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, Karanjot Jaswal, CTO & Co-founder at Cinchy, gives a candid read on what changes in 2026 — and what to watch.
On the shift that defines 2026
The shift from experimentation to production. The early push was adoption for its own sake, and that drove costs up without a clear line to value. As organizations start demanding ROI, the emphasis moves from "are people using it" to "is it delivering business value and is it operationalized properly." There's a second-order effect too: safety mattered less when AI was mostly a chat interface advising a user and generating artifacts. The moment it starts taking action in the environment, which is exactly what the push for ROI drives, the stakes change and robust controls become non-negotiable. The experimentation phase rewarded enthusiasm. This next phase rewards discipline.
On the real unlock
The governance and control layer that makes agentic action safely deployable. The market is busy pricing model capability, but the thing actually gating high-value adoption is the ability to control what AI does and what data it touches inside a real environment. The models aren't the bottleneck. The controls around them are.
On the trap to avoid
That human-in-the-loop is a cheap, simple safety bolt-on. The common assumption is that you take a "mostly right" system, put a person in front of the output to check it, and you're covered. But if that person has to redo the reasoning to verify the answer, you've duplicated the work and erased the acceleration that justified the tool. Leaders who treat HITL as a checkbox will quietly kill their own ROI. The ones who get it right will invest in tailored review experiences that surface the right context at the right moment, so a person can confirm an outcome quickly instead of reproducing it. HITL is a design problem, not a compliance step, and that's the bet most will misread.
On the organization that adapts
The core shift is doing more with less, but that's about output per person, not cutting headcount. The naive conclusion is that higher output means fewer people. The reality is that if you cut the capacity, you lose the subject matter expertise that's critical to keep innovating, and that's a short-lived benefit: your competitors who reinvest the gain will eat the opex and run faster than you. The window to compound that advantage is now. So the work shifts rather than shrinks. It splits between people driving continuous improvement, finding the next thing to operationalize and refining how AI is applied, and operators running the workflows, where the role itself changes as much as the number, since overseeing "mostly right" AI demands more judgment and exception-handling than before. The other shift is democratization: a lot of what the business used to depend on IT for, it can now do directly, which is also exactly why a governance foundation matters, since decentralized capability without control is how ungoverned spend creeps in. None of these roles disappear. What changes is the balance and distribution between them.
On the benchmark that matters
The signal to watch is AI spend discipline, specifically whether organizations can tie their AI spend to quantified value. Historically every initiative cleared a bar: dollars attached, a business case to justify them, a decision to proceed or not. AI has quietly broken that discipline. In many organizations there's now effectively unmetered spend chasing a benefit nobody has quantified, justified by the promise of value rather than the measurement of it. That's the anomaly, and it won't hold.
By December 2026, ROI will have far more importance and attention than it does today, but the majority of organizations still won't have a credible way to measure AI ROI. The intent will be there ahead of the capability. 2027 is when that changes drastically, as the discipline catches up and ROI becomes a real gate on where AI gets applied rather than an after-the-fact rationalization.
