Insight ON What Executives Are Too Embarrassed to Ask About AI Agents — Answered

Audio transcript:

What Executives Are Too Embarrassed to Ask About AI Agents — Answered

Jillian Viner (00:02):

Welcome to Insight On. The word agent has been in every meeting, every keynote, every vendor pitch for the better part of a year now. And if you're honest, the definition shifts a little depending on who's talking. That's not a small problem because the decisions organizations are making right now about how to build, govern and scale agentic AI are going to be hard to undo. So this week we're creating some clarity with three conversations that cover the full picture, what agents actually are and what they aren't, what it looks like when a large organization builds an entire ecosystem around them, including how to keep it from spiraling to chaos and what it really takes to govern agents at scale before they start multiplying faster than anyone can track. If you've been wondering how to get ahead of this rather than just react to it, this week is a good place to start.

Miles Ward (00:57):

An agent takes those same technologies, large language models and the compute infrastructure behind them and allows that kind of system, one, to persist independent of your presence and to take actions on your behalf. So not just bring text back to the screen, but go do something.

Jillian (01:16):

If that clip just made something click that hasn't quite clicked before, good. That's exactly what this episode is for. I'm Jillian Viner and this is Insight On. My guest today is Miles Ward, CTO of AI at Insight. Miles has spent eight years at Google before joining what was then SADA and has spent more time hands-on with this technology than almost anyone I've talked to. But what I love about Miles is that he has zero patience for vague. It's going to tell you exactly what an agent is with a great metaphor, what it isn't and what agent orchestration actually means and maybe most importantly, what every executive needs to go do right now, whether they feel ready for it or not. If you've been just nodding along in conversations, unsure if you really understand what an agent is, if this is the right one, this is the right episode for you.

Let's go. Miles, welcome to the pod. Hey,

Miles (02:13):

AA. How are you, Jillian? Looking good.

Jillian (02:15):

I'm good. I'm so glad to have you here.

Miles (02:17):

Thank

Jillian (02:17):

You. We have been talking to a lot of different clients, a lot of different organizations. At Google Next, we heard the word agent more times than you could. I mean, if it was a drinking game, we would all be on the floor.

Miles (02:28):

It's agent and governance and that's floor installed

Jillian (02:31):

For sure. That's right.That's right. But let's do a little bit of a reality check for listeners because you talk to clients regularly, you talk to engineering teams, you talk to people across the board and organizations. Sure. I want to level set here of what we mean when we talk about agents and agentic workplaces, agentic workflows, like all this terminology that some leaders might be bobbing their heads.

Miles (02:55):

Sure.

Jillian (02:56):

Just to start us off, how much do you think people are saying the same things when we're using these words?

Miles (03:04):

Sure. Every one of these terms, once they make it into the real world, they get sort of clobbered by marketers and the labels get confused. It's very, very clear and specific if you're hands on with this stuff all day, the differentiation. Everybody's used a chatbot before an interface that while it may record your interactions is designed to respond to you in that interface. An agent takes those same technologies, large language models and the compute infrastructure behind them and allows that kind of system, one, to persist independent of your presence and to take actions on your behalf. So not just bring text back to the screen, but go do something either in SaaS software that you have, communicate with the outside world, produce documents, transcribe objects. It's the action that's where the agent comes from.

Jillian (03:58):

Got it. Now are there different categories of agents? Because I'm thinking like a Google Gem. I can make that myself as an individual contributor. It basically is a prompt that I found myself using over and over again. I turn that into a gem and I just give it some new context when I open that gem and it just spits out the thing that I intended it to produce. So for example, I need to create an asset page for a new ebook we've published.

Miles (04:22):

Sure.

Jillian (04:22):

I give it the ebook, it reads it, and then it knows the output that I need an asset page. It gives me the meta description, the title of the page, et cetera. That's a gem, that's an agent.

Miles (04:33):

So you're right at the cusp of the difference. So I would call a gem a saved state for a chat experience. So I have the persistence requirement done, but if I put back in the chat block the text that you should copy and go paste in a document, that's still not an agent yet. An agent, you would take the request and it will hand you the URL to the document. It is taken the action of constructing the output you wanted, not part of it or not the input you might use to synthesize it. This is even more obvious for software developers. A year ago you'd ask inside of your IDE, maybe make some suggestions or use tab to auto complete. But the difference between those kinds of completions and making a request and receiving new files in your IDE, new objects that you can use to build applications to create content, that's how you've crossed over into an agentic task.

Jillian (05:34):

Okay. So I've heard of the terms like agent orchestration. What does that look like? Is this multiple agents that are now going off and doing tasks?

