Why We Stopped Running AI Pilots and Started Running Decisions

The challenge

Meridian Bank had run more than forty AI proofs of concept in eighteen months. Almost none reached production. The board was asking a fair question: where is the return?

Priya Nair inherited that portfolio when she became CIO. Her read was blunt — the bank had been treating AI as a series of experiments to be admired, not decisions to be made.

The decision

Nair shut down the standing innovation lab and replaced it with a single rule: no AI work proceeds unless it is attached to a business decision with an owner, a budget line and a date.

It was unpopular. Several teams had built their identity around experimentation. But it forced clarity about which problems were actually worth solving.

Why the decision was made

"Pilots are comfortable because nobody has to own the outcome," Nair says. "The moment you attach a decision, someone is accountable for whether it worked. That is the whole point."

What happened

Within two quarters, the number of AI initiatives dropped from forty to nine. Of those nine, six reached production. The bank's fraud-review workflow, the first to ship, cut manual review time by roughly a third.

What did not work

The first governance model was too heavy. Nair had built an approval committee that met monthly, and it became a bottleneck. She replaced it with lightweight guardrails and role-based sign-off.

"We over-corrected," she admits. "We went from no discipline to too much process. The answer was in the middle."

Lessons learned

The constraint was never the technology. It was ownership. Once each initiative had a named executive accountable for the decision it served, adoption stopped being a mystery.

What comes next

Nair is now piloting agentic workflows in operations, with the same rule applied. "An agent that acts on the bank's behalf is a decision with much higher stakes. So the ownership bar is higher, not lower."