What parts of my client work should I never hand to AI?

Five, and the word permanently is chosen deliberately, because these are structural rather than waiting on better models:

1. Final judgment on high-stakes calls. The recommendation the client will act on with real money, real careers, real consequences. AI can brief it, argue both sides, and stress-test it; the call itself needs a bearer of its risk.
2. The hard conversations: bad news, underperformance, the engagement that should end, the truth the client is avoiding. These change relationships, and only a human can spend relational capital.
3. Ethical lines: conflicts of interest, confidentiality edges, anything where your professional integrity is the backstop. Delegating the decision does not delegate the accountability.
4. Relationship repair: the missed expectation, the misunderstanding, the wobble in trust. Repair requires the injured party to feel a human choosing to show up, and no artifact substitutes.
5. Reading the person live: the hesitation behind the yes, the fear under the question, the thing not being said. This data never enters any prompt, because the client has not said it.

Every item shares one property: the value is not in the output but in who is present for it.

Because their value is constituted by a human bearing them, not merely produced by one, and that distinction survives every capability curve.

Take them in turn. A high-stakes recommendation is not just words; it is words plus someone answerable for them, and the answerability is what the client is buying. Research on AI and knowledge work adds the practical edge: beyond its competence, AI output degrades measurably, 19 percentage points more wrong in the landmark field experiment, while sounding no less confident, and client stakes concentrate exactly in that beyond.

The hard conversation works only as an act of relational spending: the message lands because someone who values the relationship chose to risk it. The same words from a system are information; from you, they are care.

Ethical calls need a party who can be held to account, by clients, by professional bodies, by your own conscience. Accountability does not pass through an API.

And reading the live person depends on data that exists only in the room: tone, timing, the flinch. No context window contains what was never said aloud.

None of this is romanticism about human specialness. It is the plain observation that some services are relationships wearing deliverables, and the relationship is the part that cannot be outsourced to anything.

It looks like the human moments getting better because everything surrounding them got carried. The pattern, applied to each:

• Before the high-stakes call: AI assembles the full picture, the client's history, the options considered, your method's read, the stress-tests, so the judgment happens on complete information instead of reconstruction. After: the decision documented, the reasoning captured, the follow-through scheduled.
• Before the hard conversation: the timeline of what happened, drafts of the key points to land, even rehearsal against likely reactions. After: the summary, the commitments, the check-in that ensures the conversation stuck. The conversation itself: just you, present, unhurried.
• Around ethical edges: the research into what your obligations actually say, the documentation trail that protects everyone. The line call: yours.
• Around repair: the facts assembled honestly so you enter knowing what actually happened rather than defending a guess.

The compound effect is the point: advisors who build this architecture report walking into their hardest moments better prepared than they have ever been, precisely because the machinery cleared the approach. The never-list does not shrink what AI does for the practice. It aims it.

Expect the temptation, because it will arrive wearing good arguments: the AI's draft of the hard message is genuinely decent, the week is genuinely brutal, and sending it would genuinely save an evening. The line holds through structure, not willpower:

1. Write the never-list down. Five lines, pinned where you configure your tools and visible when you are tired. Instinct erodes under deadline pressure; documents do not.
2. Name the tells of drift: judgment calls getting batched with routine approvals, hard messages going out without a live conversation attached, your review pass becoming a skim. Each is a small breach that normalizes the next.
3. Use the drafts correctly: AI drafting the hard message as thinking support is fine, healthy even, as long as the delivery stays live and the final words pass through you. The breach is the artifact replacing the encounter, not assisting it.
4. Audit quarterly with one question: which moments did the machinery touch this quarter that my list says it should not have? Honest answers early are cheap; discoveries by clients are not.

And tell your clients the boundary exists, plainly. 'The judgment calls and the important conversations are always me' is one sentence, it is reassuring, and having said it publicly is itself a commitment device.

The edges move; the core does not, and knowing which is which prevents both complacency and paranoia.

What genuinely shifts with capability: the preparation quality around the moments, the share of routine judgment, low-stakes, well-documented, pattern-matching calls, that your method can safely encode, and the sophistication of the briefing you walk in with. Yesterday's 'AI cannot even summarize this reliably' is gone, and tomorrow will take more of the periphery. Let it.

What does not shift, because it is not a capability question: stakes still need a bearer, ethics still need an accountable party, relational spending still requires a relationship, and the unspoken data in the room still never reaches any model. These are facts about what the moments are, not about what machines can do, and better models make them more visible rather than less: as everything else automates smoothly, the human moments stand out as the product.

The practical posture: revisit the periphery of your list annually, and treat the five-item core as settled architecture. The advisors who thrive hold both simultaneously, aggressive about the edges, immovable about the center. Watching how that boundary evolves across real practices is part of what the Collective Wisdom newsletter is for.