Agentic AI is an intern plus a manager. Most decks forget the manager.

The sentence that scores

An agent is a model that can plan a few steps and call tools (search, a claims system, a calculator) until a stop rule hits. A harness is the operating system around it: which tools exist, what it may remember, how you score a frozen eval set, where a human must click, and how you explain the decision later.

Consulting work on “agentic AI” is mostly harness design and change management. The model is the cheap part. Write-access without evals is how you get a newspaper story.

• Tools: read-only first. Write actions behind a threshold.
• Evals: 100–300 real tasks, scored the same way every week.
• Stop rules: max steps, max dollars, max customer messages.
• Logging: you will have to tell a regulator or a GC why it paid.

How this appears in an MBB case

The board wants “an agent that handles claims.” Your job is to split: copilot (draft, human decides) versus agent (act on the easy band). Then put $ on extra leakage. Then refuse to skip the copilot if the error table is ugly.

Frontier talk that helps: reasoning models for the hard tail; a small classifier plus a bigger model on exceptions; do not run the most expensive model on every ticket.

Practice this on Clearclaim Mutual. The case is built to force the copilot-versus-agent choice.