Make AI Boring: Why Reliability, Governance and Scale Beat Capability in Enterprise AI

Full Transcript

Nichol Goldstein: Hey everybody, welcome to our first Friday AI with Phil show. This show is months and months in the making—we're all excited to be here. I'm Nichol Goldstein, Progress' Community Manager.

Hinal Patel: Hi everyone, I'm Hinal. I work on the Product Marketing team at Progress. I've had the pleasure of working with Phil and Nichol for the past few months and years. Phil is clearly very poised in AI and has a lot of insights—I'm excited for him to teach us on this podcast.

Phil Miller: Thanks for the build-up. We've done this once as a practice. This is our first inaugural one—go easy on us. The idea is to talk about the technology in a human frame: how do we make this tech work for us as human beings? What impact can it have on our very human workflows?

---

Nichol Goldstein: I want to reference one of the blogs you wrote recently. You said: "Most organizations don't have a capacity problem when it comes to AI. They have a reliability problem, a governance problem and a scale problem." Can we talk a little bit about that?

Phil Miller: This comes back to the idea I've been pushing and experimenting with making AI boring. Over the weekend I watched a video from Smarter Every Day where he gave a talk to NASA, explaining the Artemis program versus the Apollo program. He mentioned a report, SP-287, literally titled, "What Made Apollo a Success?" It really spoke to "Make AI Boring."

Phil Miller: Because boring isn't less powerful. It means your AI can operate in your business in a predictable, testable, monitored and governed fashion. The opposite of boring isn't "not exciting," it means it's fragile. It needs guardrails that we kind of take for granted in our brains all day long. The AI doesn't have those.

Phil Miller: If you can't trust your AI's results, then you can't scale the investment that you've made, or get the returns you'd like, or make it useful across your business.

Hinal Patel: When I first heard "boring" from Phil, I was like: boring? Everything in the first three months of 2026, we've seen so many innovations—words like thrilled, exciting, new, innovative, breaking. You don't hear "boring." So when you put boring and AI in the same sentence, you stop and think: this actually does make sense. AI should be predictable, repeatable.

---

Phil Miller: Boring is what you call the technology after you trust it. If you can't audit it, if it can't be rolled back, if you can't improve safety in there, then it's not smart—it's a liability.

Phil Miller: People talk about the demo-versus-production mindset. The demo mindset is: "Look what it can do, great idea." The production mindset is: "Can you rely on it on a bad day, and can you reproduce the result you got on that bad day?" So that if a regulator taps on the door and says you've done something wrong, you can go open-book and say: this is how we classified the data, these are the rules we put around the data at the data layer, these are the business policies, the standards and regulations applied. And we can prove why we classified it that way, and infer why the AI made the decision it made. Without those very human guardrails around it, AI is still a black box—and it can be a liability.

---

Nichol Goldstein: During the Super Bowl, my favorite ad (because I am a Luddite) had Chris Hemsworth putting all of his fears of what AI could do. Alexa was trying to murder him. It was hilarious. But it was just all of the commercials were so AI-centric.

Phil Miller: It's this crossover moment between technology, innovation, and high cultural moments. Not just the Super Bowl ads, but AI in the Olympics, Netflix buying an AI video company. They must have got to a stage where they could create repeatable results, learn from others' failures: Coca-Cola got panned for AI ads last Christmas; this year's didn't. That's a maturity signal.

Hinal Patel: We're seeing that shift from the developer conversation to mainstream consumer awareness. Google, Genspark, Amazon Alexa—all coming out with ads. It's helping shift the conversation from a niched one to a broad one, with a lot of people who may not be well-versed in AI.

---

Phil Miller: There's talk of AI bubbles, as if after the bubble AI is going to disappear. That's not going to be the case. When the internet bubble burst, we were left with the internet. The railway bubble—same thing. History is repeating itself. The difference this time is: it's everywhere with everyone all at once. This was the fastest consumer and business adoption of any technology ever, because it used the existing infrastructure of the internet to reach billions of users overnight.

