Agents in the Boardroom
Section 13 of Agentic Banking Architecture: A Practitioner's Guide
Boards of directors are about to get their own agents. The question is what they should — and shouldn't — do.
The information problem
Most of what a board knows reaches it through the people it exists to oversee. You meet a few times a year and read what management puts in front of you. That's not a criticism — it's the structure of the role, and until now there was no alternative.
Agentic AI can change the input to the board, not just the speed at which the input can be processed. This makes it a governance question rather than a productivity one.
Seven agents, from easiest to hardest to live with
1. Board-materials assistant. Directors can interrogate the board pack directly: what changed, what's missing, and discuss the details with a board co-pilot.
2. Broader knowledge base for the board. As above, but drawing on company systems and outside sources, not only what's explicitly tabled by management. This starts addressing the information asymmetry between board and management, and makes a dent in management's ability to control what reaches the board.
3. Premise monitoring. All board decisions rest on assumptions. This agent flags when those assumptions no longer hold, helping the board to identify which decisions should be revisited.
4. Execution conformance. Monitoring whether what was agreed in the board actually was done. This one watches management directly, and requires careful calibration of governance vs interference.
5. Agenda setting. Helping to set the board agenda by identifying the outliers and trends worth attention. This introduces a risk: whoever defines what is an "exception" steers the board.
6. Management challenge. A "devil's advocate" that argues against the recommendation, surfaces the case for an alternative, and tests for groupthink. If done right, this touches none of management's work and sharpens the board's actual job: judgement.
7. Decision-to-policy. Automatically turn board decisions into rules the organisation must demonstrably meet. The board sets policy; management operates in compliance with the set rules. In architectural terms, this is the board plugging directly into the policy codification pipeline from Section 3: a board decision becomes a versioned, tested, deployable policy artefact rather than a minute that someone interprets later.
Common considerations
For all of these, some considerations apply across the board:
Who governs the agent matters more than who builds it. What's important for board independence is who governs the agent, not who builds it. A board agent built by management's technology function but governed by the board is independent; a board agent governed by management is another management channel.
Every output must be traceable to the source. A board acting on an agent's synthesis needs to know where each claim came from. This is the same decision-level observability requirement that runs through the governance architecture — applied here to the bank's most senior decision-making body.
The board — and its agents — sense and set policy, but never operate the company. The line between oversight and management doesn't move because the board has better tools. Agents 4 and 5 sit closest to that line, which is why they're harder to live with than their technical difficulty suggests.
Better sight of information raises the standard the board is held to. "We weren't told" can be a defence. "Our agent should have caught it" is not.
This section is based on a post in the LinkedIn series on agentic banking architecture. The original post closed by asking readers which of the seven agents they would let into their boardroom first — and which they would never allow.