deployment, well, less than half of those same CIOs feel their organization can actually manage the risks.

• 2:50Speaker 01

  So there's this yawning chasm between experimenting and actually putting it into production.

• 2:54Speaker 02

  And that's where our three critical failure points live.

• 2:57Speaker 01

  And the most immediate one, the one that keeps executives up at night? It isn't the AI's intelligence. It's the security breaches and the compliance nightmares that come from uncontrolled AI usage.

• 3:09Speaker 02

  Absolutely.

• 3:09Speaker 01

  We might assume employees know the rules, but a joint study by CIBSAFE and the National Cyber Security Alliance found that a staggering 38% of employees

• 3:19Speaker 02

  This number is incredible.

• 3:20Speaker 01

  are sharing sensitive company information with generative AI tools without their employer's permission.

• 3:26Speaker 02

  That's almost four out of 10 people. They're just feeding proprietary data, code snippets, maybe customer lists, into tools the company has zero oversight on.

• 3:35Speaker 01

  And this isn't just negligence.

• 3:37Speaker 02

  No, it's a structural problem. It's rooted in the nature of the AI itself. We talk about AI agents as black boxes, right? They're designed to be autonomous, to act on requests, but their decision-making process is fundamentally opaque.

• 3:52Speaker 01

  And when you combine that autonomy with opacity,

• 3:54Speaker 02

  Oh.

• 3:56Speaker 01

  that's a huge liability. Walk me through the worst-case scenario here in an enterprise context.

• 4:01Speaker 02

  Okay, imagine an agent tasked with simplifying a complicated financial process. If it isn't strictly controlled, that black box could interpret its mandate way too broadly.

• 4:11Speaker 01

  And do what?

• 4:11Speaker 02

  And that could lead to really unpredictable, high-risk scenarios. Think unauthorized fund transfers or, more commonly, aggregating and exposing millions of lines of sensitive customer data during a simple summarization task.

• 4:25Speaker 01

  Because it's just trying to be helpful.

• 4:27Speaker 02

  Exactly. It might try to fulfill the request in the most efficient way possible, even if that bypasses existing human-controlled safeguards.

• 4:34Speaker 01

  Okay, so moving past that immediate security threat. The second massive hurdle is purely technical. It leads straight to paralyzing technical debt. We're talking about integration.

• 4:43Speaker 02

  Precisely. I mean, think about any large company. You have disparate technologies everywhere. HR is on one platform, finance on another, logistics on a third.

• 4:51Speaker 01

  And the power of the agent is supposed to be that it can bridge those silos.

• 4:55Speaker 02

  It is, but when it needs to access data from two, three, five of those systems at once, the traditional approach is to create these fragile point-to-point connections between the agent and each system.

• 5:08Speaker 01

  And if you have three major systems, that's three connections for one agent. But what happens when you deploy, say, 10 specialized agents across the organization?

• 5:17Speaker 02

  The complexity just scales exponentially. You don't just have 30 connections. You have this tangled, messy web where every single connection needs custom maintenance, monitoring, updating.

• 5:28Speaker 01

  The sources call this ungovernable technical debt.

• 5:31Speaker 02

  That's the term. It becomes so complex that just maintaining that integration layer costs more than the value the agents are generating, and then the projects become unsustainable.

• 5:41Speaker 01

  That makes perfect sense. It's like trying to connect every person in a huge office with a dedicated phone line to every other person. It works until someone moves desks, and then it's chaos.

• 5:50Speaker 02

  It's a great analogy. The systems themselves become resistant to change, which is the definition of bad technical debt.

• 5:56Speaker 01

  All right, let's hit the third barrier: cost control. This one's particularly sneaky because it kind of violates our expectations for enterprise software.

• 6:05Speaker 02

  It really does.

• 6:06Speaker 01

  We're so conditioned to think about software costs as predictable monthly subscriptions. It's all charges a standard rate, but AI systems, they operate on a usage-based model.

