Summary
How should CIOs set realistic expectations for AI in 2027, plan for contingencies, and make hard choices about priorities?
Isaac welcomed special guest Jason McKay, CIO of Riggs Distler, who shared his company’s deliberate and purposeful approach to AI implementation, deploying specific agents for utility construction standards and safety information that save significant time compared to manual processes. The discussion covered key challenges including tight supply chains with 10-15% quarter-over-quarter cost increases for DRAM, $5-20 average cost increases per million tokens for top US LLMs, and 80% peak in server prices above 2025 levels. Participants discussed the paradox of C-suite leaders wanting rapid transformation while practical implementation requires careful scaffolding, data preparation, and testing – with Joanne emphasizing that 56% of CEOs report no positive ROI from AI initiatives. The group explored hard decisions around budget allocation, whether to focus on resiliency versus growth, and the trade-offs between using frontier models versus open-weight models, with Joanne advocating for organizations to “own their AI” through proprietary models built on open-weight foundations to reduce costs and improve efficacy.
Speakers
- Host – Isaac Sacolick
- Digital Trailblazers – Derrick Butts, Martin Davis, Joanne Friedman, John Patrick Luethe, Liz Martinez, Heather May, Joseph Puglisi
Discussion
- Big Expectations: Where have you seen the biggest gap between what leaders expected in 2026 from AI and innovation versus what got delivered?
- Tight Supply: How should Digital Trailblazers plan AI and other innovation work when compute costs, infrastructure availability, and evolving skill requirements can change mid-year?
- Hard Decisions: What steps should Digital Trailblazers take in planning and delivering initiatives so that, if supply tightens, their work is less likely to be impacted?
Research
Dashboard
Transcript
[00:00:00] Isaac Sacolick: Greetings, everyone.
Welcome to our 196th episode of the Coffee with Digital Trailblazers. Today, November 13th, 2026.
Hello, David. Hello, Cheryl. Welcome to Jason McKay, who is a good friend from SIM Fairfield Westchester. We’ve been trying to get Jason to be a special guest here on the Coffee with Digital Trailblazer. So he is here with us. And today we’re talking about AI’s 2027 reality gap, contrasting the big expectations our organizations, our boards, our leadership has around AI, different elements around the tight supply and the supply chain, whether it’s chips or boards or talent, and then some of the hard choices that we have to make as leaders about what we actually prioritize. Hello, Olu. Hello, Steve. It’s good to see you back from Houston. Haven’t seen you in a couple weeks.
Hello, Namatha. I don’t know if I’m pronouncing your name properly, but great to meet you yesterday at the Sim Fairfield Westchester event.
And I’m really excited to be talking about this. All I can say, everybody is I’m kind of glad we don’t cover the news very much.
Here at the Coffee with Digital Trailblazers, we try to talk about areas that our transformation leaders need to consider around AI growth opportunities, operations and resiliency. We only sneak the news in, but geez, we couldn’t stop talking about the Microsoft Outlook outage that was hitting most companies yesterday.
And then obviously the big news this week with Nvidia. And my question is, is that going to be the spark that busts the AI bubble? And that’s just a really good backdrop to our conversation today about the reality gap. Right? We plan our organizations based on the horizon. We see the horizon that we can anticipate.
And right now we’re at a point where the wind might be blowing in new directions. And we’re just not really sure how the media and how business is going to respond to all this. Different news items of the week between Microsoft and Nvidia and we have to be able to respond to it. So in terms of big expectations, 81% of enterprises report faster decisions in execution, but only 25% they use AI to deeply transform the business. So again, we’re using AI a lot of stuff in a lot of different places, but we’re not doing the types of things that are going to drive growth or fundamentally change how we’re operating. We still have ways to go around that. 40% now let AI systems execute decisions without human approval, but 23% say accountability is unclear.
When AI gets it wrong. We talked about some of that last week or a couple of weeks ago when we debated human in the loop and agentic agents and around human augmentation. We’re still in a transition period about which of our agents we’re going to actually let become autonomous. Hello, Tony. Hello Ronaldo.
Welcome to this episode.
We’re talking about AI’s 2027 reality gap. We’re experiencing tight supply. There’s a 10 to 15% quarter over quarter increase in costs around conventional DRAM contract prices as AI servers absorb supply. So those of you who are thinking about data centers or just doing AI at scale, our costs are not going to go down anytime soon.
There’s been a $5 to $20 average cost per million token increase for top US LLMs over the last year.
Server prices peaking 80% above 2025.
These are data points from BC Tree, from TrendForce and from Deloitte. And so we have some hard choices to make.
5 to 17% AI share of IT budgets and BCG 2026 survey. Right. So we look at, you know, how we’re going to balance our budget going into next year.
For most organizations, that 12% is a lot of money in the IT budget. And we’ve already had the discussion around making some hard choices around the budget. Today we’re going to make some hard choices around what areas of AI we’re going to spend on and have that discussion around how to be thinking about forecasting and how to handle the relationships with our executives about what’s changing. 71% track total AI inference and API spend, but only 51% track token volume and only 44% track per unit in first causes is coming from Deloitte. I think we have a bigger problem, folks. We’re not doing a good job of tracking value. We talked a lot about this.
Our value officer Liz is not here today. Anybody can pick up that part of the conversation. But this came up again at our Fairfield Westchester SIM chapter yesterday talking about tokenomics. And I had to pull some folks on the side and say, look, we better be talking about value and how we measure that before we start penny pitching. My talk.
My article this week on AI, shadow AI, which was really a takeoff from our coffee hour from a few weeks ago. If you start tightening the budget around what people can spend in AI, it will incur more shadow AIs. People just go use the AIs on their phone and other devices when they can’t, when the enterprise isn’t paying for it. So we’ve got a problem There. Right. We have to be talking more about AI’s value, so thank you. If you want this particular slide with this data, the links to the research star cio.comcoffee/episode-196 will have it once we complete today. And you can go to star cio.com Coffee, which has it up right now. Along with our discussion questions, I want to invite Jason McKay to the floor, our special guest. Today we’re going to talk about big expectations. First, Jason, where have you seen the biggest gap between what leaders expected in 2026 from AI and innovation versus what was actually delivered? Jason, introduce yourself and just your thoughts around the big expectation gaps.
