Council Post: Agentic AI And The End Of The Unowned Exception

Krupesh Bhat, CEO & Founder, Melento, where intelligence meets collaboration to transform business with agentic AI.

getty

​For decades, enterprises digitized functions, not journeys. Finance got an ERP, while sales got a CRM. Each tool sharpened one department, and the work still crawled between them, because nobody was ever asked to own the crawling. When something didn't fit the pattern, a person made a call and moved on. Nobody wrote it down, because nobody had to.

Agentic AI removes that absorption layer, the need for that manual buffer. It doesn’t just handle the routine cases end-to-end; it can also understand, organize and act on the messy exceptions people traditionally had to step in and manage. And that’s the part every “we redesigned the workflow” case study skips: What happens when the work doesn’t follow the workflow.

Who Owns The Exception

An automated workflow handles the normal cases well: Documents are extracted, identities are verified and approvals are routed. But real business value and risk both concentrate in the other cases, like the clause that doesn't match any known pattern, or the regulation that changed before the rule did. What happens when you route that case to a human and get an override? In most enterprises, nothing. The exception disappears into an email or a meeting and the workflow got faster. But the organization didn't get smarter.

Gartner projects 40% of enterprise applications will carry a task-specific AI agent by the end of 2026, up from under 5% a year earlier; that's an eightfold jump. But Deloitte's 2025 Emerging Technology Trends research found that while 38% of organizations are piloting agentic solutions, only 11% are actually running them in production. The gap between those two numbers is what happens the first time an agent hits a case nobody scoped for and there's no human whose job it is to care.

A pilot only ever has to handle the normal case, under the supervision of whomever already owns that function. Production means the agent starts generating exceptions nobody has assigned an owner to catch, record and convert into a better rule.

A digitized function ships once and hits its SLA every time. Exceptions arrive with new business logic attached, and "ship it once" has nothing to say about that. In a pilot, an override is a footnote someone jots down. In a live workflow processing thousands of accounts, that same unrecorded override is a policy gap, and the question is whether you find it or the auditor does.

And once a single agentic thread runs the whole journey, the enterprise inherits accountability for every case it escalated and the organization then mishandled.

Treat The Exception As The Product

Most transformation programs measure success as automation rate, meaning what percentage of cases do not need a human to handle. That number says nothing about whether the organization is any smarter than it was a year ago.

BCG's research on retail banking shows that AI agents already reduce collections costs by 30% to 40%, and BCG projects AI-first banks could run cost bases 30% to 40% lower than peers by 2030. Those gains came from banks treating fraud detection, KYC and compliance exceptions as the actual intelligence layer of the workflow, not a comment field someone gets to later. Once you internalize that, resourcing decisions flip, and the question becomes what every escalation is teaching the system that the last hundred didn't.

One Thread, One Owner, One Exception Log

This is where most redesigns quietly fail, and it has four parts:

1. Exception ownership. Someone owns the final call on whether an override is a one-off or a new rule, not "escalate and see."

2. Decision provenance. When a human accepts, rejects or overrides the agent, the organization can reconstruct why—not from a comment field that made sense to the person who wrote it, but from a structured record that holds its meaning after they're gone.

3. Feedback conversion. A correction that never becomes a rule or a training example was never really captured, it was postponed. Most enterprises postpone it indefinitely and call it human-in-the-loop.

4. Context accumulation. Without durable, cross-case memory, the 10th version of the same exception looks exactly as novel as the first.

Skip any one of these and the "redesigned workflow" is still the old crawl with a shinier front end. This is uncomfortable to implement, because someone in legal, compliance or operations has to sign their name to a permanent record of judgment calls the business used to make quietly and never revisit.

A system worth calling redesigned treats the case a rule-based system wouldn't catch as the entire reason it exists, with a named owner for what happens when it fires. The alternative—automate the predictable 90% and let the other 10% pile up in someone's inbox—is just a deferral.

The Test

The key question is: For the last hundred cases your workflow escalated to a human, can you show who owns the final call and why each override happened? How about what changed in the system because of it?

If the honest answer is "it's in an email somewhere," you just automated the normal case and left the organization's actual intelligence exactly where it's always been, in someone's memory, on its way out the door.​

Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?