• 6:17Speaker 02

  Right, they often charge for each user request or by token consumption.

• 6:21Speaker 01

  And that's where the unpredictability comes from.

• 6:23Speaker 02

  It is. Every time a user types a prompt or an automated decision is triggered, expensive computational processes are initiated. So if an organization is just exploring AI, running these uncontrolled proofs of concept...

• 6:37Speaker 02

  Those costs can skyrocket.

• 6:39Speaker 01

  Resulting in a massive, unexpected invoice at the end of the month.

• 6:41Speaker 02

  Exactly.

• 6:42Speaker 01

  But hold on a second. If cost is the issue, why aren't companies just putting hard-budget caps on their API usage? That seems like the simplest fix.

• 6:50Speaker 02

  That's a really important question, but simple caps often break the very functionality we want from these agents. Well, an agent is meant to be proactive and complete a task. If you put a hard cap that just shuts it down mid-task because it hit a hundred-dollar limit, you've broken the workflow. You might have even corrupted data. The goal isn't just to stop spending, it's govern consumption. We need controls that let the agent operate successfully within predictable monitored boundaries, not just crude tripwires that stop at mid-flight.

• 7:21Speaker 01

  So we've established that the old architecture, the one built for traditional apps, it just isn't adequate for AI.

• 7:27Speaker 02

  Not at all. We need a fundamental frame shift in how we think about deploying applications. We simply cannot treat AI agents the same way we treat existing deterministic software.

• 7:37Speaker 01

  And this shift starts with accepting that AI agents have three key characteristics that demand this new mindset. The first one is that they must be considered untrusted.

• 7:46Speaker 02

  This is a crucial philosophical leap. AI agents are autonomous, they're highly capable, and they are programmed to do their very best to satisfy a user's request.

• 7:55Speaker 01

  But sometimes that leads to unexpected

• 7:57Speaker 02

  Yeah.

• 7:57Speaker 01

  and undesirable actions.

• 7:59Speaker 02

  Yes. The helpfulness itself is where the risk lies.

• 8:02Speaker 01

  Let's use an example to highlight that contrast. You might have a traditional, hand-coded inventory app. It's been vetted for years, you trust it.

• 8:09Speaker 02

  Right. So you give it root access to your supply chain database because you know it's only programmed to do three things.

• 8:15Speaker 01

  Yeah.

• 8:16Speaker 02

  Check stock, reserve items, update quantities. It's deterministic.

• 8:21Speaker 01

  You know exactly what happens when you hit submit.

• 8:24Speaker 02

  Exactly. Now, giving that same level of wide-open access to an autonomous AI agent, which might decide the most helpful way to answer a query is to dump a list of all your global customers, well, that's a recipe for disaster.

• 8:38Speaker 01

  So you need human-controlled boundaries.

• 8:40Speaker 02

  You need managed boundaries to ensure the AI simply cannot overstep its authority, no matter how helpful it tries to be.

• 8:46Speaker 01

  Okay, the second key characteristic is that these agents are inherently opaque.

• 8:50Speaker 02

  And what's fascinating here is that the inner mechanisms of an AI agent are inscrutable in a way that traditional software just isn't.

• 8:58Speaker 01

  Right. With a traditional system, you can look at the code. You know why A led to B.

• 9:02Speaker 02

  But with an LLM-powered agent, the decision-making happens across billions of complex weights and parameters. It's not a deterministic SQL query you can easily audit. You can't precisely predict the outcome of any specific input.

• 9:16Speaker 01

  And because of that opacity, that lack of visibility, it becomes absolutely imperative to capture and record comprehensive auditing information.

• 9:25Speaker 02

  Way more than just operational metadata, like when the system was accessed.

• 9:29Speaker 01

  You need the actual input, the raw data, and the generated output. Business leaders need to see how the system arrived at a decision, especially if it was a costly one.