[00:07:17] Jason McKay: Yeah. Good morning, everybody.
Talk to my wife about the expectations of me. Try to keep them low, you’ll never be disappointed.
But I’ve been the CIO for riggs distler for 12 years and we’re. We upgrade the utility infrastructure for companies and utilities.
So the big expectations, you know, and we’re, we have the ambiguity of fog around this whole topic as well, you know, but the big expectations is everyone thinks that this is just a flip of the switch, especially at the C suite for some people that are not, you know, technically, you know, advanced, especially with AI, and they don’t understand, you know, a lot of work has to go in place to actually create an agent that’s fit for purpose, fit for use, and defining the problem you’re trying to solve. Do you have the right data and are you testing it effectively, creating ownership around it, documenting it, lifecycle, managing it?
You know, and a lot of the basics still come back, you know, your data, your processes, and defining the problem you’re trying to solve. And, you know, we’re still struggling with this. And we had a lot of expectations that people thought we could just do this, this and this.
And what we’re finding is, you know, the shadow, it was discussed, but when you look at what people can do with AI, like Isaac said, you know, on their phones, plus with things like Python, you know, what we’re trying to avoid is the MacGyver’s new tools become AI and Python. You know, we have all this shadow it trying to do all kinds of crazy stuff to cobble things together that, you know, all of us have probably run into someone does something and they abandon it. But, you know, the big expectations is, you know, how quick it can be done, can it be done? And then, you know, Isaac and Joe and I were actually talking about this.
Just because you can do it, should you, should you be leveraging other platforms? And other systems that exist, you know, but the big expectation is what’s been done and how do you measure the value. And for us, the value has been really measured in time savings. You know, things that would take three weeks now take three hours. But, you know, we’re still struggling with that. But, you know, we’re trying to document everything we’re doing and keep track of what everything you’re doing, you know, because the other question I was asked is, you know, do you centralize this or do you push it out to the edge? And unfortunately, the answer is kind of both, because you want to centralize what people are doing, but you want to give them the advantage of taking the training wheels off the bike and let them ride the bike themselves.
[00:09:45] Isaac Sacolick: Some great comments there, Jason. I’m wondering if you can share what you can share about what you’re doing at your company around agents. And we heard, we were at the CXO Spark event earlier.
Or is it called CXO Spark? I forget. I think it’s called CXO Spark Spark Executive Forum. That’s what it’s called now.
And we heard one panelist say their company has deployed 120,000 agents. That company, I think, has over 200,000 employees, just for reference.
But Gartner has come out and said by 2028, the average Fortune 500 enterprise will have 150,000 agents deployed. I’m going to guess, Jason, you’re nowhere near that. But I’m wondering what you could share in terms of what areas of the business that you’re focused on.
[00:10:34] Jason McKay: Yeah, when he said that many agents, I was like, yeah, my eyebrows almost came off the top of my head.
No, we’re very purposeful and deliberate with our approach because we do work for utilities and we do government work. The government work I really haven’t touched with AI yet. But when it comes to the utilities, it’s very specific. We build an agent to understand the construction standards and the safety standards.
We isolate it. We don’t let it go out to the Internet. We train it vigorously. And then what we do is we actually push those agents out through teams. So people in the field, if they’re in 100 foot transmission bucket truck at the top of a three phase transmission tower and they have a question about the swing arm tolerances. We’ve created agents so people can, you know, type in the question, get the reference, click on the reference, and then get the exact page that the reference might be on.
But before they’re in there, they’d have to go down in the bucket, get in their truck, thumb through a thousand pages of PDFs and stuff. But that’s where we’re really finding the value. And then creating like bill of materials for like a hundred transmission tower project. It used to take a couple weeks, that takes a couple hours. You know, we have a bunch of other things, but where we’re finding the most value is just like I said, saving time. Things that took three weeks, going through hundreds of Excel spreadsheets, hundreds of PDFs, drop it in an agent SharePoint and just start going.
But that’s where we’re finding a lot of value. And then the other one I did is I, I created a master agent called Dudley because our founder is Dudley Riggs. And you know, we’re putting all kinds of information that people ask about that all the time. Our safety history, revenue history, our acquisitions, Riggs Distiller history, all the type of information we do for pre qualifications.
And now the interesting thing is people are referring to Dudley almost as a person because they’re asking me, hey, could Dudley do this? Could let Dudley do that. But we’ve been very deliberate, very purposeful, very narrow in scope and trying to take a very crawl, walk, run approach as we continue to build them out and increase functionality.
[00:12:37] Isaac Sacolick: Very interesting. Before I let you go, when you deploy it to the field, do they know it’s AI or you or is it guised inside an app or something else and they just look at it as more technology that they’re using?
[00:12:54] Jason McKay: No, we’re, we’re very transparent so we’re telling them it’s AI and when we build these agents, we actually go to the field people and we test it with them because you know, I want them to ask the crazy questions they might have in the field. And a lot of times I’m in the field with them as we’re testing these things out.
So we want their buy in and understanding of what we’re trying to do. And it’s not, you know, we’re telling them this is very specific to this utility for these types of questions.
It’s not going to, you know, tell you, you know, give me the F top five Harry Potter books or anything like that.
So they, they recognize the value and they, they’ve told me like, you know, I used to have these safety questions and sometimes it would really be hard to get an answer.
But now they have a, you know, it’s a quick referential, I mean it’s not like, you know, crazy stuff that agentic per se, but this whole referential use of AI for us has really been a game changer for people in the field and people doing bids.
[00:13:51] Isaac Sacolick: And there you go, folks, some real examples from an active operating cio.
