Boards stopped asking how much AI you are deploying.
The question now put to executives is different:How do we create enterprise value from AI while protecting customers, employees, intellectual property, and shareholder interests?That question has a concrete answer, and it is not a maturity score. It is an inventory: the consequential actions your organization will not let anything execute without authorization, grouped by who each action protects.
The constraint is governance capacity, not model capability
Organizations cap what they delegate to autonomous systems because the blast radius of a wrong action is unbounded — not because the models cannot do the work. Every action on this page is one you would otherwise have to keep a human in front of, and often not for the judgement. For the accountability. Authorization at the execution boundary raises that cap. The organization delegates the work and keeps the authority. For a board, that reframes the whole conversation: protection is not the tax on velocity, it is the precondition for it.The four constituencies
Every action below is a canonical action carrying its own risk posture and AI-risk rating. Several protect more than one constituency; those carry an explicit note.- Customers (24)
- Employees (9)
- Intellectual property (11)
Can an autonomous system reach a customer, their data, their treatment, or their physical safety — and act — without the organization having authorized that exact consequence?24 protected actions. AI-risk mix: 5 extreme, 18 high, 1 medium.
One action restates the board’s old question
agent.tool.invoke (ACT-0029) governs the moment an autonomous agent
reaches for a tool. Because it is a protected action like any other, “how much
AI are we deploying?” stops being a number nobody can source and becomes a
governed count: how many agent invocations were authorized, under whose
authority, against which of the four constituencies.It is also the only action that can land in all four at once — an agent
holding a tool can touch a customer record, an employee’s access, a release
pipeline, and a payment. That makes it the first action to instrument, and the
one whose coverage figure a board should ask for by name.What this map establishes — and what it does not
The distinction matters more than the inventory. A register that overstates itself is worth less than no register, because a reader who finds one soft claim stops trusting the hard ones.What it establishes
The governable surface. These are the canonical action records, each
with its own risk posture, AI-risk rating, and regulatory mappings.Coverage of all four constituencies. Every group a board names has
actions defined against it — not one corner of the estate.A shared vocabulary. Board, engineering, and audit can name the same
action with the same identifier.
What it does not establish
That your organization has provisioned these. A canonical action
existing is not the same as your policy carrying that action’s
requirements.That the grouping is authoritative. Assignment to one constituency is a
presentation choice; the canonical action catalog is the authority.That coverage equals enforcement. What you actually enforce is a
separate, evidenced question — asked per action, and per execution path.
The part a dashboard cannot supply
Every vendor in this market can show a director a green screen. The harder question is whether the director has to believe it. An authorization record that is hash-linked, Ed25519-signed, and verifiable offline by a third party — without a network call, and without trusting the vendor that produced it — is the form of this claim that survives an adversarial audit. That, not the inventory above, is what turns governance discipline into a durable advantage.Verify it yourself
Start from the authorization surface: what AtlaSent authorizes, how a permit
binds to the exact execution plan, and what evidence the run leaves behind.