Your Officers Are Now Personally Exposed for What Your AI Agents Decide.
Delaware’s extension of Caremark to corporate officers is
three years old. What is new is that it is now being applied, for the first
time, to AI systems that act on their own authority. No corporate indemnity
clause rewrites who signed off on the architecture that let an agent decide
without a human in the loop.
Executive Brief
The Delaware Court of Chancery held in McDonald’s (January 26, 2023) that
corporate officers, not only directors, owe a fiduciary duty of oversight.
For most of the three years since, that holding sat outside the AI
conversation. It no longer does.
Agentic AI, systems that exercise a decision right without a human
sign-off, converts what used to be an IT implementation risk into an
officer-level fiduciary risk at the exact moment the agent acts. Before that
moment, a failure belongs to a vendor. After it, a failure belongs to the
officer who did not architect its oversight.
The SEC’s April 2025 fraud charges against the founder of Nate, Inc., over
a $42 million gap between claimed and actual AI automation, established the
evidentiary standard now being imported into officer-level Caremark claims:
the gap between what an officer represented about an AI system and what the
system actually did.
Discovery has changed accordingly. Plaintiffs’ counsel now seek agent
prompt logs, model version histories, and architecture diagrams as evidence
that an officer had the means to know and chose not to build a system that
would find out. A board minute noting that “AI was discussed” is no longer a
defense on its own.
ISS’s own 2026 research found that only 22 percent of S&P 500
companies, and just 6 percent of the Russell 3000, disclose board oversight
of AI at all, while 58 percent of investors surveyed said companies with
significant AI use should already be applying a formal risk framework. That
disclosure gap is one ISS itself is positioned to act on in a future voting
policy, which means an officer’s undocumented architecture is already a
shareholder-level exposure in waiting, not only a personal one.
None of this is a case for slowing adoption. PitchBook’s H1 2026 data
shows AI-concentrated venture funding running near $412.7 billion, and
Medallia’s incoming ownership group committed fresh capital specifically to
accelerate the platform’s own agentic AI capabilities as its turnaround
strategy, following a $5.1 billion recapitalization write-down tied to its
original 2021 leveraged buyout. The choice in front of every officer
reading this page is not caution against speed. It is documented,
accountable architecture built at the same speed as deployment, not after
it.
No court has yet issued a ruling applying McDonald’s specifically to an
autonomous agent’s decision. That is not a reason to wait. It is the reason
to build a defensible architecture while the standard is still being formed,
rather than after a court defines the floor by ruling against the first
officer who did not.
Research by Function
Board of Directors Free download
A board that satisfies its own Caremark oversight duty while its
officers remain individually undocumented has solved half the exposure in
the room. The Declarative Board Failure Pattern names the specific
failure: a board that receives an AI update and asks no further question
has satisfied the appearance of oversight and failed its substance.
CFO
Every CFO who signs a quarterly certification is attesting to the
adequacy of controls over a process an undocumented agentic system may
already be running. Private equity and strategic buyers are already
pricing that gap directly into transaction multiples, before any
regulator asks the same question. The paper names the specific
certification language that closes it.
COO
Operations owns more agentic systems, by count, than any other
function, which means it is most likely to be carrying an undocumented
escalation mechanism, the specific gap that lets individually
unremarkable decisions aggregate into harm no single review would have
caught. The paper builds the statistical escalation rule that catches
the drift, not just the single large decision.
CHRO
Undocumented agentic oversight is a specific instance of a failure
HR’s own discipline already names: the accountability conversation that
never happened. A CHRO whose own function cannot produce a clear
accountability contract for its own agentic systems has a credibility
problem enforcing the same standard elsewhere.
CRO
A risk register that lists “AI risk” as one committee-owned line item
does not satisfy the standard McDonald’s requires. Individually acceptable
agent decisions can aggregate into an unacceptable risk position over a
quarter, invisible to a review built around single transactions.
CIO/CTO
A documented decision-rights boundary that is not enforced in a
system’s actual configuration and logging is a policy statement, not a
control. The technical architecture is where the black box problem
either gets solved or gets papered over. The paper specifies the logging,
retention, and automated escalation design that makes the boundary real.
The Complete Suite: all six White Papers plus the Executive Leadership
Playbook, the complete system, saving approximately $3,000 against
individual pricing. $4,997.
Supporting Intelligence
The Duty That Was Never Delegated
Why officers keep mistaking delegation to an AI agent for discharge of
their oversight duty, and why autonomous systems break the informal
monitoring that human delegation always carried for free.
The Officer’s Caremark Exposure, Reframed for 2026
How McDonald’s (2023), the Nate, Inc. enforcement action, and 2026
proxy guidance combine into a single, current standard for what officer
level AI oversight must look like.
A Letter on Delegation and What Outlasts Us
A personal reflection on the boards and officers Glenn has watched
mistake activity for accountability, and what it actually takes to build
something that survives the leader who built it.
What ISS’s 2026 Disclosure Gap Actually Means
A compressed read on why ISS’s own finding, that only 22 percent of
S&P 500 boards disclose AI oversight at all, makes an officer’s
undocumented architecture a board-level and shareholder-level problem, not
only a personal one.
The Fork in the Circuit
One image argument: an agent’s decision looks identical whether or not
a human touched it. The oversight architecture is the only thing that
tells the difference afterward.
Work With Touch Stone Publishers
Touch Stone Publishers works with boards and executive teams through
research, through the Executive Leadership Playbook and White Paper suite
built here, and through the Residency Lab, a facilitated session that applies
this research directly to one organization’s governance architecture.
The research above is free and complete on its own terms. Every finding,
every citation, and every framework named in the six White Papers and the
Playbook is written to be usable whether or not an organization ever engages
Touch Stone Publishers directly. What an engagement adds is application: a
facilitated session that names the accountable officer for a specific
system, drafts its decision-rights boundary in the room, and builds its
escalation mechanism with the people who will actually run it, rather than
leaving that translation work to the reader alone.
The measure of an oversight architecture is not whether it worked while
the officer who built it was still in the role. The measure is whether the
officer who inherits the seat, and the agent still running underneath them,
can be governed by what was actually built, rather than what was declared.
Talk to Touch Stone Publishers about your organization’s AI governance architecture