The gap between setting unrealistic expectations and what, you know what, what Jason’s doing is setting a low bar, testing in the field, being specific and deliberate. What’s going out there.
These are all best practices that we’ve discussed at previous episodes of the Coffee with Digital Trailblazer. I’m going to go to the floor now and start with Derek. You know, again, I. We’ve talked about this, Derek. I’ve heard some CISO, some people say we don’t need CISOs anymore, that the SOC is going to be taken over by agents.
I don’t know if I want to set those expectations and pretend like I don’t need security people anymore.
How do you set realistic expectations when we need AI in the security field? But AI isn’t going to take over what people do in security.
[00:14:54] Derrick Butts: Well, I think it goes back to realistic expectations. I mean, some of the things that Jason brought up with the costs and extra things that people found out about, they weren’t expected to. The tokenomics, the things that you mentioned before, these are all real things. You know, we look at the gaps.
Definitely you’re going to need a ciso, definitely need a cio. You need people to understand how this stuff comes together. But I think they also look at, you know, what happened in 2026. We proved that AI can we work? We prove now it does the things that it can do, but now can we depend on it? I looked at a lot of different impressive pilots have gone through, introduction of governance, introduction of cyber security, data quality integration, all these things, operational resilience. But when you get to AI safely in production, it comes with different conversation. So I would look at, you know, when you’re looking at these things and the people need to be part of it. Lessons going forward in 2027, we’re not going to ask now does AI really work? The question should be asked, can the business really trust it? Can I architect AI with resilience to allow the business to recover from it and to continue to operate when something goes sideways. I need to look at now the financial impact. As we talked about the cost that people have seen with the overruns. It’s one thing where businesses have seen, you know, the models when they put them in there, they’re paying more cost now to sustain the model itself because of all the data pipes, the orchestration, observability, the logging, the data production, all these things now have become factors. I think they need to look back at what’s really important, what’s critical, what’s not, and kind of analyze and kind of chronicleize what’s going to make sense as far as moving forward. But you need somebody like a CISO or CIO to help drive that vision until the bots can become more efficient and run it on their own, which I think is still down the road. We still need some sort of intervention to make sure, sure the accountability is there.
[00:16:37] Isaac Sacolick: Thank you, Derek. John. And then Joanne, let’s just try to keep this real, right?
John, in your case, your development teams, your colleagues want to do a lot of wild things with AI.
And that sets this bar that we’re going to be using AI for tons of things and we can reduce technical debt down to a zero, we can move everything to the cloud really quickly.
How do you set realistic expectations when you’re working directly with teams?
[00:17:09] John Patrick Luethe: Yeah, it’s really hard because in an organization you’ll have 10% of the people that just really flourish with AI. They’re self learners and they’ll figure out how to use it.
And that’s really like the really early adopters. And then you’ll have 80 to 90% of the organization that really doesn’t know how to use AI in their daily activities. And so I think the thing that’s really missing is the realization at the management level that this is really an adoption, an adoption issue. Right. And so I spent a lot of time helping companies understand, like, you know, where are they using AI and you know, what can we do to move them to the next level of maturity. And so a lot of times, like in a, in an engineering organization and a development organization, the developers are the first ones that start using AI.
But there’s a lot of levels of maturity. It’s not adopting it, it’s not binary. Right. And so a lot of times the biggest return on investment is actually enabling the other groups, the business analysts, the product owners, the designers, and, and then on the QA people. And so I, I think if you can, you know, if you went into a factory and if you took the machine right in the middle and you spin it up a hundred times, that’s great. But it just creates problems before and it creates problems afterwards. And so I, I think if you really want your organization to do this right, you have to like, how do you get the whole organization like working in a different way using AI?
[00:18:35] Isaac Sacolick: I completely agree with that. I mean, John, I mean you know, there’s an interesting discussion about, you know, this extension of citizen development and whether or not, you know, or how you’re allowing your employees to create agents and how to do them in a systemized fashion. So there’s the right governance built into it, there’s the right security and resilience and built into it.
And then the flip side of it around, you know, what do you want your developers to be doing?
You know, if they’re coding less and, you know, you’re impairing more of your organizations to deliver the agents that you’re building, you know, what is a developer responsibility in here? And it really does come down to the entire scaffolding. Right. We’ve always known that building is easier than supporting and that the coding part is easier than the upfront requirements and the latter part of testing and getting your data ready and supporting what you built. We still have a lot of work to do. I’ll let you comment on that before we go to Joanne.
[00:19:48] John Patrick Luethe: Yeah, I very much agree on that one. And if your developers now two times as productive, ten times as productive, it just, it causes a problem for the, the business analysts and product owners, like, they, they’re going to like, run out of stuff to build. And so that creates an organization issue, like, do you have too many developers or can you, can you figure out like, how to, you know, keep them fed? And then it creates a massive problem for the support and operations teams and the qa because it’s like they have so much more stuff to test. And so I just, you know, I just, I recommend every company just really kind of look across its whole, whole the organizations. And if you’re going to be using AI, try to think about how you can use it in a way that goes across the organizations. Look at the seams where people are interacting with each other and the things that are passed from one group to another group.
[00:20:36] Isaac Sacolick: Thank you, John. Joanne, welcome to the floor.
We don’t have to worry about expectations, do we?
[00:20:43] Joanne Friedman: Yes, we do.
I wish I could say that it was different, but
[00:20:48] Isaac Sacolick: Joanne, I took a guess. We’re going to agree on this one.
[00:20:52] Joanne Friedman: Yeah, well, you know, miracles do happen. And by the way, it’s snowing where I am.
Yeah.
Welcome to Canada and winter.
Anyway, the.
One of the points that Jason made, I want to elaborate on a little bit. Jason, congratulations for being purposeful.
Most companies are not. And one of, one of the things that I hear more often then I would like to hear is at a board level particularly, what do we do? Where do we Start. What do we do with it and how do we measure the financial impact?
Well, here’s my first stat for the day. 56% of CEOs report that they either have no impact or no positive ROI from AI. And that’s not, according to me, that’s a PwC global CEO survey.
