The shift from operating tools to directing outcomes
McKinsey’s CEO Bob Sternfels recently revealed that the firm now counts roughly 25,000 AI agents alongside its 40,000 human employees. Eighteen months ago, the agent count was 3,000. The strategy is to achieve parity by the end of 2026. Efficiency is key here; those agents saved 1.5 million hours of work in a single year. The goal is not to replace human resources but to enable them to work more efficiently.
If the world’s most elite consulting firm has made nearly 40% of its workforce non-human in under two years, the question for bank executives and credit union leaders isn’t whether this is coming – it’s whether you’re ready.
Welcome to agentic engineering. You describe the outcomes, not the process. The skill that matters most isn’t technical proficiency; it’s clarity of intent. And the hardest problem isn’t building the capability – it’s governing it in an industry where every autonomous decision carries regulatory weight.
The new operating model
For decades, banks and credit unions invested in making their people better at operating tools: core platforms, fraud systems, loan origination software, CRM dashboards. Master the tool, master the job. That era is ending. The shift is from execution to judgment.
The instinct will be to deploy agents by department. A fraud agent, a lending agent, a compliance agent. But the real power comes from organizing agents by outcome.
What does a bank actually need to accomplish? Detect and stop fraud in real time. Approve good loans faster while managing risk. Retain customers while acquiring new ones. These aren’t departmental problems. They cut across every silo.
A team focused on “protect this customer’s financial life” doesn’t just monitor transactions. It cross-references account behavior with onboarding patterns, flags synthetic identity signals during origination, adjusts authentication friction based on risk context, and surfaces retention opportunities when a relationship is at risk. No human touching a queue.
That’s a lot of autonomous decision-making across fraud, lending, and customer management at once. Every one of those decisions is examinable. Every one needs a trail. Hold that thought.
The human-agent ratio question
Workforce planning is no longer about headcount. It’s about the right human-to-agent ratio for each outcome.
Some outcomes need tight oversight. A complex commercial loan might require one experienced lender orchestrating agents that handle data gathering, risk modeling, and covenant analysis, with the human applying judgment that no model can replicate.
Others can run with far less involvement. Fraud monitoring can operate autonomously around the clock, escalating only when confidence thresholds require it. Retention agents can detect behavioral signals like reduced direct deposit activity and declining card usage, then trigger personalized outreach without a person in the loop.
At the far end, agent teams run autonomously for weeks on regulatory reporting, BSA/AML case prep, or portfolio stress testing, surfacing results for human review rather than waiting for human initiation.
The institutions that figure out these ratios first will operate at a fundamentally different cost structure and speed. But the further you move toward autonomous operation, the more critical governance becomes. Examiners won’t accept “the AI handled it.”
The skill that actually matters now
If agents handle execution, what do your people need to be great at?
Understanding what customers actually want. Not what they asked for at the branch, but the underlying need. The member asking about a home equity loan might need a debt consolidation strategy. The business owner applying for a line of credit might really need cash flow forecasting. When agents can execute anything, the premium goes to the person who articulates the right problem.
Thinking in outcomes, not tasks. Not “process this application” but “get this borrower into the right product in under 24 hours with full compliance documentation.” The humans set the destination. The agents navigate the way to it.
Knowing when to trust and when to intervene. With agent teams running at scale across your loan portfolio, transaction monitoring, and customer engagement, the new core competency is judgment about when the system is performing and when it needs a human hand.
The governance question
Every capability described above creates a corresponding governance demand. That’s not a coincidence; it’s the central challenge of agentic engineering in a regulated industry.
Gartner’s research suggests roughly half of AI agent deployments will fail due to insufficient governance. In financial services, where every agent decision on fraud, lending, or customer treatment carries regulatory weight, that should alarm you.
When you have dozens, eventually hundreds, of agent teams across your institution, you need to know which agents are running, what tools and data they’re accessing, which models power their reasoning, and what decisions they’re making. You need complete audit trails. Not bolted on at exam time, but built as the foundational layer everything runs on.
This is the problem AI governance platforms are emerging to solve. A single control plane to govern agents, tools, data sources, and models with the traceability that regulators expect and your risk committee demands.
The bottom line
Agentic engineering isn’t about replacing your people. It’s about redefining what they do. The loan officer understands the borrower’s real need and directs agents to structure the solution. The fraud analyst sets strategy and reviews edge cases instead of clearing queues. The branch manager orchestrates human-agent teams around member outcomes.
The winners won’t have the most agents. They’ll have people with the clearest intent, and governance that makes every autonomous decision auditable, explainable, and defensible.
