Nearly half of small employer firms now use AI in some part of the business, according to the Federal Reserve’s 2025 Small Business Credit Survey. If that number surprises you into thinking you’re behind, look closer: only 7% have fully integrated it into operations. Forty-nine percent are still experimenting. The real gap isn’t adoption. It’s knowing which specific task in front of you is actually safe to hand over, and which one only looks that way.
The Research Says It’s Not About Difficulty
The instinct is to sort tasks by how hard they seem. A Harvard Business School and Boston Consulting Group field experiment, run with 758 consultants, found that instinct is wrong. On tasks squarely inside what researchers call AI’s “jagged frontier,” AI-assisted consultants completed 12.2% more tasks, 25.1% faster, at 40% higher quality than those working unaided. But on one task deliberately chosen to sit just outside that frontier, unaided humans got the right answer 84% of the time. AI-assisted humans got it right only 60 to 70% of the time. The assistance made people worse, not better, and most of them couldn’t tell.
The frontier isn’t shaped like “easy versus hard.” It’s jagged. A task that looks simple to a person can sit just outside it, and a task that looks intimidating can sit well inside. Sorting by gut feel is exactly the approach the research says fails.
Why Most AI Pilots Go Nowhere
MIT’s NANDA initiative reviewed more than 300 public AI deployments and surveyed 153 executives in 2025. Their finding: 95% of generative AI pilots showed no measurable impact on profit or loss. The 5% that did work shared a pattern: they weren’t flashy, customer-facing, or judgment-heavy. They automated defined, structured, repeatable back-office tasks: document retrieval, first-draft generation, data formatting.
The Gap Isn’t the One You Think
Census Bureau data shows AI adoption scales with firm size: 37% for firms with 250-plus employees versus 32% for firms with 100 to 249. If you’re smaller than that, you are, on average, behind the largest companies. But researchers at METR have started measuring something more useful than adoption: how long a task can run before AI’s reliability drops below 50%. That “time horizon” is real and it’s growing, roughly doubling every four to seven months. It is also, right now, still short. The frontier is moving. It has not arrived. Which means the smaller-firm gap in the Census data is less “falling behind” and more “correctly waiting for the ground to stop shifting”, as long as you’re using the tasks that already work while you wait.
What Actually Works
Morgan Stanley’s internal AI assistant searches roughly 350,000 pages of proprietary research and answers advisor questions in seconds instead of the 30-plus minutes it used to take. More than 98% of advisor teams use it. But the task is bounded: retrieval and first-draft summaries. A human advisor still reviews and sends anything that reaches a client.
Moderna built 750 internal AI tools in two months, including one that drafts a documented dosing rationale from clinical-trial data. The company is explicit that a human team still makes the final call on every one. Its contract-summarization tool hit 100% adoption in the legal department, for summarizing, not for deciding.
Neither of these is a story about AI making decisions. They’re stories about AI doing retrieval and drafting on well-defined material, with a human positioned exactly where judgment is required.
What Doesn’t Work, and Why, Specifically
Commonwealth Bank of Australia cut 45 customer-service roles in 2025, citing an AI voice bot that supposedly reduced call volume by 2,000 a week. The claim was false: volumes were rising, and staff were quietly absorbing the gap through overtime. CBA reversed the cuts within weeks and admitted the business case was flawed. The failure: measuring the wrong thing, the same trap Klarna fell into with its own AI rollback.
McDonald’s ended a three-year, 100-plus-restaurant AI drive-thru pilot with IBM in 2024 after viral videos showed the system mishearing orders: 260 chicken nuggets, bacon on ice cream. Different failure mode entirely: real-time interpretation of unpredictable human speech, a task nowhere near as bounded as it looks from a spec sheet.
Zillow shut down its algorithmic home-buying arm in 2021 after its pricing model overvalued roughly 7,000 homes as the market shifted faster than the model could adapt, a $500 million-plus loss and a quarter of the company’s workforce cut. High stakes, low reversibility, conditions that changed faster than the system could learn, the worst combination for handing something to AI.
And in 2025, multiple lawyers were sanctioned in the US for filing briefs with AI-fabricated case citations: in one, 12 of 19 citations didn’t exist. The tasks looked like ordinary legal drafting. What actually broke was verification: AI is fluent and confident on tasks requiring authoritative, checkable facts, and it fails silently when nobody checks.
The Actual Decision Framework
A 2025 Harvard Business Review framework from researchers at NYU Stern and Harvard Business School sorts tasks on two axes instead of one: how costly an error would be, and whether the task relies on explicit, structured knowledge or tacit judgment. That produces four honest categories instead of a gut call:
No Regrets: hand it over fully. Low cost of error, explicit knowledge. Data formatting, routine scheduling, first-pass document retrieval.
Quality Control: AI drafts, a human verifies. Higher cost of error, still explicit knowledge. Contract summaries, financial report drafts, the exact zone Morgan Stanley and Moderna operate in.
Creative Catalyst: AI generates options, a human picks. Low cost of error, tacit knowledge. Brainstorming, first-draft marketing copy, ideas you were always going to edit anyway.
Human-First: AI assists at most. High cost of error, tacit knowledge. Hiring decisions, performance feedback, pricing changes in a shifting market, anything where being wrong is expensive and there’s no clean rule for right.
MIT Sloan Management Review’s Benjamin Laker puts the same idea in five words worth keeping on a wall: automate tasks, not trust. That’s the actual test. Not whether a task looks technical enough for AI, or important enough to keep human. Whether getting it wrong is cheap or expensive, and whether the knowledge required can be written down or has to be judged. Sort your next task on those two lines before you sort it by instinct.
This is the Builder-stage version of a familiar problem. The same discipline that tells a founder which tasks to hand to a first hire, cost of getting it wrong, and whether the judgment involved can actually be written down, applies just as directly to AI. The mistake is not automating too much or too little. It is making that call by instinct instead of by the two questions above, for every task, every time.
So here is the real question: pick one task you handed to AI without really deciding to. Which line does it actually belong on, and would you make the same call again if you looked at it properly?