Your AI Isn't Underperforming. Your Org Chart Is.

McKinsey surveyed roughly 1,500 executives across 25 organizational traits looking for what actually drives AI ROI. It wasn't the model. It wasn't the budget. It was who does which task — and about four out of five firms haven't touched that. [1]

I read that number twice because it's the same conversation I have with every SMB owner who tells me their AI spend "isn't paying off yet." It's never the tool. It's that nobody redrew the lines around it.

What McKinsey actually found

McKinsey's people-and-organizational-performance team tested 25 traits against self-reported performance gains — things like leadership behavior, decision rights, training investment, incentive design. Out of all 25, the one with the strongest link to AI paying off was how work gets split between people and algorithms: who owns the first draft, who checks it, who's accountable when it's wrong. [1]

Not "do you have an AI policy." Not "how many seats did you buy." Whether you actually redrew the job.

And per the same reporting, only about one in five firms have made that kind of change. Everyone else bought the tool and left the process exactly where it was — same person doing the same job, now with a chat window open in another tab. That's not automation. That's a slower version of the old workflow with an extra step.

The Fed just published the boring truth

If you want the version without the consulting-deck gloss, the Federal Reserve Bank of St. Louis ran the numbers a different way. Researchers Serdar Ozkan, Aakash Kalyani, and Nicholas Sullivan read roughly 490,000 earnings-call transcripts from 5,198 public companies going back to 2000, and tracked what executives actually say about AI and productivity, sentence by sentence. [2]

The share of productivity commentary that mentions AI at all was near zero before ChatGPT, climbed through 2023, sat flat through 2024, then jumped again in 2025 to about 15% of the productivity conversation. But of the sentences that specifically link AI to a productivity gain, roughly 95% describe a gain the company expects — not one it measured. That ratio hasn't moved since 2023. Executives have been promising the same future for three years running. [2]

Meanwhile the aggregate number — utilization-adjusted total factor productivity, the metric that's supposed to catch a real economy-wide gain — grew 0.07% over the four quarters ending Q1 2026. Statistically indistinguishable from nothing. [2]

Two independent reads, same conclusion: the payoff everyone's promising isn't showing up in the data yet. That's not a reason to write off AI. It's a reason to ask what the 1-in-5 companies did differently.

The lag has a name, and it's not an excuse

Stanford's Erik Brynjolfsson has an answer for why the gap exists, and he's been making the case since earlier this year: general-purpose technologies always dip before they pay off. He calls it the productivity J-curve — a period where a company is spending on the new technology, retraining people, and rebuilding processes, all of which shows up as cost before any of it shows up as output. Brynjolfsson argues the US economy is finally exiting that trough, pointing to 2025 productivity growth of roughly 2.7%, nearly double the prior decade's 1.4% average. [3]

I don't need the aggregate number to be right to believe the shape of the curve. I've watched it happen client by client. The dip is real. Every business I've helped restructure looked worse on paper in month one than it did the month before — more meetings, more documentation, a process that used to run on habit now written down and argued over.

But the J-curve explains the lag. It doesn't excuse sitting in it. The 1-in-5 companies McKinsey found aren't the ones who bought a better model. They're the ones who did the restructuring work while everyone else waited for the tool to do it for them.

What redesigning the split actually looks like

For a five-person team, this isn't a reorg. It's smaller and more specific than that. It means picking one task that currently has a human doing all of it — drafting client emails, reconciling invoices, writing weekly reports — and moving the first pass fully onto AI, with a human reviewing instead of doing. Then rewriting the process doc so it describes the new job, not the old one with a note stapled to it.

The tell that you haven't done this: your AI subscription line item went up, and your process documentation didn't change at all. That's the giveaway every time. The tool got added. The job didn't get redrawn.

flowchart TD
  Start([You bought AI tools]) --> Q1{Did you redraw<br/>who does which task?}
  Q1 -->|No| Trough[Investment-phase trough:<br/>tool cost up, output flat,<br/>"AI doesn't work here"]
  Q1 -->|Yes| Q2{Did you rewrite the<br/>process doc to match?}
  Q2 -->|No| Trough
  Q2 -->|Yes| Harvest[Harvest phase:<br/>measured gain — the trait<br/>McKinsey found strongest]
  Trough -.->|most SMBs stop here| Stuck[Still budgeting AI<br/>as a line item, not a role]
  Harvest -.->|repeat per workflow| Q1

That's the same decision I walk through on every engagement: not "which AI tool," but "whose job changes, and did we write that down."

Investment phase vs. harvest phase, side by side

A cross-functional team working through who owns what — the conversation McKinsey found matters more than which model you bought. (Photo: Amtec Photos, via Wikimedia Commons, CC BY 2.0.)

Where this bites you if you skip it

The failure mode isn't dramatic. Nobody's AI rollout blows up. It just quietly costs more every month while nothing measurably improves, and six months in someone asks why you're still paying for three AI subscriptions when the reports still take as long to write as they did last year. That's not a tool problem. Business process automation is exactly this: finding the workflow, mapping who does what today, and rebuilding it so the split is deliberate instead of inherited. The Hub's stack-reduction engagement is the same move applied to the SaaS bill instead of the task list — the software cost didn't drop because we picked cheaper software. It dropped because we redrew who — and what — was actually doing the job.

The tool was never the bottleneck. The org chart it's plugged into is.

So what do you do with this

Stop asking whether your AI subscription is worth it. That's the wrong question — it's asking the tool to justify itself instead of asking whether you did the work around it. Pick one recurring task. Move the first pass onto AI, fully — not a suggestion box next to the old process, an actual handoff. Rewrite the doc so it describes the new job. Then check the number that changed, not the number you hoped would change. If nothing did, you didn't redesign anything. You bought a tool.

I wrote about what heavy AI users are already reporting a few months back, and the pattern holds here too: the operators pulling ahead aren't using a smarter model. They're the ones who actually touched the job description. If you want a second set of eyes on which task to move first, book a 15-minute call — that's the diagnosis I run before anyone talks about which tool to buy.

Sources

[1] McKinsey Insights — Is the AI productivity story at a turning point? — mckinsey.com

[2] Federal Reserve Bank of St. Louis (Serdar Ozkan, Aakash Kalyani, Nicholas Sullivan) — AI and Productivity: What Firms Are Saying on Earnings Calls — stlouisfed.org

[3] The Decoder — Stanford's Brynjolfsson sees AI boosting US productivity, but he also co-founded an AI consulting firm — the-decoder.com

The short version

• McKinsey tested 25 organizational traits against AI ROI. The strongest correlate wasn't the model or the budget — it was whether a company redrew who does which task between people and AI.
• Only about 1 in 5 firms have actually made that change. Everyone else bought the tool and left the job description alone.
• The St. Louis Fed read 490,000 earnings calls and found AI productivity talk is still ~95% future tense — unchanged since 2023 — while aggregate productivity growth sits near zero.
• Stanford's Brynjolfsson calls the gap the "productivity J-curve": real technologies dip before they pay off. True, but it's a description of the restructuring work, not an excuse to skip it.
• At SMB scale, redesigning the split means moving one task's first pass fully onto AI and rewriting the process doc to match — not adding a chat window next to the old workflow.
• If your AI bill went up and your process docs didn't change, you bought a tool. You didn't redesign anything.