View Recording: Halftime 2026: Where AI Went, and Where It’s Headed

Brian Haydin Well, hello, everybody. Welcome to what I think we’ll just… 0:0:13.342 –> 0:0:32.702 Brian Haydin And I’ve got a lot of content. So over the next maybe hour-ish, maybe a little less, I want to take a mid-year look at where Enterprise AI actually went. And we’ll talk about capability, cost, governance, adoption, and where I think it’s going to be headed for the second-half of the year. 0:0:33.102 –> 0:0:53.742 Brian Haydin And I want to give you a point of view, not a news recap. And so this is going to be, yeah, I’m not going to, I’m not going to like read out the news recap, but I want to like start by doing something just a little bit uncomfortable. But first, before we get to that, just a quick word, you know, on who’s behind the session. 0:0:54.142 –> 0:1:13.702 Brian Haydin Concurrency is a Milwaukee-based systems integrator, and we spend a lot of our time helping organizations do modern work with Microsoft technologies. I’m not going to spend a lot, I don’t want to spend a lot of time, this isn’t a sales pitch. I’d rather spend our time talking about what the first half of the year actually taught us. And that starts with where this point of view actually came from. 0:1:14.542 –> 0:1:34.302 Brian Haydin And what it taught us starts where this comes from. So a little context about where we started from. I’m Brian Haydin, I’m a solution architect. I spend a lot of time doing webinars like this. I spend my weeks working with manufacturing, financial services, energy, industrial organizations, 0:1:34.622 –> 0:1:55.342 Brian Haydin mostly around AI governance, agent architecture. And increasingly, I’ve been talking a lot about AI cost. If you want to follow me, there’s my QR code for my LinkedIn profile. Grab it now and connect with me on LinkedIn. Love to chit chat and hear a little bit more about what you’re doing. But everything that I’m going to talk about comes from two places. 0:1:55.742 –> 0:2:15.582 Brian Haydin They’re the customer conversations that I’m having and my own published record, things that I’ve been talking about. And that’s where I want to start first. Earlier this year, the CEO of one of the largest labs in the world sat down for an interview and he admitted that he had been, using his own words, pretty wrong. He was wrong about how… 0:2:15.702 –> 0:2:22.382 Brian Haydin quickly AI would displace entry-level white collar work. And he said he was delighted to be wrong. 0:2:23.382 –> 0:2:48.102 Brian Haydin Well, he wasn’t alone. Another Frontier Lab CEO who had predicted that half of white collar jobs were at risk later reframed his idea this way. Automate 90% of the work and people can focus the remaining 10% of their work at 10 times the scale. Meanwhile, independent labor data is showing no meaningful unemployment shift happening right now among workers in these AI exposed roles. 0:2:48.182 –> 0:3:8.302 Brian Haydin roles. Now, here’s what I want you to notice, because I think it’s going to be a pattern that we reflect on for most of the next 50 minutes or so. They were right about the technology. The models became capable of doing what they said they would do. What they got wrong was the system around the technology, how organizations would be able to absorb it. 0:3:8.382 –> 0:3:18.302 Brian Haydin how they could govern it and redesign the work around it. They were right on the technology, but they were wrong on the system. So I want you to hold on to that thought for a second. 0:3:21.182 –> 0:3:41.662 Brian Haydin In that spirit, six months ago, on this same channel, doing another webinar, I made six predictions about 2026. The talk I called from base camp to a govern summit. And I said we weren’t hiking the mountain anymore, we were climbing and we needed to start using ropes. Well, here’s a little bit about what I’ve learned since. 0:3:42.142 –> 0:4:1.502 Brian Haydin Maps, they get drawn before the season starts. So six months in where we’re at today, you’re not going to argue with the maps. You need to actually reflect on reality. You go and find out where the water actually is. So we’re going to put, you know, the mountain away and I’m going to get to all six of these predictions. I want to grade them. 0:4:1.742 –> 0:4:7.662 Brian Haydin in public with you as my witness, and I want to start with the two that I got right. 0:4:10.142 –> 0:4:30.62 Brian Haydin I think the first one is that agents moving from chat to systems of action. That’s one that landed pretty well. But I would even say it landed a little bit bigger than I expected. You know, as an example of that, a Midwest manufacturing company that I’m working with is running a pilot right now that reads inbound older emails. It extracts fields. 