The other thing is, again, to Jason’s point, there is a lot of work that goes into this. Most of you know that I have a startup doing this 18 months.
And the 18 months is not just spent tinkering with vibe coding. It’s figuring out all of the ways that you need to add things like authority, agency, how you, how things are going to be adopted, which systems of record are you going to rely on for determinism?
Because if you don’t, you’re going to have a lot of hallucination, a lot of drift, you’re going to constantly be maintaining models. So the expectation was I’m going to have all these wonderful demos, I’m going to see everything, I’m going to look at a lot of different stuff. And then there’s this giant gap between demo and production, and that gap is where everything stalls. An agent can be deployed, the authority is unclear, the execution stalls.
Well, who’s on the hook for this? Who’s accountable for it? How much freedom did we give the AI or the agent, I should say, in order to do things? And so that’s one area where expectation and reality collide.
The second is humans connect the dots and we weigh trade offs. We do this intuitively. We do this on remote, and I mean not remote distance, but on remote, like it’s automatic.
Agents have to be taught and have that built in.
If you don’t build those capabilities in the scaffolding for an agent, don’t bother.
And the more pointedly purposeful they are, the more ROI you’re going to get from them, the more you’re going to have a capability that you can then not have to do 150,000 agents. Because how much of that is duplication?
How much of that is really reaching in places where you already have an agent to do it? It’s given another skill.
So there is a lot of learning that has to go into systems Design around AgentIC AI and what you can really use it for? The last point that I want to make is agents don’t govern themselves.
Derek talks about resilience.
Well, to be resilient, you also have to have authority, agency, accountability, and so forth. All of those things need to be designed in.
They don’t just automagically happen. And so that’s another gap. So where I’m seeing them is if you’re just starting with Agentix or even with Generative AI A, be purposeful, be very selective in how you prioritize the product that’s going to give you the most.
What I would say is not bang for the buck, but rather opportunity to solve a business problem. You know, we take it from the point of view of time.
Right? Time is an operational gap that every company has.
How long do you have to wait for somebody to make a decision? How long does it take to get a record from a batch?
Where in your business processes are there very specific gaps that you can identify that are time based? And we chose time not because it’s just. Not just because it’s manufacturing, but because overall that’s where margins leak, opportunities get lost and revenue drops.
So choose a focus for your agents and then you’ll be able to meet your expectation level in 2027.
[00:25:40] Isaac Sacolick: Joe, we’ve got a paradox to deal with. The big expectations that our C Suite is looking for.
The realities that John and Derek and Joanne are talking about in terms of scaffolding, agency determinism, resiliency that we have to build. This is not anything new for CIOs to manage.
And then what Jason is doing and others are doing, which is like start small and be pragmatic and purposeful.
And you know, I’m looking at an article that my colleague@cio.com wrote the title of CEOs have had enough of incremental AI.
Okay. And quotes this data point. 70, 77% of business decision makers, including C suite leaders, directors and managers, believe their organization are dedicating too much of their AI investment to incremental improvements. What I call reshaping the business instead of transforming or reinventing it. We got a problem, Joe, of that expectation. Scrap gave you a long Runway to think about how you’re going to respond to that.
[00:26:59] Joe Puglisi: So I’m going to talk about two things. First, I really can’t see the.
I want to frame this as that the hammer and the nail syndrome. You know, when all you have is a hammer, everything looks like a nail.
I think it’s imperative that at board level, people understand that AI is just one more new technology that’s come along and will be perceived as a panacea. Well, it’s going to fix all of our problems. It’s going to make our employees faster and more productive. It’s going to make our customers happier.
You know, we can do those things, but you have to be realistic about what are the real problems that stand in our way of achieving those goals.
Just understanding and then being able to use AI is, in my view, a small part of the problem.
We have to look at the organization and find where are the speed bumps, where are the deficiencies, whether that is at the data layer or the infrastructure layer or the knowledge base of our employees themselves.
Do we have issues with our supply chain?
Because, you know, AI is not the only thing that people are spending money on, right? We have to spend a lot more money on diesel fuel these days. We have to spend a lot more money on cloud storage.
The problems are numerous.
The first thing is let’s not just focus on AI as curing everything that’s wrong with our company.
We have to figure out where the problems really lie and then bring the right solutions to bear on each of those problems.
If we determine that AI can in fact transform or perhaps even reinvent our company, then it becomes a question of how do we transition from our current culture of functions being in boxes, in an org chart where data or processes cross those boundaries, to a reinvented version of our company where I execute a process from order to cash through a chain of agents who are properly governed. Joanne. Who deliver value.
Liz.
And who don’t give away the store.
Derek we have to address all of those aspects of it.
They don’t cost us more than the employees we had before.
And importantly, we don’t disrupt the company in a way that causes us to fail because we’re still dependent on old systems and, and this new value stream or this new process, end to end process that has to fit into our current environment, at least for a period of time. So lots of things to take into account there, but those are the two points I wanted to make.
[00:30:00] Isaac Sacolick: Joe, I have to go after the softball because you’ve mentioned the word reinvent. Joe already knows that I’m working on a book.
This would be my third book. It’s tentatively titled AI Reinvention.
And it’s coming from the perspective that some of the constructs, assumptions and practices that we’ve been building up for digital transformation need some rethinking to be able to support AI.
And what I’m really referring to is I think our businesses and our careers are going to look very, very different over the next few years as AI continues to go from small pilots, lot of overhead around governance, around resiliency, to a point where not only we have organizations with hundreds of thousands of agents, but they’re going to be orchestrating work together. More of them will be doing autonomous work and they will be doing a lot more than just being our assistant. So we have a long journey ahead of us. We’ll bring Jason back to talk about the tight supply and thinking about how he thinks about it. But on the right hand side, I just want to draw everybody’s attention for our upcoming coffee hours November 20th. That’s next week we do this episode. I think this will be our third year talking about AI for social good.
