Who Steers the Tool: The Real Question About AI and Professional Work

Since early 2025, a steady drumbeat has come from the same corner. AI is months, not years, from replacing white-collar professionals. Dario Amodei. Jim Farley. Elon Musk. Now Mustafa Suleyman of Microsoft AI, who told the Financial Times that most professional tasks involving “sitting down at a computer” will be fully automated within 18 months. Accounting, legal, marketing, project management. All named explicitly.

The Fortune piece covering it surfaced the counter-evidence:

• A 2025 Thomson Reuters report found lawyers and accountants experimenting with AI seeing marginal productivity gains, not mass displacement.
• A METR study on software developers found AI actually slowed tasks by roughly 20%.
• Apollo’s chief economist Torsten Slok noted that AI productivity gains are concentrated in tech, invisible in the broader economy.

The interesting thing isn’t who’s right.

It’s that the debate hasn’t moved in sixteen months.

Both Sides Are Looking From Outside

Both positions treat the profession as a black box. One side sees a box that will collapse. The other sees a box that won’t budge. Neither opens it.

From the practitioner’s vantage, and from the vantage of those building infrastructure for the profession, neither framing matches what’s actually happening.

Inside the profession, the work isn’t a single layer. It’s at least two. And the more useful distinction is finer than that.

The Two Layers Everyone Talks About

There’s a judgment layer. The senior practitioner’s call about whether a transfer pricing position will hold at audit. Whether a Pillar Two ETR computation accounts for grandfathered Israeli incentives correctly. Whether a permanent establishment risk has been mitigated to a defensible threshold.

This work doesn’t automate. Not because the model isn’t capable enough, but because responsibility can’t transfer. A tax authority pulling a file needs a name on the opinion. Models don’t have names.

And there’s a mechanical layer. Pulling the right statutes. Cross-referencing them against treaty articles. Verifying citations. Building the comparables table. Formatting the memo. This work is migrating to agents already. Quietly. Below the threshold where macro studies pick it up.

The “18 months until full automation” position collapses these two layers and treats them as one. So does the “marginal gains, nothing’s happening” position. Both miss the actual movement, which is happening below the judgment layer, not above it or across it.

But there’s a finer distinction inside the judgment layer that the debate misses entirely. And it’s where the real economics of the profession live.

The Distinction That Actually Matters: Templated vs. Gray Zone

For the bulk of tax work, standard returns, well-mapped transactions, routine compliance, a competent system in competent hands produces roughly the same output regardless of who’s at the controls. The work is templated. The answer is known. Two practitioners with the same tools will arrive at similar memos.

That’s not where the money is.

The money is in the gray zone. The transaction that doesn’t fit a clean precedent. The treaty article whose language was drafted for a different commercial reality than the one in front of you. The new court interpretation that no one has yet tested against an old circular. The position that’s defensible under one reading and indefensible under another.

Here, the same system produces fundamentally different output depending on who’s navigating it.

Why the Tool Doesn’t Decide

Technically, a well-designed graph RAG system can surface the relevant connections. The technology is real, and improving fast. A model can pull a tax authority circular from 2019, a Supreme Court interpretation from 2024, and a treaty article from 2002, and propose a synthesis among them.

But the system doesn’t decide which sources belong in the graph in the first place. It doesn’t decide which connections are defensible versus coherent-looking-but-wrong. It doesn’t know which question to ask next, or when to abandon a line of reasoning that looks productive but won’t survive scrutiny.

Two experts working the same tool on the same gray-zone question will steer it differently. They’ll prioritize different sources. They’ll weight authorities differently. They’ll ask different follow-up questions. They’ll arrive at different positions, with different risk profiles, different defensibility.

That steering is the expertise. Not catching the system’s errors after the fact. Setting its direction in real time.

Remove the expert from a gray-zone question and the tool produces something that looks like an answer. Keep the expert, give them the right infrastructure, and they reach defensible positions on questions that previously required a team.

The Data Catches Up

This isn’t only how it looks from inside the work. The industry data is now confirming what practitioners have been seeing.

Less than a month before Fortune published Suleyman’s 18-month prediction, the same outlet published a different finding. A new global survey by SAP subsidiary WalkMe, covering 3,750 executives and employees across 14 countries, found that 80% of enterprise workers are either avoiding or actively rejecting the AI tools their companies are spending billions to deploy. 54% bypassed those tools and completed work manually in the past 30 days. Another 33% haven’t touched them at all.

The trust gap behind those numbers is striking. Only 9% of workers trust AI for complex, business-critical decisions. 61% of executives do. A 52-point chasm between the people deploying the technology and the people meant to use it.