0:4:30.142 –> 0:4:48.542 Brian Haydin validates some information, flags exceptions, and then creates orders. I would say that that’s not a chat bot. What I did not predict was how deliberately customers would limit some of the autonomy that we were talking about. The agent could do the work, but the customer kept the final approval gate. 0:4:49.102 –> 0:5:8.342 Brian Haydin And I’ve been seeing that in nearly every agent engagement that we’ve been working on. My second prediction was that shadow AI would become the governance crisis. I think that one landed too, but it was a little bit uncanny. In January, I put up a mitigation list. It was the agent registry, a policy framework, 0:5:8.462 –> 0:5:29.102 Brian Haydin sandboxes, approved connectors, and escalation paths. And as an example, in May, a manufacturing client sent me an unprompted request to review their agent registry. They had a lifecycle model and risk tiers and production readiness checklists. My slide came back as their shopping list, really, to be honest with you. 0:5:29.422 –> 0:5:50.502 Brian Haydin I’d love to claim credit for that, but that’s really what has been happening in the market. And it arrived pretty much at the same conclusion. I’ve been talking a lot about these types of conversations, like in the agent ops, you know, throughout the year. I also had a prediction around local and edge momentum. I think it was directionally right at this one, but I was a little bit early. 0:5:51.182 –> 0:6:9.582 Brian Haydin Some of the supporting technology in the smaller scale scenarios have arrived. Enterprise scale at local inference, though, I think it’s mostly just conversations and it’s not actually being purchased. So I’m grading that one. I kind of hooked it, but I didn’t land the fish. And like the sea. 0:6:9.662 –> 0:6:31.982 Brian Haydin CEO said, I’m delighted to be wrong, but for now. On the multi-modal going mainstream, you know, that capability absolutely did ship. We can do images, audio, video, and documents are not part of the normal AI experience, but my customers, they’re not leading conversation, they’re not leading in with conversations around it. 0:6:32.542 –> 0:6:52.342 Brian Haydin And I think that’s kind of the lesson and the take away for me is that a capability prediction can be completely right and still not be a headline story. Another prediction was hardware changing the cost curve. I got the economics right, but I did not get this mechanism, you know, I got the mechanism wrong. I thought that the chips would drive change. 0:6:52.422 –> 0:7:13.102 Brian Haydin But what wound up happening is that the budget conversation changed because of consumption billing, credits, and usage-based pricing. And the annoying part of that slide, you know, it was slide 18 in my January deck said, track cost per outcome, not cost per call. So I guess I was arguing the right thing, but just had the wrong mechanism. 0:7:13.182 –> 0:7:32.862 Brian Haydin And then finally, there’s this authenticity premium. And I mentioned AI slop. And that is clearly real in the broader cultural sense right now. But in the enterprise pipeline, people aren’t really talking about it. It’s invisible. Not one conversation that I’ve had with a customer has started with that. So 0:7:32.942 –> 0:7:42.702 Brian Haydin That’s my scorecard for my predictions earlier this year. Two clear hits, 4 honest grades, and I think the misses are probably the things that I learned the most from. 0:7:44.622 –> 0:8:4.942 Brian Haydin So everything that I got right was about what AI could do. Every place I was early or wrong or surprised about was what the system around it, who approves the action, what it cost per outcome, and who owns the thing after the demo is over. It’s the same pattern that we saw with those two CEOs. 0:8:5.902 –> 0:8:19.822 Brian Haydin And this is the part that really got my attention. It is the same pattern that I’m seeing a lot with the customers that I’m talking to every week. Which brings me to the argument that I want to make for the next, you know, the next half of this conversation. 0:8:24.582 –> 0:8:44.622 Brian Haydin I would say, here’s my argument, and I want to be pretty careful about how I say it, because a sloppy version of this argument is everywhere right now, and it’s mostly wrong. The sloppy version says that the model race is over, and I’m going to say that it isn’t. In the last few weeks alone, we’ve seen multiple frontier model releases, even this week. 0:8:44.782 –> 0:9:4.862 Brian Haydin you know, some big announcements. Each are pushing capability, each are pushing speed, and they’re pushing cost in different directions. The capability is still accelerating. That’s clear. And if you built your strategy on the assumption that it’s plateaued, you’re going to be wrong again later this year. But the accurate version has an and in it. 0:9:5.342 –> 0:9:25.262 Brian Haydin The model race is still accelerating, and the enterprise contest has moved somewhere else. It has moved to who can turn that intelligence into governed, economical, and continuously improving systems of work. The model is becoming a component. That’s what I would say. And the operating model is becoming the differentiator. 