What “Human in the Loop” Gets Wrong About AI in Accounting Firms

The missing step isn’t “human in the loop.” It’s “human at the helm.” People deciding what to hand off, when and how, before the work starts.

Share
What “Human in the Loop” Gets Wrong About AI in Accounting Firms
Image by Luís Eusébiolce

This article was originally published in INSIDE Public Accounting on November 16, 2026.

When a client asks, “Do you use AI at your firm?” what do you say? There’s usually a question behind that question. Is my data safe? Are you being as efficient as you can with my work? Or the hard one: “If AI’s doing everything, what am I still paying for?”

A short, thoughtful AI policy gives everyone at your firm the same confident answer. Most AI policies cover security and approved tools, and those items are foundational. But they skip the part clients are really asking about: a clear line between work that can go to AI and work that never should, even once AI is capable of it.

In June, the IRS Office of Professional Responsibility issued guidance on AI in tax practices. Its core message matches what tech companies have said for years: an experienced human should always review AI’s work. That practice, usually called “human in the loop,” is essential for any firm using AI. But it’s also not enough by itself.

Review is a necessary failsafe, and reviewers aren’t the problem. The default setup is. AI does the work, then a person checks it. That check confirms the output is technically accurate, but it doesn’t ask whether the task should have gone to AI at all. And it can’t pull client data back out of a tool it never should have gone into in the first place.

The missing step isn’t “human in the loop.” It’s “human at the helm.” People deciding what to hand off, when and how, before the work starts. That happens at two levels: the firm sets the limits, and whoever owns the client relationship applies them to the work at hand.

At the Firm Level: Draw the Lines

Every firm needs a written AI policy, updated as the technology changes. Large firms typically form a cross-functional committee to own it. At smaller firms, one person usually takes ownership. Most policies already cover approved tools and security requirements. It’s important to add a section that explains what work never goes to AI and how everything else gets handled. Some firms call it “The Human Line.”

One managing partner I spoke with established his firm’s Human Line after a near-miss. An AI tool drafted partner status emails and nearly sent them before anyone reviewed them. Now it only saves drafts. Reviewing those drafts is “human in the loop.” Deciding that AI never sends those emails on its own is “human at the helm.” You need both.

If your first reaction is “we don’t have time for that,” the IRS sees it differently. That same June guidance makes clear that, under Circular 230 §10.36, firm leadership is responsible for establishing adequate procedures for AI use, including written policies and documented staff training.

Start by Sorting the Work

To draw your Human Line, sort your firm’s work into three tiers.

Tier 1: Never Hand Off

Ask: Is part of the value we’re delivering created because a person does the task?

One firm owner told me about a longtime client who came to her distraught after getting caught in a Ponzi scheme. AI could have laid out technically accurate next steps, but that’s not what this client needed. They needed to sit across the table from someone they had trusted for years and hear what to do next.

Tier 2: AI Drafts, a Human Decides

Ask: Does the task appear routine but have edge cases that require human judgment? Does it carry regulatory or ethical exposure for the firm?

Before quarterly advisory meetings, a managing partner at a midsize firm has a purpose-built AI agent compare each client’s financials to industry benchmarks, then draft the presentation and questions to ask. The advisor still runs the meeting because only they know which topics to skip given what’s happening in that client’s life.

Tier 3: AI Does It, a Human Verifies

Ask: Is this routine, low-exposure work where checking the output before it’s used is enough?

One midsize firm owner estimates AI has cut IRS notice research from two to three hours to 10 to 15 minutes.

Notice who makes the call in Tiers 1 and 2: the person who owns the relationship. The firm draws the line. The relationship owner applies it, one client at a time.

Where to Start

This week, pick one task you’d keep human even if AI could someday do it well. Write down why and then turn that reasoning into something a client could read: a few plain sentences about what never goes to AI.

That’s your first draft of The Human Line for your firm, and it’s your answer the next time a client asks.