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, Michael Charles Borrelli, Director at AI & Partners, gives a candid read on what changes in 2026 — and what to watch.
On the capability that will reach the mainstream
Reliable agentic execution — models that don't just answer but carry out multi-step work across tools, with verification built in. The capability itself isn't new; what becomes mainstream is the trustworthy version: agents constrained by machine-checkable guardrails, structured logging, and graceful escalation when they're uncertain. That combination is what moves agents out of demos and pilots and into back-office and middle-office processes that someone is actually accountable for. The breakthrough isn't more autonomy — it's autonomy you can audit. By this time next year, "an agent did it, and here is the trace proving it did it correctly" becomes a normal sentence in enterprise operations.
On AI and decision-making
Less by replacing the decision-maker than by collapsing the time between question and evidence. Routine, high-frequency decisions — triage, pricing, credit, scheduling, first-pass review — get delegated to systems operating inside defined policy bounds, with humans supervising the exceptions rather than every case. For consequential decisions, AI shifts from "the analyst who builds the deck" to "the analyst who pressure-tests it," generating counterarguments and surfacing what the human missed. The real change is cultural: organizations that win will demand that AI-assisted decisions come with an audit trail — what inputs, what model, what reasoning, what was overridden. Decision-making becomes faster and more documented at once, which is historically rare. The losers will let speed outrun traceability and won't be able to explain their own decisions when a regulator, board, or customer asks.
On the most underestimated risk
The operational gap between a working pilot and a governed production system — and the trust collapse that follows when that gap is ignored. Most attention goes to dramatic, abstract risks; the underestimated one is mundane and immediate: organizations deploying autonomous systems without the means to verify outputs, detect drift, or recover from failure. The failure mode isn't a rogue superintelligence — it's a quietly-degrading agent taking thousands of slightly-wrong actions with no audit trail, discovered late. The predictable sequel is worse than the incident: a high-profile failure triggers a blanket ban, and the organization swings from naive adoption to blanket prohibition, destroying two years of progress. Verification, monitoring, and oversight are treated as compliance overhead when they are actually the precondition for deploying anything at all.
On the advice that matters now
Start with the process, not the model. Your constraint is almost never model quality — it's messy data, undocumented workflows, and unclear ownership. Pick one genuinely valuable, well-bounded workflow, fix the data and the process around it, and instrument it so you can measure whether the AI is right. Build the evaluation and oversight scaffolding before you scale, not after a failure forces it. Treat governance and the EU AI Act as a design input from day one, not a final gate — it's far cheaper to build in than to bolt on. And resist the pull toward maximum autonomy; earn it incrementally as the system proves itself. The companies that win won't be the earliest adopters or those with the biggest model — they'll be the ones who industrialized one workflow properly and learned to trust it, then repeated that.
On the trend worth watching
The maturing of the assurance layer — the tooling that makes AI auditable: continuous monitoring, drift detection, end-to-end traceability, calibrated human oversight. I find it genuinely exciting because it's the quiet capability that lets every other capability go live. It's what allows a regulated bank, insurer, or health system to move AI from sandbox to live operation responsibly. It reframes governance from a brake into an accelerant — the thing that lets organizations deploy faster because they can finally stand behind what the system does. After years of capability hype, watching the unglamorous infrastructure of trust catch up is the development I care about most, because it's the one that turns AI from a fascinating tool into dependable institutional capacity.