I’ll have the CIO from the Ronald McDonald House, Jared Bell here with us. I think we’ll have one other CIO or CEO from a nonprofit talking about doing more with less when the mission can’t wait. And that’ll be our episode to sort of get everybody warmed up for giving Tuesday in between. After that, the 27th will be off for Thanksgiving. We’ll be back on December 4th. 4. This was a conversation that came up last week and I said we’re going to do this one right away. We’re going to clarify business versus Digital versus AI strategies.
How do we think about the scope of them, who the audience is, and how often do we have to update them? Do not miss the December 4th episode. Our December 11th episode will be digging deep into AI governance that employees will follow policies, practices and platforms. This cannot be a Word document on your Google Drive or your OneDrive. This needs to be something that becomes part of your culture and part of your implementations. We’ll be talking about AI governance on the 11th and then please bookmark the 18th. We’ll be celebrating our 200th episode and talk about what we’ve learned, what’s changed and what’s next.
And that’s just going to be a fun episode. It’s just going to be our standing guests here. My expert advisors will have some fun with it. But do bookmark your calendars for all these episodes. You can find an add to calendar button.
Previous recordings Ability to subscribe on Apple, Spotify and Amazon. Go to drive.starcio.com Coffee.
You can get our whiteboards. You can get our research slides. They’re all available from drive.starcio.com Coffee Jason, welcome back to the floor. I want to hear your closing remarks on expectations and then let’s pivot talk about supply. How should digital trailblazers plan AI and other innovations work when compute cost infrastructure evolving skill requirements can change mid year. So you’re going to promise the budget big expectation and all of a sudden you get landmined with a cost Increase that you were an expectation or a supply or a talent issue. How do you plan for that? Jason?
[00:33:58] Joe Puglisi: But just touching on Joe’s point about the speed bumps, AI will only go as fast as your weakest part of a process.
And I think a lot of companies are still struggling with understanding that.
And when you actually sit down and put all this stuff on paper and you analyze it, you’re like, well, we can go do this, this. Oh, that’s a big roadblock.
And I think a lot of companies are starting to find that out as they evaluate I I with a process.
But from a tight supply perspective, you know, we’ve grown a lot. You know, we’re a utility construction company and we’ve grown a lot. You know, before this year was by, you know, factor of seven over 10 years and then year over year we’re going to grow 50%.
So our resources are very constrained, especially from a people perspective, you know, and that’s why we’re really trying to figure out doing these AI lunch and learns and then creating an agent builder community, you know, because money’s getting tight, expertise is getting tight. And you know, I actually I’m onboarding an AI business analyst November 23, went through a pretty rigorous process to find her. But she’s, you know, she’s relatively young and the most exciting thing for me is I’m going to learn a lot from her.
But you know, we’re really looking to see how can we leverage because I can’t go out and hire 10fd, you know, for deploy engineers and unleash them on the company because I get fired for breaking my budget.
But when I presented the business case for an AI business analyst, really looking for that person that’s solid BA skills with solid AI agent creation skills and then they demonstrated what they could do, that kind of opened eyes to my CEO and he was like, go get them.
So when we look at the constraints that we have just with people and budget, I think what she’s going to be able to do is go after those business cases that we’ve already identified that, you know, myself and a couple other people just don’t have the time to do and really throw these things, things at her, you know, but like was talked about, you know, the guardrails, the Runway, the harnesses in place to be able to make sure we’re doing the right things because just because you can do it, should you.
And then, you know, when we do do it, taking a very deliberate approach with the documentation and the process, you know, but we’re you know, I still have a pretty strict lean budget fighting for every people that I can get, you know, but I, I don’t see that loosening up just because with everything going on and you know, and you see a lot of the tech companies, you know, with the layoffs and everything, you know, you could argue about the reasons why, but every company is really looking to tightening their belt and some of them are actually saying, well we’re going to save 10 to 15% because of AI.
And when you ask well how the answer is typically because of AI, you know, because I mentor other CIOs and they’re getting struck with this pretty hard, especially with type budgets and type supply and then plus with the chips and the cost of the token with the frontier models, a lot of the current models you have now, if the frontier models didn’t keep iterating every couple of months, every couple of weeks, what we have right now, how much could we actually get done without worrying about frontier models?
[00:37:23] Jason McKay: And
[00:37:23] Joe Puglisi: when you look at token spend based on fit for purpose, fit for use, when you’re creating a
[00:37:27] Jason McKay: model, you
[00:37:28] Joe Puglisi: don’t need a frontier model all the time. So you don’t need to go blow out
[00:37:31] Jason McKay: your
[00:37:31] Joe Puglisi: tokens for context and reasoning and heavy duty type of stuff like that. Yeah. But everything else, I haven’t been really convinced yet that you need to be on the frontier models and the models we have now, could we actually use those for the next year and still gain a lot of efficiencies and a lot of automating processes without having to go to a frontier model? I’d be curious to see what other people think about that.
[00:37:56] Isaac Sacolick: Asen, what have you learned from your AI business analyst that you didn’t know? Like what did a native person come to the organization with that were oh we should do this, would any, anything come out from that? Or still.
Or she’s still learning.
[00:38:13] Jason McKay: She’s going to start November 23rd. But just the
[00:38:15] Isaac Sacolick: conversations,
[00:38:16] Jason McKay: but just the conversations we’ve had because you know, I went through a lot of people and if you’re looking for someone like this, there, there are a lot of very talented young people out there.
But just you know, explaining to her where we are with things and she’s like well did you think about this, this and this? And I’m like well no. And then she’s like well what about this, this and this? And really, you know, she’s looking horizontally and she’s thinking horizontally across the organization and about the potential, what we can do. And I’M like, okay, you know, great ideas, want to, you know, start off the game a little slow and build momentum. But the vision of what she has, of what could be done with like AI and tools like it, she’s not constrained, she’s, she’s not, you know, she’s very
[00:39:00] Joe Puglisi: curious and very open minded about what we could do, which is for me, I think I told her, I go, I am going to learn so much from you and I’m probably gonna, you know, broaden how I look at different things and how I attack different things because you know, she’s such a so AI process native, you know, and I’m caught up in all this day to day stuff of just keeping things running.
