Operations excellence leadership for AI-enabled work

Define the leader's decision and role

Start with a process decision such as whether to investigate, test, restrict, expand, revise, or stop one workflow. Describe the customer or internal outcome, steps, handoffs, sources, queues, rework, and material exceptions. Record where AI could assist and where accountable human judgment remains.

Classify the operations excellence leader as owner, reviewer, adviser, or collaborator for each part. They may own improvement method and cross-functional coordination, while the functional process owner owns output quality. Technology, security, privacy, legal, data, and frontline teams retain their own responsibilities. Improvement leadership does not automatically imply budget or system authority.

The NIST AI RMF Core calls for documented context, intended purpose, human oversight, roles, impacts, testing, and monitoring across Govern, Map, Measure, and Manage. It is voluntary guidance, not an operating verdict for a local process.

Sources: [1]

Build the friction-to-evidence improvement charter

Record the process boundary, triggering event, current steps, baseline evidence, suspected friction, affected roles, proposed AI task, authoritative sources, output, human decision, expected benefit, possible displacement of effort, failure modes, test conditions, measurement window, owner, reviewer, stop rule, and next decision.

Distinguish elapsed time, touch time, waiting, rework, exception handling, and downstream review. AI may shorten drafting while increasing source verification or escalation. Measure the complete process and preserve null or adverse results instead of reporting only faster steps.

Specify the unit of analysis before testing: case, handoff, exception, or completed outcome. Define how withdrawn, duplicated, reopened, and escalated items are treated. This prevents a faster subset from being presented as improvement across the full workload.

The UK government's 2026 employer guide to AI upskilling emphasizes practical, contextualized learning integrated with real tools and workflows. Its evidence can inform change design, but it does not prove that a particular intervention improves a local operation.

Sources: [2]

Worked artifact: a synthetic invoice exception process

Imagine a synthetic shared-services team handling invoices that do not match purchase orders. The charter maps intake, document retrieval, coding, buyer clarification, approval, and release. The proposed assistant extracts invoice fields and drafts an exception note, while an analyst determines the exception category and an approver decides payment.

The first controlled set includes a duplicate invoice, a partial delivery, an amended purchase order, and an unreadable attachment. The assistant drafts notes more quickly but treats the amended order as older evidence and sends two cases to the wrong buyer. Analysts spend extra time reconstructing source authority.

The charter records that drafting time fell inside the exercise while total handling time did not. It adds document-status labels, a stop condition for conflicting order versions, and routing verification before any expansion. A second synthetic set correctly stops the amendment conflict, but the unreadable attachment still requires manual handling.

The leader's decision is to continue restricted testing for extractable cases, not declare the whole process improved. The example is synthetic and does not represent Two Prune customer results.

Verify the improvement and its side effects

Reconstruct sampled cases from intake to final decision. Confirm source versions, corrections, escalations, review effort, accessibility, and downstream defects. Compare the charter with actual system configuration and frontline workarounds. Ask whether faster visible steps displaced work into an unmeasured queue.

Reopen the charter after changes to source systems, model, prompt, permissions, volume, exception mix, staffing, or consequence. Two Prune can provide task-bounded evidence about how participants navigate and improve a controlled workflow. It does not operate customer processes, attest to control effectiveness, or make employment decisions.