A client calls because their father has died and they do not know what happens to the accounts. Nothing in that conversation should be automated. Not the words, not the timing, not the follow-up.

Two hours later, the same firm needs a summary of eleven quarterly reports for an internal meeting. Doing that by hand is a waste of a trained person's afternoon.

Both tasks sit in the same practice, and the difference between them is not difficulty. It is whether judgment about a person is involved.

What is the working line?

The machine can prepare the work. A person decides what a client is told.

Preparation is research, summarizing, formatting, first drafts, sorting, extracting figures for a human to verify. It is work with a clear input and a checkable output, where being wrong is visible to whoever reviews it.

Deciding is different. What this client should do. How to say something difficult. Whether an exception is warranted. Which of three defensible options fits a family you have known for nine years. That work is what your clients are paying for, and handing it over is not a productivity gain, it is a change in what you are selling.

Which tasks fall on the safe side?

The boring ones, mostly, and they add up.

Turning a recorded meeting into notes that a person then corrects. Drafting internal documents. Pulling figures out of a statement so a human can check them against the source. Summarizing a long public filing. Rewriting a dense paragraph at the reading level your clients actually use. Producing a first draft of educational content you then edit.

Every one of those has a person at the end who can tell whether it is right. That is the test worth applying. If nobody in the workflow could detect a wrong answer, the task should not be delegated yet.

Where should it never go?

Anywhere the output reaches a client without a competent person having read it.

That covers advice, obviously, and it also covers the smaller things. A rescheduling note during a family crisis. A follow-up after a difficult conversation. The tone of these matters more than their content, and tone is precisely where a machine writing on your behalf will fail in ways you cannot predict.

It also covers anything where the model's confidence exceeds your ability to check it. These systems produce fluent text that is sometimes wrong, and fluent wrong answers are worse than obviously wrong ones for a firm whose product is accuracy.

Isn't keeping a human in the loop just slower?

Slower than the fastest possible version, yes. That is the trade you are making on purpose.

Your clients are not buying speed. If they were, they would have used software instead of hiring you. They are buying the fact that a qualified person considered their situation and stands behind the answer. Automating that away removes the reason you exist and saves an hour.

The right pace is one workflow at a time. Pick something internal, run it for a month, and see whether the output holds up. Then decide about the next one. We describe the same approach under Considered Build, and the underlying competencies on the AI fluency page.

For a regulated firm, bring your compliance officer in before the workflow is live rather than after. Supervision and recordkeeping obligations still apply, and yours are a question for your counsel, not for a blog post.

The firms that will do well here are the ones who can say, plainly, which parts of the work a machine touched and which parts a person decided.

If you want a read on where your own systems stand before adding anything, The Second Look is free and takes about four minutes.