That’s what I’m truly excited about. I think it’s going to be awesome.
[00:39:29] Isaac Sacolick: Eric, I need you to help this question that Jason’s raising. And Joanne, I hope you’ll comment on this too. This idea that organizations will pursue open weight, open source models, deploy them themselves, balance the cost out, figure out the orchestration between frontier and open source models.
You know, there’s a lot of, you know, I would say overhead to get to the maturity where you can get that, but maybe I’m wrong about it.
Can a company like Jason’s, you know, in the, in the industrial space play in this space where they’re going to be able to control their supply and use these other models?
[00:40:13] Derrick Butts: I think absolutely part of it goes back to realistic expectation and having the know how. The fact that Jason is hiring an AI analyst to help him kind of see this and she’s got this expansive vision of things that can be happening and things that can be done, that’s where it starts. You’re limited by your own growth or limitations of what you can think, but if you have somebody else helping to pull you along, it helps you create that, that versatility, that flexibility, that, that vigor that you need to actually push forward. When I talk to companies today about hiring personnel, some of the questions come back and say, well, why don’t you just do it? Write an agent to do that for you. You got to have the ability to know how to do that. And that’s what we’re still finding. The complexity or the, the issues where people don’t have the know how or the comfort yet and work with AI agents. So that’s still a hurdle they have to overcome. And when you look at this from, you know, the point of view of what can I do better, faster, smarter, leaner, it starts with educate yourself, become AI literate, Understand what the capabilities and tools are. We spent a lot of time, as I mentioned last year, this past 2026, understanding what these tools could possibly do. Now we need to understand what do they really can do for us. So when I look at this from a point of view of, you know, talking about tightening the supply, what does that mean? I would look at building AI resilience strategy for optionability.
We know which workloads work best and where we have the premium compute power in our services. Now we have to look at what needs to be done. What can the smaller models do? Where do I have interdependencies, what skills and critical things do I need to help keep this environment running? That’s what Jason is doing with the hiring person. I also need to look at from a cyber resilient perspective. If I have a scenario, what happens if the compute becomes too expensive? What happens now if the provider becomes constrained? I need to look at the criticality of what happens when these services are no longer available. As we experienced this with the Microsoft outage yesterday, how can we shift, how can we degrade gracefully but continue the business to run in a different way? All these questions come up and it goes back to flexibility and looking at how you change your mindset to think beyond what you’re in front of you. I need to increase the relevance of flexibility because now the cost of data centers, the major growth of engines for the compute, what that will cost, the lead time and the supply chain of all these different things now become a concern.
How will I plan for them? It starts with resilience. And resilience is going to be an intentional mindset where now these, these leaders need to figure out what can they do differently to, to plan for these hard times, plan for the interoperability, plan for these things that may not be available and how AI can help them get through it.
[00:42:43] Isaac Sacolick: Thanks, Derek. I like you need to the know how is a good starting point. It just reminds me going back to the days where people were looking at part public cloud, sitting on top of their data centers and saying, well, how do I get elasticity and the ease of onboarding and offboarding workloads
[00:43:05] Derrick Butts: when
[00:43:05] Isaac Sacolick: I’m doing it in my cloud and back then when I do it in my data center. And it took a generation of hybrid computing infrastructure, you know, to be able to enable that across enterprises and midsize computing.
Joanne, set some expectations here. I mean is the, you know, if I’m starting to get into, you know, the hundred to thousand AI agents, you know, from the gigabytes the terabytes.
Should I be focusing on use cases, resiliency and operations, or should I be looking at frontier mod, going from frontier models and giving me myself some model flexibility because of current supply costs and future supply costs? I’m just set my expectations here because I think that’s hard. I actually think going to open weight models would not be my first thing that I’d be working on right now.
[00:44:01] Joanne Friedman: Well, I think to your point, a couple of things. The first is, and, and this may be an outrageous thing to say, own your AI, because if you don’t, you’re immediately creating risk.
The open weight models are very effective, but in terms of reducing token cost and so forth and hallucination and drift. However, what’s wrong with creating your own ontological stack and your own taxonomy and creating a specialized language model based off of an open weight model not only to reduce your cost, but improve your efficacy? Because in doing that, what you’re taking away from a, from the overhead of AI is the, the natural language processing, right? I mean, you’ve heard me say this before. The word yield in manufacturing means how many products did I produce? It also means go through a roundabout slowly and look for other traffic.
AI does not know how to differentiate between those two.
So when we speak of the open weight models, you can take an open weight model that will cost you very little, mostly nothing, and start using that as a foundation on which to build a specialized language model that is for your business. In the lingua franca of your business, which may also be partly industrial, meaning whatever industry you’re in, people refer to the same thing. You know, there’s a vernacular that people talk, that is one of the ways to reduce the overhead. But the more important reason for doing that is because if you do it, you’re doing it based on your own systems. Less risk, lower cost, faster roi. Three good reasons to do it. And the other reason is because you will find that it’s a way for you to bring in tribal knowledge. Now, pardon me for putting it that way, I should say knowledge management or institutional knowledge of your number one asset, your employees.
And so I wouldn’t be surprised that if in six months or a year from now, if Jason returns, he’s not making a comment about, well, we’ve started to not only give people the information that they need so that they can get their job done, like on the spot with safety regulation, but we’re also asking them, what do you know from your years of experience? And bringing that into Dudley so he can do Right. Over and over and over again.
[00:46:49] Isaac Sacolick: So Joanne, I want to bridge this into this question about hard decisions, you know, because again, and honestly I haven’t done a lot of research on these open weight models and what it takes to go from the idea of using them to actually deploying them either yourself or with a third party, making them resilient and balancing load balancing between frontier and open weight models or your approach, which you’re saying basically I can build a proprietary model focused on my data and you know, let’s just call that door number three. I’ve got a finite number of resources, I’ve got, you know, it that’s got to change its operating model to support this better. I’ve got my employees that will, if you shut the door on their ability to use agents or to build agents, they may go rogue or they may go shadow AI.
