Are today’s AI tools actually doing the work, or are they still sitting on the sidelines offering advice that humans have to act on? In this episode of the AI at Work podcast, I sat down with Oren Michels, Founder and CEO of Barndoor AI, to explore why so much enterprise AI still feels stuck in what he calls “advisor mode.” We talked about the gap between AI that summarizes and AI that acts, and why that distinction matters far more to knowledge workers than most leaders realize. Oren drew on his experience building Mashery during the early days of APIs, drawing a clear parallel between then and now, when powerful technology exists but remains inaccessible to the people who actually need to use it. We spent a lot of time unpacking what true agentic AI really means inside the enterprise. For Oren, it is not about smarter chatbots or recycled RPA workflows, but about agents that can safely take action inside systems like Salesforce, CRMs, and other tools of record. We discussed why so many AI initiatives fail to deliver ROI, and why the missing skill is often not prompt engineering, but the ability to break real business problems into clear, executable instructions that an AI agent can actually follow. Governance became a central theme in our conversation, especially as we dug into the Model Context Protocol, or MCP. While MCP is emerging as a powerful standard for connecting AI to enterprise tools, Oren explained why it also introduces new security, cost, and control challenges if left unchecked. We explored why governance should act as a launchpad rather than a brake, how least-privilege access changes the conversation, and why the most important question is not how a model was trained, but what it can do with access right now. If you are thinking seriously about agentic AI, enterprise adoption, or how to prevent “bring your own AI” from becoming the next wave of shadow IT, this episode will give you a grounded, experience-led perspective on what actually needs to change inside organizations. As AI agents begin to operate at speed and scale across core systems, are your guardrails designed to stop progress, or to make it possible to move forward with confidence? I would love to hear your thoughts after listening. How close do you think we really are to AI that acts, not just advises? Useful Links Connect with Oren Michels Learn more about Barndoor AI Thanks to our sponsors, Alcor, for supporting the show.
Audio and video play straight from the show's own hosting. PodSubs stores and serves no media.
Transcript
From the show. Tap a time to play from there.
Welcome to AI at Work, a podcast which is part of the Tech Talks Network. And in this podcast, we're going to venture into the transformative influence of artificial intelligence inside the workplace. And our discussions will focus on both the remarkable breakthroughs, but also the complex challenges of integrating AI into our everyday business functions and workflows. Have you noticed how every wave of technology seems
to arrive with a lot of excitement but equally a lot of confusion and often a nagging sense that have we seen this movie before? Well today's conversation starts in New York and this is an audio podcast so you won't be able to see this but my guest backdrop is framed by Broadway posters and we'll quickly move into familiar territory for anyone who has lived through APIs, mobile, cloud and now AI. Because my guest today is Oren
Michaels, he's a founder and CEO of Bondor AI and they are a builder who helped shape the API movement long before it was mainstream and despite stepping away to produce Broadway shows He's now pulled back in by what he sees as unfinished business in the world of enterprise AI. So today we can talk about why AI is still stuck in advisor mode, what agentic AI really means when software
finally starts taking action and why governance, security and trust are just a few of the things that need to move at the same speed as innovation. So if you're wondering why AI still feels powerful but underwhelming inside real workflows, this one's for you. But before I get my guest on today, I want to give a quick thank you to my friends at Dinodo, because one of the questions I hear more and more from listeners on this podcast
is, why does AI succeed or why does it fail? Now, most projects fail not because of the AI, but because the data foundation isn't ready. This is why organizations are increasingly turning to Denodo. So if you're ready to understand why your AI projects fail and how to succeed with AI, simply visit denodo .com and take control of your data world. But enough for me. Let me introduce you to my guest right now. So a massive
warm welcome to the show. Can you tell everyone listening a little about who you are and what you do? Sure. Great to be here. Thanks for having me on, Neil. I'm Oren Michaels. I'm the co -founder and CEO of a company called Barndoor AI based in New York City. And I have been a technology entrepreneur for a long time. Back in 2006, I noticed there was an opportunity to take these very technical sounding things called APIs and
turn them into a way for businesses to grow and interact with each other and build larger businesses. But unfortunately, there was no technology, there was no ecosystem of technology providers to give you the tools you needed to expose these APIs. So I started a company called Mastering. And we brought API management into a place where business people and really anybody in the company could use APIs and turn it into something they
