What the Technology Actually Returns

Every AI investment is a bridge with two towers. One tower is the technology: what AI does to the work. The other is your people: whether your workforce can actually put it to use. Two towers do not make a bridge. What carries the value across the gap between them is the architecture: the workflow redesign, governance, and adoption that turn a capable tool into a result on the income statement. This series takes each tower in turn, then the span that connects them. This part is about the first tower, the technology.

Most leaders I work with have already been handed a number. A vendor deck promised a 40% productivity gain, or a respected peer quoted a figure at a conference that sounded close enough to true. The number goes into a board slide. Then it sits there, waiting for someone to ask where it came from.

When that question comes, most AI business cases fall apart. Not because AI does nothing. Because the number was never sourced, and a number you cannot defend is worse than no number at all.

So let me start with what the technology actually returns, according to the strongest evidence we have rather than the brochure.

The technology works. The controlled studies are clear.

When researchers run real experiments, with random assignment and measured output, the productivity gains from AI are genuine and, in places, large.

Across three field experiments at Microsoft, Accenture, and a Fortune 100 firm, developers using an AI coding assistant completed 26% more tasks.¹ Among customer-support agents, AI assistance raised issues resolved per hour by 15% on average, and 30% for the least experienced agents.² Professional writers finished tasks about 40% faster, and the weakest writers improved the most.³ Management consultants working inside AI's current capability produced work rated substantially higher, while doing measurably worse on tasks that sat outside that capability.⁴

Two things run through all of it. The gains are real. And they concentrate at the bottom: AI narrows the distance between an organization's strongest and weakest performers. That second finding matters more than the first, and I will come back to it, because it quietly turns AI from a technology story into a people story.

The honest ceiling

Here is the number the vendor decks leave out.

Of the organizations investing in AI, only 39% report any impact on earnings at the enterprise level, and roughly 6% attribute more than five percent of profit to it.⁵ A separate survey found only 3% of companies report a return in the 10–20% range; most land between 1 and 5%.⁶

Hold those two pictures side by side. Controlled studies show double-digit gains on individual tasks. Company-level results show a thin slice of firms capturing real profit. The distance between the lab and the P&L (the profit-and-loss statement) is the whole story of AI right now.

Where the return leaks out

That distance has a shape, and it is measurable.

Nearly two-thirds of organizations have not begun scaling AI across the enterprise.⁷ Pilots run, demos impress, and then the work stalls before it reaches the income statement. One widely cited and openly debated analysis put the share of enterprise generative-AI pilots with no measurable profit impact at roughly 95%.⁸ Treat that figure as directional rather than precise; the methodology is contested, but the direction is not. Most pilots stall before they pay.

The failure mode is not model quality. It is the integration gap: the distance between a tool that can do a task and an organization redesigned so that it does. That distance is architecture, and it is the subject of Part 3.

There is a hopeful number hiding in the same research. Top-performing mid-market firms scaled a working pilot in about 90 days, while large enterprises took nine months or more.⁹ Smaller companies are often more able to close the lab-to-P&L gap, not less. The architecture is more buildable when fewer layers sit between the decision and the desk. For a 100-person company, that is the advantage worth pressing.

Why the technology tower cannot stand alone

Return to the finding I flagged. AI compresses the skill distribution. The least experienced support agent gains 30%. The weakest writer improves most. The developer's gain depends on the developer.

That means the return on AI is really a return on how well your people take it up. The same tool placed in two organizations produces two different numbers, and the difference is the workforce and the managers guiding them. That is the people tower, and it is Part 2 of this series.

None of these productivity gains land on the income statement on their own. Someone has to redesign the workflow around the tool, decide what good looks like, and bring the team with them. That work is leadership work, and it happens before the technology does anything.

See it in your own numbers. The Living ROI Calculator at changeadvisor.net models both towers and the span between them from your inputs: the value as typically realized, and the value with the architecture in place. For a 100-person mid-market company investing about one fully loaded salary in AI, it puts the realized figure near $157,500 a year and roughly double that, about $315,000, once the architecture is built. Add the people tower and the supported total reaches roughly $1.06M. Every figure traces to a graded, published source. Nothing is stored on the server.

The honest number is a range

The technology returns more than the skeptics think and less than the vendors promise. Both are true, and living with both is where an honest AI strategy begins.

The number in your board deck should be a range, and it should carry its sources. When it does, it stops being a claim you have to defend and becomes a conversation you can actually have.

The next question is the one most AI strategies never ask out loud: what are your people worth to that return? That is Part 2.

---

What Actually Returns is a three-part series.
Part 1 — The technology tower: what AI actually returns. (you are here)
Part 2 — The people tower: the most undermeasured number in business.
Part 3 — The span: the architecture that closes the gap.
Run your own numbers at changeadvisor.net. Not sure where your organization stands first? Start with the AI Readiness Assessment.

References

1. Cui, Zheng, et al. "The Effects of Generative AI on Software Developer Productivity." Field experiments at Microsoft, Accenture, and a Fortune 100 firm (n≈4,867 developers), 2024 (revised 2025). [T1 · via Change Advisor evidence ledger, changeadvisor.net]
2. Brynjolfsson, Erik, Danielle Li, and Lindsey Raymond. "Generative AI at Work." Quarterly Journal of Economics, 2025 (NBER working paper 2023). [T1]
3. Noy, Shakked, and Whitney Zhang. "Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence." Science, 2023. [T1]
4. Dell'Acqua, Fabrizio, et al. "Navigating the Jagged Technological Frontier." Harvard Business School working paper, 2023. [T1]
5. McKinsey & Company. "The State of AI." November 2025. [T2]
6. Forbes Research. AI ROI survey, October 2025. [T3 · figure previously mis-attributed to Wharton; corrected July 2026]
7. McKinsey & Company. "The State of AI." November 2025. [T2]
8. MIT NANDA initiative. Enterprise generative-AI pilot analysis, July 2025. [T2 · CONTESTED — methodology debated; robust takeaway is that a large majority of pilots stall before measurable financial impact]
9. MIT NANDA initiative. July 2025. [T2 · mid-market ~90 days vs enterprise 270+ days to scale a successful pilot]