I need to be able to keep up with their demand. And sitting in the center here is this sort of, you know, supply costs. So I’ve got some hard decisions. Am I going to go and get more people in my organizations to use AI and start transforming their business processes? Am I going to focus on resiliency? Am I going to focus on infrastructure to give me some, you know, potentially some cost versatility or benefits around this?
I think the question for you, Joanne, is like, under what circumstances does it make sense to focus on the supply side versus the growth and expansion slide?
[00:48:32] Joanne Friedman: I would say given geopolitics, given economies, given economies of scale when it comes to supply of internal resources and even materials, one should probably concentrate on what is the business problem we are actually trying to solve. First it’s about process redesign and companies find tremendous efficiencies by identifying where their operational gaps are, what I call execution layer problems, and figuring out which of which of the supply side issues they want to tackle. But taking the holistic view to it. You know, to Jason’s point about the business analyst, she may be very broad minded and looking across the board simply to reduce the number of agents that are going to be required and looking at a firm foundation that’s broadly applicable. I don’t think it’s a question of open model or frontier model. I think it’s more about what do I really need to generate the resiliency, the return on investment, the return on data that I actually want. Because in the supply, let’s not forget that we have petabytes, zettabytes of data stored in clouds that is basically digital dust. Why aren’t we leveraging some of that?
And that’s really the question that I ask every CIO that I meet. How much are you spending on data that you’re not actually using to create business value? I mean, to Derek’s point, that’s the core of resiliency. So if you took that data from where it’s sitting, doing pretty much nothing other than a historical record, and use that to help you fashion where your small language model can help you or your specialized language model and use the open source, that’s how you start using your supply chain, if you will, for AI production more efficiently and more easily. And I just want to make one little quick point after that. When it comes to actual supply chain and goods and materials moving from a supplier to a manufacturer or manufacturer downstream to a customer, that supply issue cannot be addressed solely by historical data. And this is where institutional knowledge, decision making, trade off management, which is going to be a huge issue in 2027, start to come into play. Those things have to be built into agents.
Don’t expect the agent to carry a task and then you’re going to build another agent to do another part of the task and so forth. And so, Solomon, keep the number limited to solving the whole problem. And then you can take that notion of here’s my solution with 10 agents or 50 agents, whatever, that are running as a swarm, that’s how you start to duplicate the modeling of those agents into other areas. And then, then you start connecting the dots. As I said at the outset, humans think in connections.
Agents have to be trained to do that. Even, even language models have to be trained to do that. If you change topics with an agent, with a, a chat or anthropic or anything else, it goes off into different directions and you end up with a lot of fluff.
Be very focused, be very careful about how you think about these things. And literally a piece of paper and a pen are your best friends.
Because you start thinking system wise, you don’t think task wise.
[00:52:27] Isaac Sacolick: Thank you, Joanne. Derek, hold off a second. I want to bring Joe in here because we’re getting into what I think you wanted out of this conversation. Around our, around hard choices. Right? I’m hearing on one hand, you know, that our CEOs want more.
I’m hearing our best practice again, which is go incrementally slower, focus on the areas of value. And in particular, because we don’t know the, you know, what cost or what limited resource is going to slow us down either during construction or later once it’s in production. I better forecast that and plan on the areas that if there’s a 30% increase in cost or if I lose the wrong person or if I have to cut costs somewhere, I could still operate what I’ve built. So help me deal with this paradox, Joe, where everybody wants to do more, faster, and I have a whole bunch of constraints that require me to think through what to focus on. Go ahead, Joe.
[00:53:38] Joe Puglisi: I think a lot of points have already been made.
Isaac, you know, John started earlier and, and even Rick, early in the, in the comments was saying you, you have to fund, you know, three or four bets, not 30.
Choose your battles, identify where you have the highest probability of success or the greatest return.
That, that’s kind of point one.
We all know that you, you have to estimate the cost and you have to be willing to pull the plug when things go.
So we’ve talked before about, you know, the proverbial kill switch. When is the plan going so far off the rails that we need to recognize this one’s not going to happen, you know, and pull the plug. So the kill switch is important.
Joanne and others have always said, and I will say it again, you got to take care of the boring stuff first. You can’t go for the glamour right away if you don’t have the right data, the, the retrieval mechanisms, if you, if you don’t have the right algorithms, whether that’s a frontier model or a small language model, you know, it’s got to be tested and built and you got to be prepared to take care of that boring stuff, get it, get it nailed down and, and then you can get to, to the really high value stuff.
We’ve talked about resiliency. This is Derek’s favorite topic. You know, you got to design for a downgrade.
As you pointed out, costs can escalate, whether it’s the cloud storage or the frontier models. The token, tokenomics. Right, all that. Let’s figure out what do we do when this starts to get too expensive. Don’t lock into one technology, be able to be flexible and move around if possible.
And then finally, Jason raised this, this question. I think it’s brilliant that he brought in some young talent who can do the business analysis and who can help facilitate the leverage over that new technology that all those old timers don’t have as much experience with. Right.
We’ve got the gut, but they’ve got the tech. So train the translators, bring, bring in the right business analysts.
And the last piece of advice is tell the board the bad news.
Don’t hide the failures. Learn from them and share them and say hey, you know, this didn’t work and here’s why. We just didn’t have the right data or that, you know, we didn’t have the right combination of steps. And, and there’s a fundamental flaw in the business. You got to bring that to the board’s attention or the C suite, whoever’s appropriate.
Don’t, don’t hide the problems. Expose them and fix them.
[00:56:12] Isaac Sacolick: That’s interesting comment for all the news this week going on around bad news. Derek, quickly, because I want to bring Jason back to close out the session. Go ahead, Derek.