0:9:25.982 –> 0:9:29.822 Brian Haydin And I can show you that the field is going to agree without quoting any of the opponents. 0:9:32.222 –> 0:9:51.422 Brian Haydin Take a look at these two rooms. They never really talk to each other. Room one is the AI engineering room. It’s in San Francisco on the conference stages, mostly in open source repositories. In about 18 months, that room renamed its own core skill three times. It started out with 0:9:51.462 –> 0:10:10.942 Brian Haydin prompt engineering. Then the conversation shifted into context engineering, what the model can see, what it can retrieve, and what it remembers. And by the middle of this year, the conference agendas has started to move again. And what we’re hearing about is loop and harness engineering. That’s the scaffolding around the model. 0:10:11.582 –> 0:10:33.702 Brian Haydin Now, in the second room, that’s your room, in January, my customers were asking me which use cases, where do we start? What’s actually realistic for us to do? But by Q2, those same people were saying agent registry, kill switches, credits, evals. And by this summer, the question that I’m hearing the most is who owns the agent after the pilot? 0:10:34.542 –> 0:10:52.502 Brian Haydin Now, nobody coordinated these conversations. Two completely different populations, different timelines, but they’re moving in the same direction, away from the model and into the systems that it’s built around. And that, and when the vocabulary starts to shift, that’s when the money is going to follow. 0:10:52.622 –> 0:11:3.822 Brian Haydin It always has. Now, is everyone convinced about this? I wouldn’t say so. And the best counterargument probably deserves, you know, a hearing for us to listen to. 0:11:7.822 –> 0:11:28.22 Brian Haydin Before I go any further, let me give a strong, the strongest argument against everything that I just said. And it goes like this. If the models were truly general, they wouldn’t need all of this scaffolding. The fact that we’re building up harnesses and context layers and control planes is evidence that the models simply aren’t 0:11:28.102 –> 0:11:41.502 Brian Haydin there yet. All this may be some sort of a temporary bridge, something that gets torn down when the next generation of models arrives. Maybe, you know, I would say, but that doesn’t change what you start that. 0:11:42.382 –> 0:12:3.342 Brian Haydin That maybe isn’t going to change what you do on Monday. Whether it’s a bridge or whether it turns into a permanent layer, you need to do it either way. Nobody gets to skip the scaffolding on the theory that it might be obsolete in just the next couple of years. So let’s get to what you actually came here for. I’ve got 4 dimensions and four verdicts. 0:12:5.982 –> 0:12:12.542 Brian Haydin These 4 dimensions are capability, cost, governance, and adoption. 0:12:14.622 –> 0:12:18.702 Brian Haydin The same for that the abstract promised you and let’s take them in order. 0:12:21.662 –> 0:12:39.742 Brian Haydin Where is, why am I off on my slide? Well, all right, let’s talk about order entry, the first, the capability first. I predicted that agents would move from chat systems to systems of action. And that’s exactly what’s happened. Order entry, document processing, proposal assembly, reporting, 0:12:40.62 –> 0:12:59.742 Brian Haydin Real work, not just the demos. But here’s the finding that I didn’t expect, and it has held up across most of the agent engagements that we’ve been doing this year. The moment a customer sees an agent that can act, the very next thing they do is they’re putting a gate on it. Not because they’re timid, but because they’re running a business. 0:12:59.902 –> 0:13:12.382 Brian Haydin Running a business has business rules and breaking those rules has consequences. So the honest read on the capability is not that the agents became autonomous, it’s that the autonomy became the dial and not a destination. 0:13:13.582 –> 0:13:24.62 Brian Haydin And the interesting question, like design question that I have in 2026 is where we’re going to actually put that dial, which raises the next question, who is actually doing what? 0:13:28.222 –> 0:13:47.542 Brian Haydin So here’s the shape that like I’ve started to think it’s taking. The agent owns the inner loop. It’s doing searching, it’s drafting, it’s extracting, it’s validating, and it’s got retries. That’s a tight cycle. That’s the repetitive middle of the work that’s happening. But you as a human, 0:13:47.662 –> 0:14:6.302 Brian Haydin You own the outer loop. We are setting the objectives. You define what good looks like. You’re handling the exceptions. You judge the outcomes, and you own the results. In January, I made the comment, human in the loop becomes human on the loop. And I think it was directionally right with that. 0:14:7.182 –> 0:14:26.342 Brian Haydin but I was about half a level 2 abstract. This is what it looks like with six months of like evidence behind it. And notice what sits on the boundary line, because everything expensive in your AI program lives there. Your controls live there. Your costs live there. Your accountability lives there. 