Miles (05:43):

Sure. I've checked of no businesses that decided to have exactly on spreadsheet in which all of their data and all of their tracking coexist sort of works to divide the context ever so slightly and set up teams to use the specialized interfaces that they need to be able to make sense of the business problem they have. In the same way, agents rely on a narrowing of context around a specific part of a problem or a specific problem area to be able to increase the accuracy of their responses. Everybody's heard of an AI hallucination. The real core driver of that in most cases is asking extremely broad questions to extremely broad models where it can be difficult to anticipate which places they are strong and which places they don't have enough context. By building an agent, you're narrowing the context to just the individual task you want to do and the data you need to do that task that really radically reduces the hallucination problem.

It also means if you think about the difference between for software developers, maybe you've heard this word monolith. I build one big giant application with all the different buttons and functions and when I want to make an update to it, I must update the whole thing. Everybody everywhere decided about 15 years ago, that's a terrible idea, makes development very slow. If we had on big agent for the whole business, we wouldn't get the specialization advantage and we wouldn't get the independence of improvement and innovation. So you need agent orchestration to say, "Well, this agent takes my input from customers and converts it to a dossier. This agent takes dossiers and converts it into proposals. This agent takes proposals and puts them into our primary systems for tracking. This agent takes those tracking systems and turns them into instructions for the next team. Orchestrating that workflow so that it happens seamlessly is observed and monitored so that quality stays high.

You can detect model drift and the other risks to these kinds of systems doesn't happen automatically. You need an orchestration platform to do that.

Jillian (07:58):

Okay. You've given me a couple examples there, but give me something even more concrete. Where have you actually seen this play out in a business? What's a good something I can latch my ideas onto to understand how this

Miles (08:11):

Agent works

Jillian (08:11):

For this?

Miles (08:12):

We were on stage. I'm wearing the most fabulous jacket of all time. It is quite

Jillian (08:16):

Fabulous.

Miles (08:17):

Because I was on stage with the folks from Mattel. If you can't go Barbie, you got to go even bigger. So we worked together with Joseph there. He's a brilliant leader on their business. They bring in data from quality control signals like what's happening in the manufacturing plant, who's complained about a toy online, where are their returns for products to get a sense of what's working and not working in their environment. And their old method of this is literally like clipboards and paper and tally sheets of how many times somebody complains about a given thing. All that's been digitized, all that runs in Google's environment and BigQuery, but the workflow they want is to be able to go from a signal in their dataset that says," There's a problem with this toy we should inform to the manufacturer and maybe their suppliers what needs to change to solve the problem with that toy.

"You could maybe take one agent and synthesize the first part of that report, maybe build a PDF document that shows what is the nature of the problem. Put some of those places are in different countries. Some of them have contracts that determine exactly what that document's supposed to look like. So now you need another agent that takes the input PDF that you and I can read and turns it into one that's written in a different language formatted for a different client. You also need to deliver it to them and confirm that they've racted on that action. So check in with their management systems to ensure the cycle was closed to end to end. So if you build each of those agents on their own and they are not aware of each other and they don't have connections to each other, each of the handoffs from tool to tool are done by you and me and that sounds time consuming.

So having an orchestration platform means they can tell that all of the agents are doing their job. They can observe the flow of information between them and confirm that the end-to-end cycle's actually working how they want.

Jillian (10:19):

Interesting. That makes perfect sense and it sounds like the kind of judgery work that no human wants to do anyway.

Miles (10:25):

That's right.

Jillian (10:26):

And definitely increases efficiency to make sure that those complaints are actually addressed. I mean, changes are made. So that's agent orchestration. Another term that's get kind of newer but getting thrown around a lot is this idea of like AI harnesses.

Miles (10:41):

Yeah.

Jillian (10:41):

Explain that to me.

Miles (10:43):

Sure. If you've ever ridden a horse bareback, you recognize that while they are of high horsepower, it's somewhat an uncomfortable way to consume said horsepower. And so we created these structures called harnesses that allow us to put that horsepower in gear in a way that's just ever so more comfortable for the writer. Everyone, it's my view that every job, every function in every role, in every business, everywhere will have AI augmentation over time, but that can be either comfortable or uncomfortable. If your task, say you're a video editor and you're going to put together the best clips for our podcast today, you could pull open a video editing suite and individually alter the frames that are involved and double check it yourself. You could also use tools like Google's video. There's a thing called VIDS that does some of that stuff automatically. VIDS combines LLM calls, individual parts from the model with software that's already been built to use that model for that use case.

