As part of The State of the AI Frontier 2026, AI Frontier Network invited leaders building and deploying AI in the real world to share where the frontier is actually moving. In this contribution, Dhivya Nagasubramanian, Lead AI Solutions Architect at U.S. Bank, gives a candid read on what changes in 2026 — and what to watch.
On the capability that will reach the mainstream
The single most important capability going mainstream in the next 12 months is auditable tool-use orchestration — agents that don't just take actions, but can prove those actions were correct. Most "agentic AI" today is still demo-grade: a model chaining a few API calls together. What's missing is the ability to catch silent failures, where an agent looks successful but actually did the wrong thing. That's the real blocker to deploying agents in finance, healthcare, and insurance, where a bad action has real consequences. The shift isn't about smarter models anymore, rather it's about trust and verification whoever cracks that unlocks the next wave of enterprise adoption.
On AI and decision-making
By 2027, I think the biggest change is that organizations stop asking "what can AI decide for us" and start asking "what can we actually defend." Right now a lot of decisions get handed to AI because it's fast and convenient, without much thought to what happens when someone has to explain that decision later, to a customer, a regulator, or a board. That's going to flip. Companies will start building the explanation and accountability into the decision before it's made, not after something goes wrong. So you'll see faster decisions on the small stuff, where the stakes are low and being wrong costs almost nothing, and noticeably slower, more deliberate decisions on anything consequential, because someone finally has to sign their name to it.
On the most underestimated risk
The most underestimated risk right now isn't a model going rogue or saying something harmful — it's the quiet stuff. An agent that completes a task, reports success, and is actually wrong. Nobody's watching for that because it doesn't look like a failure. It looks like a normal day. Most of the safety conversation is still focused on what a model says — bias, toxicity, hallucinated facts in a chat window. But the moment you give a model the ability to act, such as file something, move money, update a record, send a message, the failure mode changes completely. The dangerous case isn't the agent that visibly breaks. It's the one that finishes the task, looks fine on every dashboard, and got something wrong three steps into a process nobody's checking that closely. I think this gets underestimated because it's hard to demo and hard to put in a slide. "Our agent processed 10,000 claims" sounds great until you ask how many of them were quietly wrong and nobody noticed for two weeks. That gap between looking done and actually being correct is where the real risk lives, and almost nobody's built the tooling to catch it yet.
On the advice that matters now
Honestly, my advice is don't start with the model. Start with the decision you're actually trying to improve, and get really clear on what "correct" looks like for that decision. If you can't define that up front, you won't be able to tell later whether the AI is actually helping or just quietly making things worse. Most companies starting out get excited about capability first: what can this model do, how fast, how cheap, and governance ends up bolted on later, usually right after something's already gone sideways. I'd do it the other way around. Build the evaluation and audit piece before anything else, even though it feels slow and unglamorous, because trying to retrofit trust into something that's already live is a much harder problem than just building it in from the start. And start smaller than feels comfortable. Pick one real workflow with clear edges, not the flashiest thing you could do, but something where you can actually tell whether it worked or not, and get that right first. Most of the companies that struggle later scaled fast on a use case they never really defined "done" for in the first place.
On the trend worth watching
What I'm genuinely most excited about is watching agentic AI move from "interesting demo" to something that actually has to hold up under scrutiny. For years the conversation was all about what a model could say. Now it's about what it can responsibly do, and that's a much harder and more interesting problem. I find it exciting because it pulls in everything I care about at once: the technical side of building systems that work, and the harder question of how you prove they work, how you catch them when they quietly don't, and how you build enough trust that people are comfortable handing over real decisions. That intersection of capability and accountability is where I think the most meaningful progress happens this year