0:14:27.102 –> 0:14:46.22 Brian Haydin And one more thing about the outer loop kind of surprised me. When we’re encoding into these systems isn’t just data, it’s lead time logic, customer specific agreements, the way your best analyst handles an exception, which means that your real moat isn’t simply what your AI knows, it’s the encoded judgment governing on 0:14:46.222 –> 0:14:55.742 Brian Haydin on how it uses what it knows. Now, I’m going to give you an example of what a customer did that drew in line themselves, and they did it before we even suggested it. 0:14:56.862 –> 0:15:18.302 Brian Haydin So this is a commercial real estate firm. They have an RFP to proposal process that took 7 to 10 days. It interprets the requirements, finds the right material, it drafts the response and assembles it, and then routes everything for approval. Their target is to get to a same day review process, a review ready output. 0:15:18.862 –> 0:15:37.342 Brian Haydin We mapped out what the agent would own versus what stayed like within the lanes of the human, and they wound up actually drawn the line themselves. They wanted an agent that interprets, retrieves, drafts, and assembles. But the human team is going to keep the pricing. They’re going to define the commercial terms. 0:15:37.622 –> 0:15:57.422 Brian Haydin and then an executive is going to sign off on it. Nobody sold them on that as a prospect. That was their first instinct. In their words, roughly, the winning agents are going to do the middle work, but they’re not going to be responsible for the final judgment. And honestly, that’s the pattern that I’m seeing more broadly right now. The gate into production is no longer just 0:15:57.582 –> 0:16:18.782 Brian Haydin you know, on a good demo. In a parallel engagement, something a little bit different, a manufacturer that I’m working with defined acceptance this way. Process 25 to 30 documents end to end in test mode and catch every error that they deliberately planted in there and match what a human would have produced. That was how they did it. That’s not a demo, that’s an eval that was written. 0:16:19.582 –> 0:16:38.702 Brian Haydin And a great demo, that’s a hypothesis, but when we start to incorporate evals into there, that becomes evidence. And that’s the second-half bet that I have on capability. It’s probably ahead of the charts right now, and it arrived with an invoice attached. And that’s what we’re going to talk about a little bit more. 0:16:39.982 –> 0:16:40.542 Brian Haydin So. 0:16:41.742 –> 0:17:3.662 Brian Haydin Dimension 2, that’s the cost. And this is the one where I was wrong in a more interesting way that I kind of alluded to. I predicted that the hardware would move the cost curve, that we’d get faster inference, that we’d get cheaper compute, and more viable on-premise options. Directionally, I think that’s still happened. We’re getting a lot of really cool announcements around the hardware. 0:17:4.62 –> 0:17:25.582 Brian Haydin But that’s not what changed the budget conversation. What actually changed the budget conversation is that the meter started turning on. Consumption billing on agentic work, the idea of credits, whatever credits are worth, usage-based pricing and developer tooling. For many organizations, that was the first genuinely variable AI invoice that landed. 0:17:26.222 –> 0:17:47.142 Brian Haydin And that started happening just in the last few weeks. Well, here’s the counterintuitive part that catches the finance teams off guard. Per token prices have actually been falling, but the total spend can still climb because agentic workflows do not just make one single call. They make dozens of them. They have retrieval calls. They have reasoning. 0:17:47.502 –> 0:18:1.502 Brian Haydin They make calls to different tools, they retry, and then even more so, they pass all the different works off to sub-agents. So at a unit price, things are getting cheaper, but there’s just a lot more of those units that we have to account for. 0:18:4.462 –> 0:18:25.902 Brian Haydin If you think about it, I have one sentence that’s really useful for the CFOs that are out there. Licenses are the cost that you approve, and tokens are the cost that you’re going to wind up discovering. And that’s from a cost perspective, that’s the whole shift in one sentence. A license shows up in a procurement cycle with a signature attached. 