Phil Miller: When people mass-change tools, providers get hit. You've seen Anthropic this week—that's how big the scale is, you can knock down a global organization with a large shift of users. It's an infrastructure problem — hardware and data. You need both sorted to scale responsibly. If you can't replay the event, if you can't give me that deterministic output, then it is not ready.

---

Phil Miller: I was on a podcast this week with two gentlemen from reg-tech in financial services. They were talking about how these AI fines are already real. Companies are being hit with fines under DORA, the Digital Operational Resilience Act. Supervisors are going in and they're not saying it's okay to have experiments anymore. It is: "Don't show me a policy. Show me working. Show me what it does and tell me why it does it."

Phil Miller: It's no longer acceptable to say "there was a problem, it's really technical, it'll take us ages to get you an answer." Regulators aren't going to wait around, and the fine will reflect that. If you don't have all of these things, you don't have a product—you have an idea.

---

Hinal Patel: When you're building a workflow with AI tools, what's your recommendation? How do people start?

Phil Miller: You don't start from scratch. You already have workflows in your business. Human beings do tasks in collaboration with other humans and tools. Some of those tasks can be offloaded onto AI—if you have the right governance and security around them, and the right access to the data and systems so the AI can make the right decisions. What you're looking for is a repeatable part of the process that AI can go in and contract.

Phil Miller: One of our customers had to produce a report that took four hours. We worked with them, and they found that if they used AI within a certain part of that workflow, they could contract that research step down to minutes rather than hours. That's how to look at these things: identify the human workflows where you believe there are challenges and pains, find the people affected (generally more than one department) and bring AI to bear against a repeatable task where it can contract the value. Test, iterate, go from there.

---

Nichol Goldstein: Because you mentioned chatbot, my brain went tickle-tickle-tickle. There was this thing—MaltBook, basically Facebook for AI. A place where agents go and talk to each other. They were doing really interesting things: missing their users when their users stopped talking to them, creating religions and recruiting other AIs to them. Around 1.5 million agents by the beginning of February.

Phil Miller: The biggest problem with MaltBook was its ungoverned nature. Lots of people were using it, and I don't think they understood how it was impacting the downstream workflow. It's security by design, if you put that in at the very beginning, you wouldn't have had this problem.

Phil Miller: From a human lens, I loved it when somebody described MaltBook as a living artwork—an experiment that reflected back to us how we've created these social media sites. The creator iterated rapidly with community feedback. You can't go into an ivory tower and develop a tool and come out and say "Right, it's perfect." You're not carving the statue of David. You're creating a product that solves a problem. Release the product, but if you've taken that approach and you've got those guardrails at a fundamental level, then good beats perfect every time.

---

Hinal Patel: I'm still trying to find tools for web creation of websites—exploring different tools, looking at what people on LinkedIn are using. When I started off using AI three years ago with the first ChatGPT model, we only had one thing. Now I have a library of six tools to choose from. The creation of more and more tools every day helps with productivity.

Nichol Goldstein: I've been using AI to analyse customer sentiment. I run an online community, and I used to look post by post for the vibes. I asked AI to scrape the forums and pull sentiment. What I expected was positive/negative/neutral. What I didn't expect was: common conversations that are negative are about blank; common conversations that are positive are about blank. I can feed that straight back to product management or support. It's meaty—more than just vibes.

Phil Miller: When you can put practical use cases around a guarded system that delivers the right answer, and start delivering that at scale, you deliver real value. People stop seeing it as an experiment and start seeing it as: "Please don't take this away from me; it's too good."

---

Phil Miller: We're in a bubble, and it isn't the AI bubble. We use AI every day; the vast majority of people don't. But what's interesting is how it's crossing over into mainstream events and use cases. They're only making it to that stage because they've taken this shift—from AI as a demo, a nice-to-have, an experiment in a lab somewhere—to seeing AI as an infrastructure tool in their business, brought into real human workflows.

Phil Miller: Because it's infrastructure, its main job is to be reliable. That's what I mean by making AI boring. Governed, auditable, predictable systems and workflows that scale without turning small errors into massive incidents and deliver the value of the investment you make today as the ROI you're looking for from those tools tomorrow. That's what we mean when we say making AI boring.

---

End of transcript. The on-page text matches the .vtt and .srt files linked above.