would use to grow their business. Ran that company for seven years and sold it to Intel and was just enjoying being an angel investor and sitting on boards and producing Broadway shows. And a couple years ago, I started noticing that there was this move toward bringing AI into the enterprise, but we were having similar problems that really reminded me of the problems we had back in the API days. We had this great thing, chat GPT, it was a chat interface, but it was really mostly
useful for people who were, let's say, writing code. You could take all your bad code that didn't work, throw it into chat, and it would fix it for you. which is great. You talk to this brilliant PhD and it tells you what to do, which is lovely if you have that kind of job. But most people don't have the kind of job where they chat with a PhD. Most people actually use tools and do things, interact with Salesforce, they create documents, they write emails, they do all kinds of things in their job. And the AIs out there really weren't able to do those tasks. And so
we started looking at what it would take to allow AIs to safely and securely be trusted. to go do these things. And what came out of that was this company, Barn Door AI, that we started a bit over a year ago. What a great story. And as you said there, you helped shape the API movement with Masher, enjoying the shift to mobile and cloud. Then that period of producing Broadway shows there. But from that moment, it almost feels like there was a Godfather 3 moment, you
know? Just when I think I'm out, they pull me back in. So I mean, I had those experiences influence the way you see the current turn towards agentic AI and workflow automation for, hey, one last job. Well, you know, it's funny. Part of having done this a few times is that you, you know, history does repeat itself when you see some patterns. So hopefully there are things that I, you know, screwed up last time around that I'll get right this time. I'm sure I won't get them all right, but it also, you know, it really
helps you understand what kind of team you have to put together, what kind of, you know, what sort of service and support and involvement, you know, sophisticated enterprise customers need to deploy something like this. Because we had to build that last time, you know, we had in the API days, when I started the company, I thought it was going to be for, I had been at a small company myself, I thought it was going to be something that little companies would use to open APIs and big companies could come and work with them. And then the iPhone happened
about three months after we started the company and suddenly everybody needed a mobile strategy. And pretty much overnight we went from serving little startups to serving Walmart and Best Buy and the New York Times and all these amazing companies around the world. And i learned pretty quickly what was involved in doing that and so hopefully this time uh we get to you know we get to have the benefit of that experience and one of the things that you said is that ai is still stuck in advisor mode that was something
i found when i was doing a little research on you i completely agree but when did you first realize that that real change would come only when ai began taking real action rather than just summarizing information. And we're starting to see this with the future of agentic AI that everyone's talking about now. But when did you, you first notice this? Well, people started talking about agentic AI. And when I started digging into what they meant by agentic AI, they either meant things like customer service chat bot agents,
which certainly are things and they're lovely and they can be very helpful, but that's not what we really had in mind. And then they would start talking about, well, you can have an agent that will go out and, you know, pull in a bunch of data and summarize it for you. And that's not really what an agent does. If I have a, let's say I have a customer success agent, they actually take actions. They do things by the, uh, uh, agent who is going out and talking to customers, they will log that their, their calls in Salesforce, they will actually take actions. And so for me,
I sort of summarized it as. If an AI isn't actually writing something or impacting something, all it's doing is gathering information and turning it over to a human with some suggestions of what to do. In my book, that's not really being agentic. That's just being a chatbot. You can be very sophisticated at that. And we had a trend six, seven years ago of robotic process automation or RPA. And a lot of these so -called agentic workflows really are just Rewarmed rpa and not
nearly as interesting as as they could be so what i look at it and i say okay i'm human and i use a i to hopefully make myself a bit more productive and i automate certain things i do and that's great that's interesting and it's hopefully useful and maybe is productive for my company but what's really interesting is when we get to a place where because of what. AI can do and what these agents can do and the tools they have access to, they can actually do things
that humans can't do. It's not just a faster Orin, but it's doing things that Orin can think of wanting to be able to do, but could not possibly get done in any rational period of time without something that can actually take action and do things. And as those use cases come into the enterprise, we're going to see some really amazing growth and some amazing productivity enhancements and hopefully businesses will grow because of it. And throughout 2025 we did hear a lot of