0:18:26.302 –> 0:18:46.182 Brian Haydin But consumption, it’s going to show up later on an invoice that you read after things have been done. What I’m seeing with organizations right now that are getting ahead of this, they’re asking for show back budgets, cost dashboards before they scale, not after the first month where the bill arrived. One of our own delivery playbooks has started to put 0:18:46.302 –> 0:19:6.302 Brian Haydin It puts it in a little bit more blunt terms than I would. Budget predictability is gone. One warning while we’re here, because this is a trap that I’ve watched a lot of teams walk into, is do not let usage become the target. If your adoption metric is just the tokens consumed, congratulations, you are going to get tokens. 0:19:6.462 –> 0:19:24.942 Brian Haydin Like, that’s what’s happened. I’ve read a bunch of stories about organizations that have like had token races and stuff like that. But that’s not really adoption. That’s just an expensive way for people to be busy. And, you know, what I predicted before was to measure cost per successful outcome. 0:19:25.342 –> 0:19:31.182 Brian Haydin not just the cost per call. And that turns a finance question into an architectural question. 0:19:33.382 –> 0:19:33.822 Brian Haydin So… 0:19:35.582 –> 0:19:53.982 Brian Haydin And this is where the cost, you know, six months ago, the questions that I heard were, which model was the smartest and which platform should we standardize on? The question today is a little bit different. It’s where does this workload belong? Because not every task needs a frontier model. 0:19:54.382 –> 0:20:13.262 Brian Haydin Some work, it needs to run as asynchronously. Some work should be batched overnight when nobody’s waiting for it. And some should not leave your environment at all. And so those decisions all come down to a combination of risk, latency, privacy, and cost. And I think that’s my second-half. 0:20:13.342 –> 0:20:34.502 Brian Haydin bet on this particular dimension. Workload placement and model selection are going to become the runtime policy, not like a one-time procurement decision. And I think I’m a little bit ahead here, you know, from what the evidence is saying. I hear the conversations constantly. I don’t see many organizations doing this one. 0:20:35.262 –> 0:20:56.662 Brian Haydin you know, formally. So that’s the bet, not an observation. What the cost that the meter is running, and it moved from procurement to operations, and it’s moving really fast in that direction. So the third dimension that I wanted to talk about was governance. And I want to start with the January slide that I’m 0:20:56.942 –> 0:21:15.702 Brian Haydin that actually I think was, that resonated the most with people that I talked to. I’m proud of it, but I think that like I’m a little bit unsatisfied with it as well. And it was the one that I talked about, the Trust Act. I had seven layers, identity and least privilege, data boundaries, 0:21:16.142 –> 0:21:36.462 Brian Haydin prompt injection defense, human gates, observability, continuous evals, and incident response. I still stand by every single one of those. And in fact, it became like a talk that I do probably every two or three weeks. And it’s a shopping list for a lot of the organizations that I work with. 0:21:36.502 –> 0:21:55.302 Brian Haydin with. But here’s what six months of customer conversation started to clarify for me. Everything that you see in that stack can be defined as a capability. And most of the organizations, they just implemented it as a document. They wrote the policy, they formed committees, and they published the principles. 0:21:55.902 –> 0:22:1.102 Brian Haydin And then the business unit shipped an agent that could send an e-mail, update a record, or create an order. 0:22:2.262 –> 0:22:21.742 Brian Haydin So it’s just a policy document if you’re not acting on it. And a policy document can’t interrupt the dangerous tool call. And that is the sentence that I would add to the January deck if I could go back to it. So let me show you a little bit of what that gap looks like inside an organization that was doing almost everything the right way. 0:22:23.982 –> 0:22:42.142 Brian Haydin It’s a global manufacturer that we work with, and they had an AI digital workforce governance proposal and a production agent review checklist. So they did the things that we talked about. And it was pretty good work. And most, like I would even say, like ahead of most of the organizations that I talked to. 0:22:42.902 –> 0:23:2.222 Brian Haydin And they asked us to take a look at it, not because the framework was wrong, but because they realized that it stopped short of their operational control. What it was missing is they had no registry showing that which agents existed. They had no life cycle for moving an agent from a prototype to production. And eventually when things… 0:23:2.902 –> 0:23:21.542 Brian Haydin when it became antiquated to retirement. And they weren’t measuring any telemetry as a baseline to show whether any of the things that they were doing were working or any of the rules were actually being followed. So my point is that a framework is not the control plane. A framework describes what should happen. 