The AI productivity gap: why brands move faster with purpose-built agents

One statistic dominated boardroom AI conversations for a year: 95% of enterprise generative AI pilots deliver no measurable P&L impact. It came from The GenAI Divide: State of AI in Business 2025, a preliminary report from MIT's Project NANDA led by Aditya Challapally, based on 300+ public implementations, interviews at 52 organisations and a survey of 153 senior leaders.

The number deserves its critics. The data was never released; Wharton's Kevin Werbach publicly questioned where the 95% comes from; the six-month ROI window is short for enterprise change. Treat it as a directional finding, not a law of nature.

But buried under the headline is the detail that actually explains the productivity gap — and it barely made the coverage.

Bought and purpose-built beat built and generic

In the same report, purchased or partner-built specialised tools succeeded roughly 67% of the time, against about one third for internal general-purpose builds. The failure driver was named the "learning gap": tools that do not retain feedback or adapt to the workflow they sit in. Not model quality. Workflow fit.

The second overlooked finding: the largest returns showed up in back-office automation, while more than half of budgets went to sales and marketing. Organisations were spending where AI is most visible rather than where it compounds.

What the gap costs, precisely

Two independent studies price it.

BetterUp Labs and Stanford's Social Media Lab, writing in Harvard Business Review (22 September 2025), surveyed 1,150 US desk workers and named the failure mode "workslop": AI output that looks finished and isn't. 40% had received workslop in the previous month, each instance costing about one hour fifty-six minutes to resolve — an invisible tax of roughly $186 per employee per month, near $9M a year for a 10,000-person organisation.

Glean's Work AI Index 2026 (10 June 2026, 6,000 workers) measured the same leak from the other end: AI saves around 11 hours a week, but workers spend 6.4 hours a week supplying context and correcting output. Despite 87% regular usage, only 13% say their organisation performs significantly better.

Both numbers describe one thing: generic tools push verification onto humans. That is the gap.

What closing it is worth

IDC's global study for Microsoft (The Business Opportunity of AI, November 2024, 4,000+ decision-makers) puts average return at $3.70 per $1 invested, with leaders at $10.30 and value typically realised inside 13 months. Gartner expects 40% of enterprise applications to include task-specific agents by the end of 2026, up from under 5% — the market is converging on narrow, embedded systems for the same reason the MIT data does.

The operator's version

• Buy or build the specific thing; never build the general thing. Two-thirds versus one-third is the widest lever in the data.
• Ground it in your institutional data. Unretrieved context becomes human hours — 6.4 of them a week.
• Start where volume is high and variance is low. Back-office first, however unglamorous.
• Instrument rework as a defect count. If humans quietly fix agent output, your ROI slide is fiction.
• Define the number before the pilot. Handle time, first-contact resolution, days sales outstanding — the 5% start there and work backwards.

The gap between AI that pays and AI that doesn't is not a gap in models. Everyone rents the same frontier. It is a gap between systems built for a workflow and systems bought for a headline.

Sources

• MIT Project NANDA, The GenAI Divide: State of AI in Business 2025 — report PDF; critique via 80,000 Hours
• BetterUp Labs & Stanford Social Media Lab, AI-Generated "Workslop" Is Destroying Productivity, HBR, 22 September 2025 — hbr.org
• Glean Work AI Institute, The Work AI Index 2026, 10 June 2026 — glean.com
• IDC for Microsoft, The Business Opportunity of AI, November 2024 — blogs.microsoft.com
• Gartner press release, 26 August 2025 — gartner.com