talk around AI productivity and yet on the flip side of that many leaders were still scratching their heads wondering why they weren't seeing ROI on their investment, why they weren't coming out of test into production, why they wasn't seeing outcomes that match the hype. So what patterns did you see behind that frustration? Was it that's a certain amount of bandwagon jumping or not focusing on the problem first and going tech first or was it something else? else. Oh, it's, it's, it is something of that ilk. It's, okay, we have to have AI. Okay. Well, which one
we'll use chat, cheeky, cheeky. We'll use anthropic. Okay, great. We'll get that and we'll put it out there. And nobody knew what to do with it. That basically, if you, you know, I still talk to you, I was in the cab yesterday and the cab is like, well, what should I use AI for? And we started talking about it and he finally came up with some really great ideas of what, you know, a New York city cab driver could do with it. But it's, um, you know, just by putting a tool out there, doesn't mean you're going to start really understanding what to do. You need to have examples. And you also need to be someone
who is good at explaining to someone, whether it's a human someone or an artificial someone, what to do, how to do it, and what the work product is that you want. And, you know, if you look at a lot of people who work in, you know, in knowledge worker jobs, They may be great individual contributors, but they may not have the skill of saying, okay, I've got this task. I've got this job. I'm going to break it down into things and be able to tell someone who's super smart
and super energetic, but doesn't really know and have context exactly what I want. And that skill and being able to, you know, we have been trained to call that prompt engineering, but it's more than that. It's more than just engineering the prompt. It's actually being able to understand and be one of those humans who's able to sort of dissect a business problem and turn it into a set of instructions that someone can go do. And the fact that you give people this powerful AI doesn't mean they're going to get stuff done
with it any more than if you give, you know, an individual contributor, you say, okay, look over there, you got six interns, go put them to work and make you more productive. They may or may not know how to do that. And a lot of people have likened it to the iPhone moment, but I would argue it feels more like the App Store moment. Do you remember when the App Store launched, but every iPhone app was turn your, I don't know, your phone into a pint of beer or a flute or a chainsaw. It wasn't until developers had a bit of time with it and thought, okay, we've had a bit of fun now. What real problems
can we solve? It feels very similar to that, doesn't it? Very similar. And it's, you know, I'll even take it back a step further. So, so you're absolutely right. And, you know, at Mashery, we were fortunate the iPhone came out and that sort of turned us into a catalyst for a lot of mobile apps. So we got to see a lot of, you know, good ones. And it turns out a mobile app that actually does anything generally needs an API to talk to, right? That's sort of what made our business. But, you know, right now we're sort of in this phase where you have this chat, these
chat clients that are mostly text. And yeah, they can put some pictures up and you can do some things. We just learned that chat GPT and, you know, open AI released an SDK to put little I -frame apps in their chat interface. And they're kind of kludgy. They're kind of wonky. It's a, it's a step in that direction. To me, it reminds me of when flash came out. So you had the websites and suddenly they became quote rich media and they had. interaction and things like that but
it was all really tough because you had to be running this flash thing that was kind of weird and and you know eventually people figured out how to actually extend the power of the capability of both browsers and servers and the languages and all those things to turn them into the rich sites that we have now and i think we're we're sort of in that moment people don't know what to do and they want to make sure that these apps are are deployed with the proper guardrails with safety so they don't go running amok and doing
things they shouldn't do. And that is a step in the direction that I think will evolve fairly quickly and we'll see a significant evolution over the course of the next 12 months. I was also reading that you describe a coming moment when AI agents could actually outnumber human employees. The prospect of that makes me a little bit nervous from a security point of view, but
how should leaders be rethinking workforce management when digital entities begin to operate across so many different core systems with speed and scale that humans cannot match and we all have our own agent almost working on our behalf? Well agents plural and I think that you know part of it is there's not going to be one single AI to rule them all in any organization, right? It's just that we hire different kinds of humans to do different jobs. There are going to be different AIs, different agentic platforms that are particularly good at doing things, but I'll take it a step further, which is that it's all also about trust
that if you want to have an agent that is going to go do something particularly that involves writing into your systems of record or talking to a customer or you know these sorts of things. you're not going to have a single all -seeing, all -acting agent. You're going to create something that has significantly narrow guardrails that it can't get out of, that it will go do the task at hand, and it may hand that task off to a different agent who has its own guardrails. But ultimately, you want to... These agents are non -deterministic.