0:23:22.62 –> 0:23:41.422 Brian Haydin And the control plane makes it happen, you know, when nobody’s watching. And here’s my favorite data point from like the first half this year. Back in March, I had a specialty chemicals customer ask me in a meeting about kill switches. Nobody asks about kill switches for software that they 0:23:41.462 –> 0:24:2.862 Brian Haydin that they do not think can act. One more example, and this is a little bit less fun, is that another manufacturer, 2 agentic projects that they were working on, quietly died earlier this year. Not because the technology failed. They died because nobody had established the value beforehand, and nobody was in charge of owning them afterwards. 0:24:3.422 –> 0:24:14.782 Brian Haydin So governance is not how you keep agents safe, it’s how you actually keep them alive. And this is becoming urgent in the second-half because the number of agents that we’re building is about to scale dramatically. 0:24:16.622 –> 0:24:30.182 Brian Haydin In January, I called Shadow AI the governance crisis of 2026. And I want to update that prediction a little bit because phase one is mostly over and phase two is starting to look a lot different. 0:24:32.942 –> 0:24:51.582 Brian Haydin Phase one was, who is using the AI tools that we didn’t approve? Most organizations have at least some answer to that now. But phase two is a little bit harder. The barrier to creating agent has, the barrier to creating these agents has collapsed. Pro code, low code, it doesn’t really matter. 0:24:51.662 –> 0:25:11.182 Brian Haydin They’re both becoming pretty easily and effective. So the question is no longer whether people can build agents. It’s which agents actually exists? Who owns each of the agents that were built? What can they reach? What tools can they call? What do they cost? And when does anyone decide that it’s ready, that we need to retire it? 0:25:11.742 –> 0:25:30.542 Brian Haydin And that’s my second-half bet on the governance aspect. Agent sprawl is becoming the next shadow IT. It’s the access database of our time. And the organizations that succeed are not going to be the ones that simply lock everything down and stop it. It’s the same principle that I argued for in January. And I think it’s still believed. 0:25:30.702 –> 0:25:49.182 Brian Haydin I think it still is true today. It’s visibility without killing innovation. Enable distributed creation, but centralize the identity, the visibility, and the evidence. So the Halftime call for governance is the work has started, but the controls are still developing more slowly than the capability. 0:25:49.742 –> 0:26:8.622 Brian Haydin Governance isn’t absent, but it’s one step behind. And I think an evidence of this is Agent 365 just not quite having the clamor for adoption that people think that it has. And that brings me to the last one, the last dimension that I wanted to talk about, and probably the one that matters the most. 0:26:9.782 –> 0:26:29.982 Brian Haydin And that’s, it’s adoption. And this is where the gap between activity and impact has been the widest this year. And here’s what most organizations are counting, you know, in terms of adoption. They’re looking at measures like seats deployed, active users, prompts submitted, agents published, 0:26:30.142 –> 0:26:39.742 Brian Haydin training completed. And I think we’ve all built some sort of report that looks at these numbers. And then there’s the other line, what actually changed. 0:26:41.342 –> 0:27:0.462 Brian Haydin So you can report every number on that first list, every, you know, every target, and you still don’t move a single business outcome. The tools went mainstream earlier this year. I think everybody is using some of these tools, you know, on a day-to-day basis. And some of the largest AI newsletters in the world are now 0:27:0.622 –> 0:27:21.142 Brian Haydin teaching agent workflows to millions and millions of people, and most of these people have never written code in their life. Access isn’t the constraint anymore. Usage, though, is not transformation either. So the unit of transformation isn’t like how many prompts, and it’s not how many agents, it’s how many workflows are we starting to change. 0:27:21.902 –> 0:27:29.902 Brian Haydin And the best evidence that I have for that is what changed, I think, in my own e-mail, in my Mailbox, things that I’m seeing people ask questions about. 0:27:30.862 –> 0:27:52.142 Brian Haydin It’s not a survey. This is what changed in my own calendar between January and June. In the first quarter of this year, the questions were, which use cases should we start with? Where do we begin? What is actually realistic? And I think that was, you know, a lot of valuable questions that led to a lot of valuable conversations. 0:27:52.542 –> 0:28:13.702 Brian Haydin But it’s shifting in Q2. Those same organizations, and honestly, some of the same people, they’re asking different questions to me. They’re saying, who’s governing this? Who funds it? And how do we prove that it’s actually working? So the question that I hear most now is who owns it after the pilot? One client that I’m working with captured the whole story. 