At 2:47 this morning, while your executive committee slept, an AI agent acting for one of your corporate clients renegotiated the terms of a treasury facility with an AI agent acting for you. The exchange took 11 seconds. No human read it. Multiply that by a million, and you have a picture of financial services in 2029.
Boards have spent the past two years asking sensible questions about AI: data, models, copilots, cost. But those questions assume AI is a tool your people use. The next wave is different. Agents don’t advise—they act, they accomplish goals; they will negotiate, price and contract. And even more critically, they won’t just act inside your institution; they will transact with everyone else’s agents. Your customers’ agents will comparison-shop across the market in milliseconds, with none of the loyalty or inertia that has quietly protected your organization’s margins for a century. Your suppliers’, counterparties’ and—eventually—regulators’ agents will deal with yours at machine speed.
If your organization is in the midst of agentic development and use, three areas deserve board attention now, not in next year’s strategy cycle.
First, the risk is shifting from your AI to the space between everyone’s AI. When thousands of agents interact, markets acquire emergent behavior nobody designed. We have already seen what algorithmic feedback loops do to equity markets in a flash crash. Imagine similar dynamics in deposit flows, FX hedging or credit pricing—or agents converging on pricing patterns that look, to a regulator, exactly like collusion no human intended. Your institution can have impeccably governed agents and still be blindsided by the ecosystem.
Second, agent design is strategy work, not just IT. What authority you delegate, what negotiating posture you encode, what your agent discloses— these decisions will be made by agents multiple times a day. An agent that squeezes every basis point may win a transaction but lose the client relationships. In an agent economy, your agent’s character is your institution’s character. This is the most consequential delegation of corporate authority since the power of attorney, and in many organizations it is getting designed by employees three levels removed from the executive committee.
Third, human review is being dissolved by physics, not negligence. “Human in the loop” is the comfort blanket in every AI policy a board has approved. But humans cannot meaningfully review interactions occurring at machine speed and machine volume. Oversight is quietly degrading from human-in-the-loop to human-near-the-loop to human-informed-afterwards. The honest response is to redesign control around hard boundaries—what agents may never do, when they must halt and escalate. This must be done as part of agentic design.
Which brings us to the boardroom itself. Today’s directors carry deep instincts for human organizations: incentives, politics, culture, conduct. Almost none carry instincts for machine ecosystems. So, ask three questions at your next meeting: What can our agents commit us to today? What happens when they meet a smarter agent on the other side? And who around this table could genuinely interrogate the answers?
The agent economy will not wait for your governance calendar. You need to start answering the questions and creating the environment now.