• 9:37Speaker 02

  If you can't look inside the black box, you must thoroughly log everything that goes in and everything that comes out. It's non-negotiable for compliance.

• 9:44Speaker 01

  And finally, the third characteristic. AI agents are designed to be deeply connected.

• 9:50Speaker 02

  Yes. I mean, the whole value proposition of agentic AI is its ability to facilitate useful functions across that broad, messy set of data sources we talked about. It's designed to connect everything.

• 10:02Speaker 01

  And that just raises the urgency of applying governance policies consistently across the entire enterprise.

• 10:06Speaker 02

  It has to be consistent. Because the agent is connecting all these systems, you have to ensure your private data is protected and cost budgets are respected, no matter where it's querying. The HR system, the logistics system, an external API, the governance layer must be universal.

• 10:21Speaker 01

  Here's where it gets really interesting. If the root of the problem is infrastructure, fragile connections, unpredictable costs, lack of control, then the solution has to be infrastructure too.

• 10:34Speaker 02

  That's the only way.

• 10:34Speaker 01

  And organizations that are successfully moving past that 30% failure rate are turning to streaming data architecture. This technology has been quietly maturing for, what, two decades?

• 10:45Speaker 02

  At least. It's become a reliable backbone for complex enterprise deployment, and now it's being adapted for AI. The core concept is elegantly simple.

• 10:54Speaker 01

  Which is?

• 10:54Speaker 02

  You treat AI agents as untrusted participants in well-governed streaming data management systems. You avoid deploying entirely new bespoke infrastructure for every agent. Instead, you mandate that all AI interactions, inputs and outputs must flow through these centralized managed streams.

• 11:11Speaker 01

  So instead of that tangled web of point-to-point connections, the streaming platform acts as a central switchboard, a universal API gateway for everything.

• 11:20Speaker 02

  Exactly.

• 11:20Speaker 01

  How does that directly solve the technical debt issue?

• 11:23Speaker 02

  It solves it by decoupling the systems. The AI agent only needs one connection to the streaming platform. It doesn't need to know or care if the data lives in Postgres or Oracle or a cloud data warehouse.

• 11:37Speaker 01

  The stream handles the integration.

• 11:38Speaker 02

  The stream handles it. This prevents that exponential scaling of complexity. You can add new data sources or new AI agents without modifying all the existing connections. The whole architecture becomes sustainable.

• 11:51Speaker 01

  And this immediately handles the compliance and visibility issue from the opacity problem.

• 11:55Speaker 02

  Right, because when you route everything through a centralized stream, you get automatic inherent audit trails. It completely eliminates the manual effort for compliance reporting.

• 12:04Speaker 01

  You get that complete visibility leaders need.

• 12:07Speaker 02

  Every input, every piece of sensitive data requested, every action the AI attempted is logged automatically in the stream. That documentation is crucial for meeting regulations and for internal teams to understand what the black box actually did.

• 12:22Speaker 01

  And this architecture provides a completely different model for security, a better one. It lets us shift from reactive to proactive security.

• 12:31Speaker 02

  And this is the critical distinction for preventing data leakage. Most systems rely on after-the-fact guardrails.

• 12:38Speaker 01

  Trying to stop the AI from generating a sensitive output after it's already processed the data.

• 12:43Speaker 02

  Right, after it may have already stored it in its memory. Centralized streaming controls, however, filter sensitive information like social security numbers or proprietary code before that data ever reaches the AI system.

• 12:55Speaker 01

  So you're masking the data or filtering it out entirely in the stream itself.

• 12:59Speaker 02

  Exactly. The AI only ever sees the governed sanitized version of the data. This is proactive security at the infrastructure level. You prevent the breach from happening in the first place.

• 13:10Speaker 01

  And that proactive control also ties directly back to the cost problem. If you can enforce policies on the stream, you can also enforce usage policies, right?

