I’ve been watching the AI-is-going-to-take-jobs conversation for a while now. Most of it has felt like hype — big claims, small data, and a lot of extrapolation from coding benchmarks that have nothing to do with what accountants actually do all day.
This one is different.
At the beginning of April 2026, MIT FutureTech published a study that evaluated AI performance across more than 3,000 real labor market tasks drawn from the U.S. Department of Labor’s O*NET database — the federal taxonomy that defines what people actually do at work. They collected over 17,000 evaluations, ran tasks through 41 different AI models, and had domain-expert workers score the outputs themselves. Not AI judges. Real people who do those jobs.
The question they were trying to answer: is AI disruption coming like a crashing wave, or is it rising like a tide?
The Number That Should Get Your Attention
93%.
That’s the share of Business and Financial Operations tasks flagged as having meaningful AI time-savings potential — defined as at least 10% time reduction. That ties accounting and finance with computer and mathematical occupations for the highest exposure of any job category in the study.
Before you close the tab: the finding is more nuanced than that number suggests, and honestly more reassuring than most AI headlines.
Crashing Waves vs. Rising Tides
Up until recently, the dominant narrative in AI research came from a group called METR and related work on coding and software engineering tasks. Their data suggested AI capability arrives in sudden bursts — models that can’t do a task at all suddenly, with a new release, do it almost perfectly. They called this a steep logistic curve. Think of it like a cliff: fail, fail, fail, then succeed.
Think about it as the difference between a market correction and slow-building inflation. A correction hits hard and fast — your portfolio is fine Tuesday, down 25% by Friday. The crashing wave hypothesis says AI is a market correction. You won’t see it coming until it’s already happened.
The MIT team tested that hypothesis with a much broader dataset — 11,500 real occupational task statements across management, healthcare, legal, education, construction, and accounting. What they found was a surprisingly flat relationship between task complexity and AI success. AI doesn’t perform dramatically worse on a four-hour task than it does on a fifteen-minute task. The drop-off is gradual. And when newer models come out, the whole performance curve shifts upward equally across short and long tasks.
That’s a rising tide, not a crashing wave. And it changes the planning calculus significantly.
What the Data Actually Shows
The scoring threshold the researchers used for “success” was a 7 or above on a 1-to-9 scale — defined as: a manager would accept this output without requiring edits. Usable as-is. That’s a bar most of us in accounting and finance can relate to. It’s basically what you’re asking when you hand something to a staff associate: can I use this, or does it need significant work first?
At that bar, here’s where things stand for Business and Financial Operations: roughly 57% success rate today, improving by 8 to 11 percentage points per year. The feasible task duration AI can handle at a 50% success rate went from about three hours in Q2 2024 to about one week by Q3 2025 — doubling roughly every 3.8 months.
One important nuance from the transcript: 50% is still a coin flip. The researchers are clear that getting AI success rates into the 90s — where you’d actually need them for most accounting and finance work — takes considerably longer. By 2029, the MIT projection puts most text-based professional tasks at 80 to 95% success. Near-perfect performance is several years beyond that.
What This Means for Your Practice
I spent a lot of time early in my career watching firms adopt citizen-led technology tools — Alteryx, Power BI, UiPath. Every single rollout started the same way. People said it wasn’t quite good enough yet. Six months later it was handling half their workflow. AI is following the same pattern, just faster.
Here’s a simple test. If you handed a deliverable to a junior associate and they handed it back usable — no edits required — 57% of the time, that associate is already carrying real load for your team. AI is at that level across more than half of your text-based work right now. The question isn’t when this will affect accounting. It’s which tasks in your workflow are you still doing manually that AI could handle at 57% accuracy today — and what happens to your staffing model when that number hits 70% in two years?
The good news buried in all of this: the rising tide pattern means you can see it coming. The water is already at your ankles. The firms that navigate this well are the ones paying attention to the waterline now — not the ones that notice it when it hits their waist.
And the research is clear about where human expertise retains the longest runway: complex judgment, client relationships, novel situations, and work where the cost of error is high. Getting AI from 80% to 99% on those categories takes years, not months. That’s your window.
Key Takeaways
- AI automation of knowledge work is a rising tide, not a crashing wave — broad, simultaneous improvement across task types, gradual enough to see coming.
- 93% of Business and Financial Operations tasks have meaningful AI time-savings potential. This field is not insulated.
- Current AI success rate on finance and accounting tasks: ~57% with no edits required, improving 8–11 percentage points per year.
- By 2029, most text-based professional tasks are projected to hit 80–95% AI success. Near-perfect performance takes longer — giving a runway for high-judgment, low-error-tolerance work.
- Build the workflow muscle now. Firms integrating AI today will adapt more smoothly than those who wait until the tide is already high.
Want the CPE credit? Take the full lesson on EverydayCPE and earn 0.2 CPE credits: [lesson link]
Leave a Reply