So rather than forcing folks to sort of figure out how to type into the chatbot the correct incantation to get what they need, instead you go to a tool that feels maybe a little more familiar software that's shaped and has the right kinds of buttons to feel like the workflow you're used to, but on the other side, there is an LLM making magic happen. The place where that's happening most aggressively is in software development because every company building LLMs is building them out of software. And so they really need this harness really badly. And so we are happily inheriting the largess of their investment. Software engineers everywhere now have many options, whether that's anti-gravity or cloud code or they're using the building blocks and codex or parts of GitHub have these same kinds of functions where multiple models, multiple agents encapsulated in a user interface so that as a user, it's much easier for me to get the best value out of that intelligence, get as much horsepower as possible.

Jillian (13:03):

All right. I can't wait for your metaphor for this next one. Are

Miles (13:06):

You ready? I'm ready.

Jillian (13:07):

Headless agents.

Miles (13:09):

Sure. So a great example, if you've gone into any of the chatbots and I think a lot of users early on you would type a prompt and then immediately it would sort of spit out the answer. You'd watch it sort of assemble that answer over time. As we got to what are called thinking or reasoning models, there are many steps to that analysis and synthesis and customers got confused, worried. I don't know what the model is doing. I'm worried that chain of thought doesn't make sense. And there was really strong feedback from users. I want to see what's happening. Explainable AI was a pretty big label for a while. So now in all the interfaces you go, I am thinking, and then it shows you some of its work. Usually it's in like a lighter text color and there's all this text that's going by.

Those are tokens. You're paying for those tokens. That's exactly as expensive as the end output. Your software like a harness does not need to know how it got to the answer. It just wants those answers. So a headless agent doesn't produce all of this user interface and feedback to people to help them feel comfortable with the workflow. It's designed to be talked to by software that does not care. It just wants the answer right now. And the result of that is pretty important. One, all those tokens that were used to help us feel comfortable, you're not paying for them anymore and these are the spendy kind of tokens, so that makes a big difference. All those tokens take time to construct. So that's latency in each request. If you're building a software application for millions of users, latency is the dark side. It also makes the output that you're reviewing that much smaller.

And so the quality control exercise, the work to make it so that the results are as high accuracy as possible has a narrower problem space to operate in, so it improves the performance. So that's the difference between things like using Gemini as a front end on the web and calling Google's Gemini CLI where all you get back is straight on the terminal direct text of the specific output you want.

Jillian (15:26):

Okay. Thank you for explaining all of those.

Miles (15:28):

Yeah.

Jillian (15:29):

Great terminology for us to understand.

Miles (15:31):

Appreciate it.

Jillian (15:31):

Which ones do leaders ... I'm talking like CTOs, CEOs. They're talking to the board because the board's asking, "What's our AI strategy?" Sure. Which ones do they really need to be tuned in on and really understand, paying attention to and asking their teams about?

Miles (15:45):

So I think it's pretty critical at this point given the steps forward that had been made around governance. I think you kind of had a pretty wild, wild west in 2025 around these tools. I don't know how many times at this conference I have heard and read the word governance. It is clear that our customers want guardrails, want controls and it's not enough to sort of per company, per team, figure out how to sort of glue this together and get to a reasonable tracking. It's very useful to have trillion dollar global multinationals invest in assembling exactly the kinds of controls that you need because they had to assemble them for themselves. So there's a new set of features. We actually did a lot of advisory, submitted some bugs against these features in early trusted testing for what's called Gemini Enterprise Agent Platform. What that does is one like Gemini Enterprise allows you to build agents easily, make that simple for people to use, but it adds all of the evaluations because agents are non-deterministic.

You can't actually test them. You have to do evaluations instead. It's going to answer the question slightly differently each time, but doesn't mean really that it's wrong. That's the difference that you have to check in on. It also does cost control, has all of the structures for understanding which of your users or power users are not. Maybe that means you need to assign more training or do more of the technical enablement. That broader platform I think will set up businesses to start to identify the areas where maybe the workflow they have isn't as big as all software development and they could start to work to build a harness of their own. So I think there's going to be harnesses for a whole bunch of the kinds of software applications that we're used to, but the majority of software in the world by count is weird stuff built in one company for their one weird thing.

And if that's what your business earns money on every day, it should have a harness too. I think we're going to start to see businesses rebuild their internal tools to take advantage of all of these new skills and capabilities that come frankly at incredibly low cost as a result of all of the AI innovation that's happening so that they really can put in gear this newfound intelligence.

Jillian (18:10):

If you're talking to a client and they're like, "Miles, tap you on a shoulder and lean in. I got to ask you something. This is embarrassing." What are they most likely to ask you? What's the thing that they're kind of confused about?