They do their best to act within how they interpret the instructions they're given. And so... by narrowing these things down and making it so that I might have the agent that logs my call after I talk to you and takes my notes. So if we talk again, I remember what it is. That might be a different agent than the one that logs a call with a prospect. That would be a different agent than the one that goes through all of my salespeople's pipelines and perhaps ask them
questions about the status of various deals. Each of these things will act differently. I don't think that in the near term we'll just have here's Oren's agent. It just sort of sits next to him and does all the things he wants it to do and, and never screws up and never embarrasses him. And, you know, that, that just, I think we have a bit of time to get there. A quick thank you to the sponsor that supports every podcast across the tech talks network and every episode, because their help allows me to publish 60 interviews
a month with founders and technologies who are keeping this industry moving. And this month I'm partnering with Alcor and if you've ever tried to hire engineers in another country, you probably know just how painful it can be. Different laws, patchy support and partners who don't truly understand engineering roles. So Alcor approaches this from a different tech point of view. They specialize in Eastern Europe and Latin America and they're able to combine EOR capabilities with recruiting. So you get one partner handling
everything and they help you choose the best location for your stack, find developers with the right depth of experience and run proper assessments so they can onboard people quickly. And they also give you a model that respects both transparency and margin. Most of your spend goes directly to your engineers and the fee will decrease as the team expands. And you can even transition everyone in -house at that time when you're ready. without having to worry about a penalty. And that structure is why a mix of early
stage and unicorn stage companies use them as they scale. So if you want to take a look visit alcor .com slash podcast or tap on the link in the show notes. But now on with today's show. and bring your own device. We both remember clearly, I think, changing the entire era of corporate security and governance. Then came shadow IT. I'm curious, looking back at these things, what parallels do you see with bring your own AI? Bring your own AI is a thing and people are doing
it all the time. And the parallels, it comes down to, let's leave aside intentional bad actors, because that's not what this is. This is people who want to get their job done, they want their company to compete with their competitors more effectively. They want to do a good job. They see this technology out there and they say, gosh, if I had access to this, I could be a whole lot more effective. And so, you know, We've all seen it a bunch of times. Let's go back to mobile email, right? The BlackBerry came out and the
security people are, we will not have mobile email. That's not secure. And the sales guys are like, I'm sorry, I need to hit my number and my competitors have this. So they would just go and put down a credit card and bring BlackBerry and sort of dare IT to say, no, you can't use it. And then, you know, that's how wireless internet came in. We always wanted to be able to sit down at a conference table and be on our laptop and look something up. And so we would put a Wi -Fi access point in the conference room, even though IT was like, oh my God, that's not secure. And, you know, and on and on and on. So now we have
AI. And a lot of the initial, those, you know, 50 or whatever percent of bad AI proofs of concept, They would basically give you AI, but nerf it to a point where if you wanted to get anything done, you had to go outside. And the challenge with narrowing this down, and we see it in agentic AI as well, we see, all right, we'll give you MCP, which is the protocol I'm sure we'll talk about. We'll give you MCP, but it's read -only.
