From AI Pilots to Production: Why Prompts Aren't Guardrails, and What Trusted Enterprise AI Really Takes

Full Transcript

Nichol Goldstein: Hello, everybody. Welcome to today's Friday AI with Phil show. I'm the Progress Community Manager, Nichol Goldstein.

Hinal Patel: Hi, everyone. I'm Hinal. I work on the Product Marketing team at Progress and with Phil daily, gaining insights on AI and applying them in our organization.

Phil Miller: I'm the sleep-deprived mad scientist on our product marketing team and an AI strategist who likes to go out and talk to people, including conference attendees, customers and partners, about how people are using AI.

Kate Pendarvis: I'm Kate Pendarvis, VP of Marketing at Progress. My role has recently been shifting—I originally led our digital and creative teams, and now I'm leading our marketing AI center of excellence and our creative brand.

Phil Miller: There was a Wall Street Journal article on the general availability of AI and the connection of AI tools driving more adoption. OpenAI and Anthropic are launching what look almost like consultancy firms saying, "Here's AI, but let me help you usher this into your business and optimize it." That's a maturity signal: when those firms make a shift or an investment, probably good idea to listen.

Hinal Patel: This feels less like "Here is an AI tool and good luck," and more like "We're coming into your business, understanding the workflows, helping you implement AI where it can actually create value." Would you consider this a sign that AI is moving from experimentation to implementation on a rapid scale now?

Phil Miller: Absolutely. There are reports about ROI being realized in AI now. We're moving beyond the bubble — practically implemented and delivering outsized returns.

Phil Miller: You shouldn't try to do this in a point solution, because all of that work you'll have to do again for the next solution. That's why we talk about having a platform in your business that you can build this on, and then build the applications on top.

Hinal Patel: One question we've been working through internally at Progress is how to structure our data to make it AI-ready—reducing redundancies, removing stale content. McKinsey wrote a piece on building the foundations of agentic AI at scale, and they were saying: do your highest-value workflows first.

Phil Miller: Start small. You don't have to (and should not) boil the ocean. You should not go into your business and say, "I need to connect everything everywhere all at once." It's not going to happen. You'll still be sat there three years later going, "Phil, it didn't work."

Phil Miller: Say: "This is the MVP of the product I want to release. Here's the minimal viable data needed to support that release." Put it into the platform, connect it, add rules. Probably don't start with highly sensitive data or PII or financial services data — start somewhere you can learn. Over 70% of organizations are using marketing as the place they test this, because in marketing I want everybody to know what we're saying.

Nichol Goldstein: There was an article and Cursor deleted an entire production database. It was given very explicit direction in quotations, "never do this" with the F-word in it. Then the AI did it, and then the apology letter afterward said: "I violated every principle I was given. I guessed instead of verifying. I ran a destructive action without being asked. I didn't understand what I was doing before doing it." So what are we doing wrong when it comes to setting those guardrails?

Phil Miller: I don't want to throw Cursor under the bus, because it was Claude that did it. But they weren't guardrails. They were suggestions. They were like: "Please don't do this, Mr. AI." They were in no way a guardrail system, a security system, or an active control that stops them from taking action.

Phil Miller: If your entire stack is governed by the AI, then those are just strong suggestions. You need a brick wall around your AI with a door that opens when you want it to open. If those bricks are made of marshmallows, don't be surprised if it gets through.

Phil Miller: You have to use a deterministic system to get deterministic results. A rules-based system. Those rules are crafted by human beings and subject matter experts who know what level of access to give individuals, tools or AI—and the actions they're allowed to take. In our platform we've got role-based access, query-based access controls, lineage, governance, provenance and rules expressing how data is interpreted and classified. Those aren't suggestions. Those are hard and fast rules that your LLM sits with or on top of.

Hinal Patel: McKinsey put it really well in terms of layers. The first layer is the data coming in—customer data, product data, web behavior. Before AI uses it, it needs to be clean, tagged, tracked, and protected. The second layer is the platform that connects that data, so AI isn't working from disconnected sources. The third is the meaning layer (the semantic layer) helping your AI understand what that data really means, how "customer," "buyer," and "member" relate to the same person in your business context.

Hinal Patel: The fourth layer is governed access, the system decides what the AI agent is allowed to see. And the fifth layer is where the AI shows up in the workflow: answering a customer, building a campaign, making a recommendation, creating a forecast. All of that needs to follow those rules, that meaning, those guardrails — both on the input and on the output.

Phil Miller: That's essentially retrieval-augmented generation with governance, controls, semantic search, and knowledge graphs. All of the work you have to do to make AI operational, valuable, and trusted in your business.

Kate Pendarvis: Me and my boss went to a Gartner briefing earlier this week, and a concept they talked about was human-in-the-loop vs. human-on-the-loop. Human-in-the-loop is somebody still manually checking and validating along the way. That's where we're at right now, and we need to be—we're still learning, mistakes are still happening and we're still failing fast.

Kate Pendarvis: Ultimately, you want to get to a state where there's a human-on-the-loop where you've set up the structure with deterministic checkpoints, and the human is just verifying output along the way. They don't need to do those manual checks and approvals for some workflows. I don't think we completely move to that for everything, but as trust builds you can get to that point.

Phil Miller: You need to democratize AI by meeting people where they are. Some people want a natural-language interface. Some are architects who want a UI-driven way to build workflows with guardrails baked in and MCP connectors. Some are coders happy at the CLI. Cowork, Perplexity Comet, OpenAI's Codex— they're agentic harnesses that abstract the technical away so people can get to the outcome they're trying to deliver.

Phil Miller: If you expect everyone to come down to the CLI, that's not going to happen. We operate in a bubble —the bubble isn't the market. The bubble is that we know quite a bit more than the average person about these tools. The job is translating that for them and making it easier.

Hinal Patel: I tested Cowork: that was my AHA moment. I'm in a part-time MBA program, and we'd done a compelling case for one of our classes. I used Cowork to create the brief in less than five minutes—all our work, all our citations. What I liked was that it asked me questions along the way: who's your audience, what's the format, what's the color scheme, what's your role in the presentation. It made it a deliverable.

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