Miles (18:22):

I mean, I'll tell you they go right to the top. What are we supposed to do about AI? I mean, it's really that big. If there isn't board anywhere that isn't being held accountable to demonstrate that they're doing something to move the business forward and my response to that is fairly straightforward You have to do, must do. You are out of your mind if you are not doing these three things. If your software engineers are not using artificial intelligence harnesses to build software, you are now more than a year behind and you will not catch up. There is no level of practice and acumen that you can assemble that will outrace the access to artificial intelligence in this workflow. There is an impossible amount of investment in making that better and it will only get better than it is now. These are the worst harnesses we will ever have.

These are the slowest invitations we will ever have. And I'm already watching just totally ridiculous levels of increase in performance for teams that would call themselves 10 years behind and teams that are right at the front edge. It does not matter. Old code, new code, weird broken leaky data systems, brand new modern stuff, everybody benefits from this improvement. Your employees are experimenting and interacting with AI, whether you know it or not. There used to be this thing called Shadow IT, then there was Shadow Cloud, now there's Shadow AI. If you're not putting basic tools into hands of users, they will find their own tools and that's just a really material security and governance risk for every business. So I think we have to encourage companies to get a first party deployment of these systems, make sure that they have folks onboarded to those tools. We have a lot of customers and employees that are ... This is kind of an entirely new way of working.

And so the training and onboarding and enablement and adoption management, the change management in that cycle is almost as important as the technical purchase. The third thing that's happening is while those kinds of systems, both the software development environment and the environments where every employee gets access to the power of AI, both of those systems absolutely and totally depend on your data platform. And there's really rich tools to connect into some of the SaaS resources and you can use some federation tactics to create some opportunity, but all of that pales in comparison to a well-structured data warehouse with the goods in it. That data warehouse for every business just got 10 times more valuable and they're probably behind. I was at dinner last night and one of our customers leaned over and goes, "Have you ever really seen good data? Does it exist?" That's a fair question.

I've been in some hot startups where, yep, everything is well structured. I've been in departments where, yep, everything's really incredible. I think Mattel has good data. I saw it in purpose, but I would say the vast majority of us do not. And I think there's this hope that maybe AI will make it so I don't have to do all that data integration and data science work and all this sort of rationalization and structuring of that information. I promise you, that's just not the deal. You're going to have to do that work. That work just got a whole bunch more valuable. It's probably time to dive into it. It's easier than it's ever been. The tools are making that faster just like everything else, but it's a step that has to be done to set up your business for the next stage of growth.

Jillian (21:50):

With that being said, what in your view is the single most important thing a leader needs to know or go do right now?

Miles (22:01):

That's a long, long, hard test. I think probably a recognition that the change rate is exceptionally high and it's going to stay high for a while. So looking at the teams in your business and identifying who really has that experimentative spirit, who are the builders, who are the creators, who is absolutely furious about the current state of things, who can't help but get in trouble with security and fight with the finance people, those cranky folks, if you can sit with them and start giving them tools, I promise the dividends are material, they will teach you and learn for you the kinds of change that have to happen at your business and set the first examples that everyone will learn from. I think as an executive that demands some humbleness, right? This is brand new. It's learning. No one's a 20 year expert at generative AI.

It doesn't exist. So we all have to kind of put our learning cap on, get out of the Dunce's corner, get into gear. You got to hit the books, but I think this is one of those places where the dividends come fast and heavy.

Jillian (23:20):

This feels like a moment for the kid who was always disruptive in class.

Miles (23:24):

Oh yeah.

Jillian (23:24):

This is his moment to shine.

Miles (23:27):

Did you watch my childhood videos? What is the deal?

Jillian (23:31):

Miles, thank you so much for clearing the air. Hopefully you've answered a lot of questions that our execs have been wondering about but too afraid to ask. And if not, we know how to find you.

Miles (23:41):

I'm right here. Happy to help.

Jillian (23:42):

Thank you for the time.

Miles (23:43):

Cheers.

VO (23:44):

Thanks for listening to this episode of Insight On. If today's conversation sparked an idea or raised a challenge you're facing, head to insight.com. You'll find the resources, case studies, and real world solutions to help you lead with clarity. If you found this episode to be helpful, be sure to follow Insight on, leave a review and share it with a colleague. It's how we grow the conversation and help more leaders make better tech decisions. Discover more at insight.com. The views and opinions expressed in this podcast are of those of the hosts and the guests and do not necessarily reflect on the official policy or position of Insight or its affiliates. This content is for informational purposes only, should not be considered as professional or legal advice.