0:28:14.382 –> 0:28:34.942 Brian Haydin with this. In January, they made a deliberate and I think a defensible decision to defer governance until they had proven value. That’s kind of smart sequencing at that time. And so I agreed with it. But by Q2, governance was the thing that they’re leading the conversations with. Nothing about the underlying technology suddenly forced the change. 0:28:35.302 –> 0:28:49.502 Brian Haydin What changed is that the work that they were seeing was actually starting to get real. And so they went from curiosity to accountability in just six months. And underneath all those Q2 questions is the one question that they’re all really asking about right now. 0:28:51.822 –> 0:29:2.302 Brian Haydin They stopped asking whether the agent can do the work. They started asking who’s accountable when the work crosses from recommendation into action. That’s it. 0:29:4.302 –> 0:29:25.182 Brian Haydin That is the Halftime insight underneath, you know, all the noise. And notice what kind of question it is. It’s not a capability question. It’s a system question. Things are right on the technology, but they’re wrong on the system. And this is the third time that we’ve landed there today. First it was the CEOs, then it was my commentary. 0:29:25.222 –> 0:29:45.102 Brian Haydin And now it’s the market making those same commentary. So the second-half bet on adoption is pretty straightforward. AI projects get funded as workflow redesign, not as just the simply the tooling. I think executives are going to stop counting pilots and start asking which one, you know, what one successful outcome costs. 0:29:46.222 –> 0:30:4.222 Brian Haydin So, like, you know, I mentioned this before about measuring a, you know, the cost of 1 successful outcome. And I think that’s where executives are going. And what’s going to wind up happening with that is that the science fair starts losing its funding. So adoption at Halftime, my call is that there’s been lots of activity, but there’s been 0:30:4.302 –> 0:30:12.302 Brian Haydin uneven impact. And so those are the four dimensions, my 4 verdicts. Here’s the scoreboard. 0:30:14.582 –> 0:30:35.262 Brian Haydin Capability, ahead of the charts, moving toward action and doing so definitely with the human gates. On the cost side, the meter is starting to run and the conversations move from procurement to operations. And that’s moved pretty quickly in just a short, like in one quarter. On the governance side, I just mentioned 0:30:35.422 –> 0:30:54.142 Brian Haydin that it’s trailing the capability. Policy is starting to become the runtime, but it’s doing so pretty slowly. And on the adoption part, there’s high activity, but uneven, you know, uneven impact. Usage is turning into the workflow redesign at the leading edge and almost nowhere else. So 0:30:54.382 –> 0:30:54.942 Brian Haydin Um… 0:30:55.822 –> 0:31:14.622 Brian Haydin The story those four rows tell, I would say that the story that these four rows are telling together is the same one that I kind of opened with. AI is winning the capability game, absolutely, 100%. Enterprises are still learning, though, how to play the system game. So with that in mind, what are you going to do with it? 0:31:14.702 –> 0:31:34.62 Brian Haydin And I promised a little bit of a verdict on what you could do. So here’s where it is. I want to double down on three things first. Pick 1 bounded workflow. Measure the baseline before you start to adjust it, before you build, you know, before you build the agent, and build eval harnesses around it. 0:31:34.862 –> 0:31:55.822 Brian Haydin One workflow that goes from seven days to the same day, if you can measure that, beats 50 pilots every single time. Second, have some sort of a runtime governance, identity, registry, approval gates, that kill switch I talk about. And name a human who owns each agent from the request all the way through retirement. 0:31:56.542 –> 0:32:13.422 Brian Haydin Not the policy, put the plumbing together. Third, workload economics. Cost per successful outcome, having a real budget, and real routing decisions. Do this before you start getting invoices that are going to start making strategy for you. 0:32:14.782 –> 0:32:34.862 Brian Haydin So I’ve got two things that I would say hold, like, you know, on for now. And notice that both are my own January promotions or predictions that I had that are getting demoted. That should tell you that I’m not doing this as a self-serving exercise. I’m not puffing my own chest. But the 4th one is this local and edge inference. 0:32:35.582 –> 0:32:54.142 Brian Haydin Honestly, I think that the direction is still valid, but the market is not quite where I expected it to be right now. I would say buy it where regulation and latency demands it today, or build it in a local scale, but don’t build a broader strategy around that yet. It’s not quite there. 