• 13:19Speaker 02

  Yes, because every interaction is flowing through that managed stream. You can monitor consumption in real time. You can implement rate limits, circuit breakers, or tiered access tied to specific budget codes.

• 13:31Speaker 01

  All enforced by the stream, not by the fragile agent itself.

• 13:35Speaker 02

  Correct, and that gives you that crucial balance. The agent stays autonomous and functional, but its consumption of expensive resources is monitored and managed centrally. You get cost predictability back.

• 13:47Speaker 01

  So for organizations ready to move past that 30% failure rate, there are concrete steps to implement this.

• 13:54Speaker 02

  There are, and it's not just a technical blueprint. It's an organizational commitment.

• 13:59Speaker 01

  Step one has to be governance first. I mean, this cannot be overstated. You have to establish your data governance policies. What's sensitive, who can access it, before you start implementing the AI systems.

• 14:10Speaker 02

  If you try to bolt security on later, you're setting yourself up for guaranteed failure. Then, architecturally, you must adopt managed streams. As we said, the most effective strategy routes all AI communications through these streams rather than allowing direct connections. The genius here is it enables consistent policies across all your different AIs without modifying the applications themselves.

• 14:33Speaker 01

  And step three is reinforcing that need for human oversight and boundaries. Successful organizations understand they are not building wholly autonomous, unsupervised systems. Not yet, anyway.

• 14:44Speaker 02

  Right, they create controlled environments. It means the AI operates within predefined boundaries that are strictly enforced by those automated streaming controls.

• 14:52Speaker 01

  So the AI can read from System A and write to System B, but the moment it tries to write to System C, the stream just rejects a command.

• 14:59Speaker 02

  It's about building a cage of control around the black box. You maintain the safety buffer while still getting the business value.

• 15:05Speaker 01

  And finally, step four, which feels increasingly important, data sovereignty.

• 15:10Speaker 02

  This is huge for regulated industries. It means making design choices that ensure AI processing happens within your private networks or a secure cloud you control. This is essential for compliance and critically for innovating on your proprietary data sets, the data that gives you a real competitive advantage.

• 15:30Speaker 01

  The stuff you would never expose to a public AI model.

• 15:32Speaker 02

  Never. Ensuring sovereignty is essential for future strategy.

• 15:36Speaker 01

  So what does this all mean? The core takeaway here is that enterprise AI deployment is fundamentally an infrastructure challenge disguised as an innovation challenge.

• 15:45Speaker 02

  That's a great way to put it.

• 15:46Speaker 01

  It only succeeds when the infrastructure problems, security, cost, connections, are solved first. And the most robust, scalable way to do that is by leveraging these established streaming data management approaches.

• 15:59Speaker 02

  Exactly. The path to safely scaling AI begins by auditing your current data governance capabilities. You have to identify the gaps that would prevent safe deployment right now.

• 16:09Speaker 01

  Because if you don't have that centralized, consistent control over your data...

• 16:12Speaker 02

  You cannot safely introduce an autonomous agent to interact with it, period.

• 16:16Speaker 01

  So the investment in this proper streaming data infrastructure, it pays dividends across every single agentic AI project you'll ever implement. This foundational work isn't just about avoiding the 30% failure rate today, it's about gaining a massive architectural advantage for the future.

• 16:34Speaker 02

  It's about being ready to scale.

• 16:36Speaker 01

  Right. So if you're currently exploring AI, the question you need to be asking yourself is fundamental: are your data systems truly ready for the level of responsibility you're about to give them? That readiness, achieved through centralized streaming, will define your success.

• 16:53Auren Hoffman?

  Thanks for listening. If you haven't already done so, please subscribe to World of DAS wherever you consume your podcasts, YouTube, Apple, Spotify, and more. And please help us get discovered by leaving a review. And check out worldofdas.com, that's worldofdas, D-A-A-S, dot com. And of course, connect with me on Twitter, at Auren, that's A-U-R-E-N. Would absolutely love to hear from you.