Heads of AI and AI-enabled knowledge work

Define the leadership boundary

The Head of AI may own enterprise direction, technical standards, platform coordination, portfolio visibility, or enablement. Functional leaders still own the quality and consequence of their work. Security controls access, privacy and legal specialists interpret obligations, data owners govern authoritative records, and designated business leaders approve releases.

Classify the Head of AI as owner, influencer, reviewer, or collaborator for each use case. Name the actual decision owner and escalation route. Avoid a centralized approval model that gives one team formal responsibility for decisions it cannot inspect or operate.

The NIST AI RMF Core organizes AI risk work through Govern, Map, Measure, and Manage and calls for documented responsibilities, contextual understanding, measurement, and risk treatment across the lifecycle. It is voluntary guidance, not proof that a portfolio is safe or compliant.

Sources: [1]

Build an AI operating portfolio

Create one row per proposed workflow. Record the objective, affected people, output, source systems, AI action, human judgment point, evidence retained, reviewer, release authority, current stage, stop condition, and next decision date. Use stages such as discovery, controlled test, restricted operation, and broader operation.

Require a stage-specific evidence package. Discovery may need a defined problem and source hierarchy. A controlled test needs representative cases, failure categories, and reviewer dispositions. Restricted operation needs actual permission controls, monitoring, and a fallback. Broader operation needs evidence that the control design operated under realistic load.

The NIST Generative AI Profile extends the AI RMF with generative-AI-specific risk considerations and actions. Use it to inform risk identification and testing, while tailoring the portfolio to the organization's actual data, users, systems, and consequences.

Sources: [2]

Worked artifact: a synthetic three-use portfolio

Imagine a synthetic professional-services firm considering three uses: drafting internal research notes, reviewing proposal claims, and automatically sending client status updates. The portfolio places internal notes in controlled test, proposal review in discovery, and automatic sending on hold.

For research notes, the evidence package contains source-link accuracy, unsupported-claim dispositions, and confidentiality checks. For proposal review, the team has not yet defined which requirements are authoritative or who resolves disagreements. For client messages, the stop decision reflects external consequence, weak source reconciliation, and no approved release authority.

At the next gate, the Head of AI does not approve “the model.” The portfolio records three different decisions: continue the note trial with sampling, fund a proposal source-hierarchy exercise, and redesign the client workflow so AI may draft but cannot send. The scenario is synthetic and does not establish business benefit, customer demand, or Two Prune performance.

Balance capability, control, and operating evidence

Do not treat policy completion as deployment evidence or adoption as capability. A workflow can have strong access controls and poor output verification. Another can produce useful drafts while lacking an accountable operating owner. Track each dimension separately: task fit, evidence quality, human capability, technical reliability, control operation, review capacity, and business outcome.

Include null and adverse findings. If a pilot produces no measurable change, record it. If review time increases, show where. If users avoid the approved workflow, investigate whether the task, interface, incentive, or control design is wrong rather than reporting low usage as low proficiency.

Verify the portfolio against reality

Sample use cases and reconstruct the decision from retained evidence. Confirm that the described sources, permissions, reviewers, and stop mechanisms exist. Compare intended use with actual logs and outputs. Reopen a gate after changes to model, prompt, retrieval, data, autonomy, affected population, or downstream consequence.

This method supports portfolio governance and learning. It does not establish regulatory status, certify controls, or prove return on investment. Two Prune can provide task-bounded assessment evidence from realistic simulations for human interpretation. It does not operate an AI portfolio or make employment decisions.