0:32:54.502 –> 0:33:14.222 Brian Haydin And the other one is the multimodal one. Use it opportunistically inside of your workflows that you’re already redesigning, but don’t lead with it. And finally, there’s three of these things that I would probably just let go. Generic chatbots that are looking for a problem and counting pilots as the number that, you know, as if the number was the point. 0:33:14.622 –> 0:33:33.502 Brian Haydin I would let those go. Adoption measured in seats and logins, not relevant, not important. And governance that ends the day that PDF gets published. If you’re not going to follow through on it, you might as well just let it go. So that’s the verdict that I came up with for this webinar. 0:33:34.462 –> 0:33:54.102 Brian Haydin And I think that leads to five questions that you can take to your next leadership meeting and ask these ones out loud. The first one would be, which business workflow are we materially changing? And what is that measured baseline? If you can’t state the baseline, then you’re going to have a hard time later proving that 0:33:54.222 –> 0:34:13.102 Brian Haydin the change worked. Second, what is our what is what can our AI actually do? Not just say, but actually do, and who approved the action service? Third, what context does it receive and who owns the context? Who’s in charge of generating it? Fourth, 0:34:13.622 –> 0:34:32.702 Brian Haydin What does one successful outcome cost? And would we notice if that doubled down next month? Fifth, who owns this agent after the demo is done all the way through the retirement? If those five questions are easy for you, then I think you’re further along than most of the companies that I’ve been working with and definitely further along than most of the market. 0:34:33.262 –> 0:34:45.742 Brian Haydin And I think I’d like to hear from you so I can learn a few things. But if one of those made you uncomfortable, that’s not a gap. That is your second-half priority. So let’s go back to a little bit where we started. 0:34:47.102 –> 0:35:5.702 Brian Haydin In January, I told you that the summit was there and the path was clearing. Six months of this year gave me quite a sharper version of that. The first half of 26 asked how capable the models were going to become. And the second-half asked how capable your organization can become around them. 0:35:6.702 –> 0:35:25.462 Brian Haydin So, to use my little fishing analogy, the captains that are filling the cooler in August, they’re not going to be the ones who just looked at the fishing charts in the spring. They’re going to be the ones that actually went out, checked the water, and, you know, adjusted accordingly. And I’ve got three ways that we can do that. 0:35:26.542 –> 0:35:46.62 Brian Haydin If you couldn’t answer one of those five questions before, you already know which door you’re in. Couldn’t answer question number one, which workflow? What was the baseline? That’s A readiness and a use case triage conversation. If you couldn’t answer two, three, or five, what can it do? Who owns the context? 0:35:46.302 –> 0:36:8.862 Brian Haydin who owns the agent? That’s a control plane conversation. And if you couldn’t answer 4, what does one outcome cost? That’s A purpose-built pilot with an eval harness on it that we can, that we will work on together. So I’ve got a QR code here. This is a, fill out the, there’s a couple little questions. What do you want to hear from us next? 0:36:9.822 –> 0:36:31.142 Brian Haydin And we can send you all five of these questions in the scoreboard, you know, in one page, unbranded. And for next steps, if you want to hear a little bit more from us or you want to give us a little bit of feedback on this presentation, I’m happy to have a conversation around your AI strategy and do a little bit of a 0:36:31.222 –> 0:36:54.702 Brian Haydin pressure test. We can evaluate your AI roadmap, help you plan for the second-half of 2026. A lot of companies are asking us to do an AI governance readiness assessment, identify gaps in security, what are some of the things that they can be doing next? And then I had a webinar a couple of weeks ago that really took, that really got a lot of attention and led to a lot of conversations. 0:36:55.62 –> 0:37:14.222 Brian Haydin And I mentioned it a little bit today in terms of some of the cost, but AI tokenomics and the consumption review, how do I plan? How do I justify some of the costs that are happening? And love to have a conversation with those as well. So again, Paige, I think dropped a link in the chat, or you can use this QR code. 0:37:14.862 –> 0:37:23.102 Brian Haydin fill out, you know, fill out the, you know, the, um, fill out the form and let us know if you want to have a conversation. 0:37:24.302 –> 0:37:24.502 Brian Haydin Kayla. 0:37:25.902 –> 0:37:31.102 Brian Haydin All right, well, thanks very much for joining us today, and enjoy the rest of your week.