A few days ago, part of our leadership team attended Ai4 2026, America's largest enterprise AI conference. Six talks, a stack of notes, and one conclusion that kept coming up session after session: the enterprise AI conversation is maturing. It's no longer about who has the newest model or the flashiest agent; it's about who can actually turn that into sustainable business results.
Here are the key takeaways, organized around the themes that resonated most.
One of the first talks, "Building the Enterprise AI Stack: Strategies for Tools, Platforms and ROI," opened with a simple but worth-repeating idea: if you already have a solution that delivers real value, you don't need to chase whatever just launched. The pressure to adopt "the most advanced" or "the wildest" new thing is real, but keeping a stable, reliable stack, even if it's not the most cutting-edge, is usually the right call.
One data point the speaker shared stopped the room cold: they asked the audience how many people felt genuinely comfortable with their current AI stack. Out of a packed room, only ten people raised their hands.
That's not an isolated data point. Recent industry surveys paint a similar picture: 97% of executives have already deployed AI agents in the past year, but only 29% report significant ROI. Gartner goes further, projecting that more than 40% of agentic AI projects will be canceled before 2027. The gap between "we implemented AI" and "AI is delivering measurable results at the corporate level" remains the number one pain point, and it lines up exactly with what that room showed.
Why this matters to us at Onetree: it's the exact problem we see most often with clients. The anxiety of "falling behind" drives teams to stack up tools without a clear strategy, and the result is a fragmented stack that costs more than it delivers. Helping clients audit what they already have, before adding anything new, matters just as much as the implementation itself.
The second talk centered on something we treat as a working philosophy: to get real economic value out of AI, you have to start by understanding the company's workflows, not by picking the trendiest technology. The right question isn't "which model should we use?" but "which process, if automated, actually moves the needle?" The technology gets chosen afterward, based on that impact.
This lines up with what the latest enterprise adoption research shows: the organizations getting results are the ones that connect AI directly to revenue or measurable efficiency, and treat adoption as an organizational redesign, not a software rollout. The ones that struggle tend to have an "AI strategy" that exists more to look good than to guide real decisions.
One of the most original talks of the event, and the one that most surprised our CTO, covered "agent dreams," a concept gaining traction across the industry. Anthropic shipped a version of this for Claude-based agents under the name "dreaming" earlier this year, though the broader approach is being explored by several players.
The idea, in plain terms: at the end of the day (or a work cycle), a process analyzes everything the agent did, meaning its logs, identifies what worked and what didn't, and decides what's worth remembering. That curated data gets saved to the agent's memory, so it's available for future use, without a human needing to manually retrain anything.
It's essentially the difference between an agent that repeats the same mistakes every week and one that actually accumulates experience. For any company moving from "testing agents" to "operating with agents," this kind of memory and continuous-learning mechanism is going to be a real differentiator.
Another talk, by Srini Venkatesan, EVP and Chief Technology Officer at PayPal, showed how one of fintech's biggest players is using AI agents to speed up software development without growing headcount. The numbers shared during the event (which line up with what PayPal has been publicly communicating) are striking: the company rolled out AI coding tools across thousands of developers, and high-adoption teams moved from one- to two-week release cycles to daily deployments. One specific example: a Java migration affecting 3,000 applications, originally estimated at a year of work, was completed in two months.
The underlying message wasn't "we used AI to shrink our teams," but "we used AI so the teams we already have can do a lot more." That distinction matters, both for the internal conversation about AI and for how we communicate it to our own teams.
Closely tied to point 1: another talk focused on the real obstacles to taking agents from pilot to production, namely accuracy, cost, and ROI. This is the stage where most projects stall: an agent that performs well in a controlled demo can become expensive, inconsistent, or outright risky at scale. It reinforces the idea that points 1 and 2 aren't separate talking points; they're two sides of the same problem: without understanding the business workflow and without a reliable stack, scaling agents into production is much harder than the demo makes it look.
The final talk combined two themes that look unrelated at first glance but are more connected than they seem:
Physical AI models represent the real world, used for example in autonomous vehicle development and robotics. What's most interesting is their use as simulation environments to train other AIs that need real-world data, today one of the sector's biggest limitations. This lines up closely with the strong push NVIDIA has been making in this space throughout 2026.
AI-driven software development is how AI is reshaping development processes end to end. This is something we've been actively working on at Onetree, and seeing it feature so prominently on the agenda of an event like Ai4 confirms we're headed in the right direction.
What we're taking away as a team
If there's a throughline across all six talks, it's this: enterprise AI is moving past the "wow" phase and into the "show me the result" phase. The companies that win this next stage won't be the ones with the newest model, they'll be the ones that understand their own workflows, keep a reliable stack, and treat every agent in production with the same rigor as any other business-critical system.
The honest exercise for any leadership team right now might not be picking the next tool; it's asking whether the stack you already have is being asked to prove its value, or just assumed to have it.
.svg)