Well, they can't do anything. If you insist that nothing bad can happen, you're also insisting that nothing good can happen. And so people will go right around that. They'll set up their own MCP. They'll just go directly in with the API. There's all kinds of things you can do when you bring your own AI. And the bottom line is if IT wants to actually be a catalyst and a facilitator of growth and innovation and good things happening, Smart risks need to be taken and you have to
give people the ability to use these tools or they're going to go figure it out themselves. And we are at that time of year where everyone's starting to think about what's going to dominate 2026. And you used an acronym there. Technology loves a good acronym. MCP. We're not talking about Microsoft certified professionals. No, we are not. Though I did hear that the other day is a thing I'm told. Yeah, it used to be, but obviously the model context protocol is what I'm talking about. The connective layer for AI
workflows across systems in everything like Salesforce and Snowflake. So what excites you about this shift? What do enterprises need to understand as they open these connections? Because I do expect throughout the year, we're going to hear MCP more and more, aren't we? Absolutely. Absolutely. And so MCP is is a amazing thing that conveniently for us, again, the last company I did, we started doing APIs and then the iPhone came out this time. About a month after we incorporated and cached our first, you know, series seed check,
MCP comes out, which is great because it gives you a scaffold and a protocol around which AIs can sort of understand the tools that they have at their disposal and know how to use them. And so if you ask an AI to go do something and it has access to some universe of MCP tools, it can go out and say, okay, well, it's like you sort of, you got something to do in your house, you go to your garage, you see what tools you got, and you go, okay, I'll grab that screwdriver over there, it'll work and I know how to use
it. If I need instructions, it's right there. These MCPs in turn generally, hook up to some sort of an API. So you have Salesforce. Salesforce exposes their API. You layer MCP in front of that, and the MCP sort of interprets that API, for lack of a better word, for the AI. So what's great about this is it really simplifies the process of explaining to an AI what tools it has and what it can go to. But what it doesn't
do, it doesn't secure them. It doesn't limit them. It doesn't keep track of who's doing what to whom. It doesn't ensure that the AI even uses the right tool. And the challenge with it, it's like, here's a pipe, but there's no governance, there's no analytics, and critically, there's no sort of directing the AI with what it should do. So if you open, let's say, Salesforce, MCP to an AI, and maybe you use ours so that it actually
provides the ability to write, since most of them are read -only, you have this universe of tools at your disposal. Well, let's say you also then turn on the HubSpot MCP and you turn on the Google Drive MCP. Well, now you've got all these massive quantities of tools. And when you ask chat to go and do something that happens in Salesforce, the first thing it does is it pulls in every single one of these tools and
analyzes it essentially. By the time it has done that, it has consumed hundreds of thousands of tokens and it generally runs out of space before it even does anything for you. And so what we've had to build is what we call tool IQ. It basically understands the context of what the AI is trying to do and it only shows it the tools it should. So rather than saying to it, please go look up the the IBM account in Salesforce and it starts by saying, I'll check Google Drive to see if
it knows anything about it. It won't do that. And so there's all kinds of things that enterprises can use to manage these and this MCP access to make sure that it is cost effective. It actually works. People can get stuff done and it's secure. So for instance, I am allowed to go to Salesforce and delete an opportunity. I have that privilege, but My AI should not have that privilege. Even though I'm allowed to, I know it's stupid. And I know that if I do it too often, I'll probably
get fired. My AI, if I say to it, go clean up these opportunities and give me a report, it might say, well, clean up. What does that mean? Okay, yeah, I'll get rid of them. That'll do, right? And so it's more than identity. It's more than API management is sort of a new paradigm of a combination of security, making sure bad things don't happen from outside actors and governance to manage the, the pros and cons and quirks of
each of the agents and each of the humans that are involved. And I was also reading before you join me today that you often argue that governance must serve as a launch pad rather than a break again, completely agree, but absolutely. What does good governance look like when an agent is probabilistic, capable, and operating inside production environments? Because it will make some people nervous. Oh, it will. It will, absolutely. And, you know, you look at your self -driving cars right now, right? So you have your self -driving cars, but most of them still require
you to hold the steering wheel. Got to have your eyes on the window. If you don't do that, it's going to get angry at you, right? It's kind of the same thing here. You want to... narrow the edge cases, right? And make sure that essentially when you ask the agent to go do something, that it has guardrails that allow the things that are consistent with doing that thing and disallow everything else. Some people call it least privilege, right? It's how do you make sure, and this is
again why you're going to have lots and lots of different agents, because depending on what they are for each context, least privilege will look differently. Right? And so from my standpoint, what it is, is governance is make sure the agent has the capability of doing the things that are consistent with the job at hand and nothing else. And what you often do is you don't even give it those privileges at first. You first say,