[00:56:23] Derrick Butts: Sure. Thanks. When you look at the hard decisions, the things that come to mind is the separation. They have to be laser focused. As Jason and Jaron both articulated laser focus on those things. When it comes to what are the business critical things we need to focus on versus those things that are nice to have protecting those AI capabilities tied to revenue generation. Your customers, your critical operations, your AI and cyber risk. First, these all need to be architected as capabilities to overhaul all the dependencies of one model. You can’t just have one model in to be your primary infrastructure. There has to be conglomerate duration of the sbfu. Also look at identifying AI resilience strategies around the dependencies before they become the disruptors. This requires leadership to look at the long term visions and hard things. They need to look at what’s going to help me continue my focus on revenue generation, continue business operations without having to worry about AI disrupting these kind of things. And lastly looking at particular things, my AI infrastructure continue this significance and the impact, the cost is going to be associated with this. How are all these things going to be relevant? What are these things I need to focus on now? They’re going to get me to where I need to be two to three years down the road. That requires insight, requires laser focus and it requires intentional conversations need to happen now.
[00:57:33] Isaac Sacolick: Thank you, Derek. Let’s bring back Jason. Thank you for being our special guest. Jason, I’m sort of wondering, you know this, you know, from where you sit going into 2027, are you going to go after more use cases? Are you going to do more of the boring stuff?
Are you going to be focusing more on people or costs dominating things where they have to be a number one concern above these other areas.
Where are your hard decisions leaning toward?
[00:58:08] Jason McKay: I’ll
[00:58:08] Joe Puglisi: just answer yes and that’s all.
[00:58:10] Jason McKay: No.
[00:58:13] Isaac Sacolick: Welcome to the world of being a cio, right?
Yes,
[00:58:17] Jason McKay: yes. Have a good day.
You know, so a couple things here, you know, like Joe kind of alluded to you Know, highlight the failures, you know, fail fast, fail quick fail, cheap pivot.
You know, I had a mentor Tell me 25, 30 years ago, and I still use it today. Seize on the momentum of chaos.
There’s a lot of chaos out there. And you know, sometimes our job is to bridge that chaos and figure out how to contain it and then, you know, buy box it in and pull the box in closer, you know, but we’re, we’re going to focus on everything. You know, I, I have to be budget minded, but, you know, how do we innovate with the constraints? You know, what can we do? What don’t we do?
Really identifying the things just because we can, should we, you know, leaning on vendors, you know, making our platforms more AI native and AI focused and if they’re not doing it, someone else will.
And, you know, we’re going through that evaluation process right now. And then even, you know, we, we used to talk about technology roadmaps, three to five years, that is forever right now. And, you know, what does it look like in six months, one year, 18 months? And you know, trying to figure out what, what are we looking like in 18 months to two years from now?
And you know, based on some of the platforms that we use now and based on what we’ve been able to do, do we need those platforms in the future? And is it something that we can build on? As long as there’s not a regulatory requirement, we have to be very cognizant of or, you know, some other risk tolerance we have to be, you know, cognizant of. So, you know, we’re really, I think like everyone else, there’s so much ambiguity in the, the windshield used to be so far out from your face and now it’s like up close.
And, you know, how are you dealing with that and mitigating that? You know, and I think, you know, this, this type of discussion is great for everybody because one, we’re all in the same boat together, but two, you know, you get to find out a bunch of different perspectives and ideas because I, I’ve taken a lot of notes today, you know, about how to approach things. So, you know, we’re, I’m just trying to figure it out like everybody else.
[01:00:25] Isaac Sacolick: Thank you, Jason, for being here. Joanne, we got 15 seconds. Go ahead,
[01:00:30] Joanne Friedman: Jason. Start looking at AI as counting it in dog years.
[01:00:35] Isaac Sacolick: Oh, my gosh, way to simplify that.
I’m just going to leave this for everybody here today.
We’re giving you a lot of different advice in a lot of different areas, but you need to put this into your context and your own organization’s context. I know some organizations are still in learning mode and learning has a track to it.
I think other organizations are either in early experimentation or rapid experimentation mode. And I think this notion that Joe introduced, finding, you know, going from 30 to 40 experiments to a few big bets is where you need to be.
And I think that’s really important if you are on that journey from, you know, several hundred agents in production to thousands. You know, now you’re going to start looking at big areas of how you’re executing and that’s going to get into cost and supply and architecture. You know, something that Jason and, and Joe and I heard earlier this week, getting to 120,000 agents. What Joanne recommended own your AI. I think what we heard was own your AI harness and you start getting into okay, now I need to build the engine of AI for our organization. My sense is many of our listeners aren’t there yet and you’re somewhere in that learning scale rollout and doing the things that really matter. Set the expectations around that with your board going into next year because we know it’s not going to be a smooth ride. We’re going to have a bust and we’re probably going to have some speed bumps along the way. Thank you folks for joining this week’s coffee with digital trailblazers. We’ll be next week back next week for a discussion on AI for social good. Our discussion on the 27th will be off for Thanksgiving. We’ll be back on the 4th for our discussion around business, digital and AI strategies.
Back on the 11th for for our conversation around AI governance. And then we’ll be celebrating our 200th episode on December 18th. So just a few more episodes to close out the year. Go to drive.starcio.com Coffee to see the schedule to get access to the research slide to get this dashboard to listen to episodes that you missed and then click on add to your calendar so you never miss an episode that you can attend.
Folks, you heard me talk about my book coming out. I am starting to collect names for folks who want early access to different areas. If you want to be part of the early access team, go to drive.starcio.com AI reinvention and there’s a skeleton of a landing page there. You could sign up there for news about the launch of my book. I hope to have it ready for Pre sales in Q1 and launch in Q2. But I will be sharing some insights around the book for anybody who signs up there. And you’ll also get access to the Driving digital newsletter. Folks. Everybody, have a great weekend. Thanks for being here. We’ll see you here next week.
Leave a Reply