The cost is quantifiable. Workers lose the equivalent of 51 working days per year to technology friction, nearly two full months, up 42% from 2025. Goldman Sachs economists, separately, reported that AI saves workers who use it correctly an average of 40 to 60 minutes per day. The math is almost perfectly symmetric. The productivity AI gives skilled users roughly matches the productivity it destroys for unskilled ones.

The cause isn’t mysterious. Dan Adika, CEO of WalkMe, told Fortune the problem is structural. Companies hand out the tool without giving anyone the ability to use it. He used the analogy of giving every employee a Ferrari without teaching them to drive, providing fuel (the context), or building roads (the APIs and integrations the tool needs to work).

Brad Brown, Global Head of Tax Technology and Innovation for KPMG in the US, used the same analogy unprompted in a separate interview. The F1 car, he said, is impressive. But without “a skilled and talented driver, that tool’s not gonna do much.” Two veterans, a software founder and a Big Four tax partner, converging on the same description, separately, says something. The tool isn’t the bottleneck. The driver is the missing piece.

For tax practice specifically, this matters more, not less. The gray-zone work where value lives is precisely where steering matters most. A skilled driver in a Ferrari outperforms an unskilled driver in the same Ferrari by orders of magnitude on a winding road. The same logic holds for cross-border tax positions navigated through AI.

The Generational Shape of the Shift

The resistance shaping these numbers has a generational layer the debate hasn’t fully reckoned with.

A 2026 survey of 2,400 knowledge workers, reported in Medium, found that 29% of employees admit to actively sabotaging their company’s AI strategy. Among Gen Z employees, the number is 44%. These are the most digitally fluent workers in the workforce. They are not rejecting the technology. They are rejecting how it is being introduced.

Researchers from Peking University and Hong Kong Polytechnic University identified a related dynamic, what they call the competence penalty. Workers who use AI tools are perceived as less competent by their peers, regardless of their actual output. So they hide their AI use, or avoid it entirely, even when it would help them. A practice culture that treats AI engagement as a signal of weakness will not produce the kind of integration the technology requires.

The picture from the top is no clearer. In a 2026 Writer survey of C-suite executives, 75% admitted that their company’s AI strategy is “more for show” than actual guidance. 69% reported their company is doing layoffs related to AI. 39% conceded they have no formal strategy for how AI generates revenue. The people most loudly insisting on adoption are often the ones with the least coherent vision of what they are adopting.

For tax practice, the generational dimension is especially sharp. The career bargain that built every professional services firm was implicit. Junior people do the routine work, learn the business, and earn judgment over years. AI automates exactly the work that previously served as on-the-job training. If the templated layer goes away without a redesigned path to expertise, the profession runs out of senior practitioners within fifteen years. Architecture isn’t just a technology question. It’s a workforce question.

What the Debate Has Backwards

The real question isn’t whether AI replaces accountants and lawyers in 18 months.

It’s whether the profession designs the layer-by-layer transition, or has it designed for it.

Suleyman’s framing makes the decision for the profession. The skeptics pretend the decision isn’t being made. The executives admitting their strategy is “for show” are designing nothing at all.

All three leave practitioners as passive recipients of someone else’s architectural choices.

That’s the part worth refusing.

What Gets Gained

What the public debate misses entirely is what gets gained.

Move the mechanical work to agents. Move the templated judgment work to agents as well, the routine determinations where the answer is essentially known and the expert is mostly confirming what the system surfaces. What’s left isn’t less work. It’s the work that always mattered but never had room to breathe.

The gray-zone questions. The strategic positioning. The novel arguments built from three sources nobody has yet connected. Deeper client conversations. Multi-year structural thinking. Mentorship of the next generation of practitioners. The questions that don’t get asked in a practice that spends seventeen weeks a year on a single corporate return.

The technology narrows the gap between practitioners on the templated work. A junior associate working with the right infrastructure can produce output that previously required senior review. But it widens the value of expertise on the gray zones. The floor rises. The ceiling rises with it.

The pricing implications are worth sitting with. Standard work will compete on price and speed. Gray-zone work will compete on judgment. And judgment will be worth more, not less, when the templated layer underneath becomes commoditized.

What This Means for Practitioners

The seventeen weeks a year a large corporation spends complying with a single tax form were never the practice of tax.

They were the cost of practicing it.

That cost is what’s actually being removed. What remains, what practitioners will spend their hours on, is the part of the work that called them into the profession in the first place.

That’s worth getting right.