I'm giving you no privileges. Go do these things and you actually watch the MCP and see, okay, are the attempts that the agent is making consistent with what doing that job looks like? Well, if so, then okay, we'll open that up and we'll let it go do them. And you often say the core question is not how the model was trained, but what will it do with the access right now? And how should enterprise build a repeatable way to test, measure and trust agents before they touch real customer
data and revenue systems? We've kind of, I would hope most enterprises have learned the lesson through not measuring things. That's how they ended up with no ROI. But now we're in firm IT territory, aren't we? Testing, measuring, et cetera. Well, they're certainly, depending on how dangerous a thing is that you're doing, how catastrophic it could be, certainly sandboxes are lovely things and testing is great. But we're actually interesting to see that... A lot of the customers we're dealing with are actually
fairly high -secure financial institutions and other folks that want this to be running, want MCP and want our governance to be running on -premise or in a private cloud, which we're capable of doing, because they do see that opportunity, but if they're going to go do that, they very much need to ensure that the outcomes that can happen are very, very limited, are tracked, and
are accessing things in a way that they can be undone if bad things happen. And I think that is a powerful moment to end on. But before I do let you go, anyone interested in exploring the world of AI, APIs and MCPs? So many acronyms there, but they want to make sense of it. Where should they go to find out more information about all things bond or AI? barndoor .ai and you can come and see all the stuff we're doing and check
it out. Awesome. Well, I'll add links to absolutely everything. Make it easy for people just to click and pop over. But before I let you go, I've got to ask, obviously you've got quite an eclectic mix of experience here and I love how you're frequenting the worlds of entertainment, stage, Broadway and technology. How do these worlds collide for you? Is there any lessons you've learned along the way? Do they? Is there any crossover there? Oh, there's tons of crossover. There's tons of crossover because bringing technology to a a fairly traditional and not, you know,
Broadway's a funny place. There are some parts of it are incredibly sophisticated. If you go to a Broadway stage and you go under the stage, they've basically built an entire data center there, right? It is incredible because all the lighting and video and sound, these are all highly, highly technical and computerized, and they do incredibly sophisticated things. And then on the business side, they're pushing paper, right? And it's really interesting to see that what the audience sees is often very, very sophisticated, but what's backstage is very, very traditional.
And I've worked with some of my favorite, you know, longtime producing partners to really help them reimagine that in a way that makes everybody a bit more efficient. But it's also, you know, it's all about storytelling. You know, if you want to see a great show, someone, you know, you arrive feeling a certain way in the... the cast and crew's job is to have you leave feeling a different way. And if I'm pitching a venture capitalist or I'm talking to a customer or I'm
on a panel at a trade show, it's all part of that. Well, I love chatting with you today and I will urge everyone listening to check out Bond or dot AI, anyone, especially if you're interested in having complete visibility and control over how AI accesses your enterprise data through MCP connections and that phrase alone, you will hear a lot in 2026. And I think maybe we should get you back on later in the year and learn more about the entertainment side of your Broadway life there too. But thanks for joining me today.
Great. Thanks, Neil. Wow, one of the many things I really enjoyed in this conversation today was how we connected past tech cycles with what is unfolding right now. We have been here many times and the older I get, the more I realize that everything happens in cycles and whether it's APIs and mobile to MCP and agentic workflows. or that AI is actually doing things. I think my guest brought a very grounded perspective on what it takes to move from demos to dependable
outcomes. So I will include links to Bond or AI so you can all dig a little bit deeper in how governance, access and control fit into the next phase of enterprise AI. And as always, this episode was about asking better questions, not selling easy answers. So hopefully it has helped you rethink how AI could operate inside real businesses your business please if you think it'll be useful share it with someone who is wrestling with similar challenges and more than
anything though thank you for listening and stopping by and i urge you to check out techtalksnetwork .com we have nine shows over 4 000 interviews with yours truly there and a whole heap of names in every industry you can imagine but i hope you enjoyed today's episode and i hope to speak with you again on AI at Work very soon. Bye for now.
What happens when AI changes the work long before a company changes the job description? In this episode of AI at Work, I speak with Vishnu Shankar, Chief Data Officer at Draup, about a workforce problem that is easy to...
Can a finance team trust an AI answer if the systems underneath it disagree about what revenue, cost or performance actually mean? In this episode of AI at Work, I speak with Matej Trbara, co-founder of Farseer, about...
If your organization switched off its AI systems tomorrow, would customers, employees or the finance team notice any meaningful difference? In this episode of AI at Work, I speak with Frank Böhmer, founder and CEO of...
In this episode of AI at Work, I speak with Rohan Kodialam, co-founder and CEO of Sphinx AI, about the difference between an AI system that sounds capable and one that understands how a business defines truth. Sphinx AI...
What should a finance leader be able to prove before giving an AI agent permission to act? In this episode of AI at Work, I welcome Danny Fields, Chief Technology and Customer Operations Officer at Avalara, for a...