What's new in AI for work: the features that actually accelerate productivity

2026 was the year AI at work stopped being a survey question and became telemetry. Four substantial datasets landed within months of each other — and read side by side, they draw an unusually clear line between the features that accelerate real work and the ones that just generate activity.

1. Delegation beats conversation

OpenAI's enterprise report From assistance to execution (12 August 2026) is the clearest signal. Among enterprise customers in June 2026, its agentic coding surface accounted for 64% of combined output tokens — the majority of enterprise AI output is now delegated execution, not chat. The growth is sharpest outside engineering: weekly active use since February 2026 rose 108x in legal, 41x in sales and recruiting, 26x in marketing, against 5x in engineering.

The same report finds the top decile of firms produce 8.3x more output per active user than typical firms — a gap that widened from 2.6x in January. Those firms use plugins and skills roughly twice as often. The differentiator is not model access. Everyone has that. It is whether the AI is wired to do things.

2. Agents are scaling; agency is the constraint

Microsoft's 2026 Work Trend Index: Agents, Human Agency, and Opportunity (5 May 2026), drawing on 20,000 workers across ten countries plus Microsoft 365 telemetry, reports active agents in M365 growing 15x year over year (18x in large enterprises). Around one in five workers now qualify as heavy AI operators, and 58% of AI users say they produce work they could not have produced a year ago — rising to 80% among the heaviest users.

The finding leaders should sit with: organisational factors — culture, manager support, workflow design — explain roughly 67% of AI's measured impact, against 32% for individual behaviour. AI productivity is an operating-model outcome, not a personal-skill outcome.

3. The context tax is real and measurable

Glean's Work AI Index 2026 (10 June 2026), a survey of 6,000 full-time digital workers across the US, UK and Australia, quantifies the other side of the ledger. Workers report AI saving about 11 hours a week — while spending 6.4 hours a week "botsitting": feeding context, checking outputs, fixing mistakes, switching tools. That is 37% of AI time. Despite 87% regular usage, only 13% say their organisation performs significantly better.

Botsitting is the symptom of ungrounded AI. Every hour a human spends supplying context is an hour the system should have retrieved for itself.

4. Measure production, distrust perception

Two studies should be read as a pair. METR's randomised trial (July 2025) found experienced open-source developers were 19% slower on AI-allowed tasks — while believing they had been 20% faster. Perception is not a metric. Meanwhile DX's Q1 2026 impact report, covering 400+ companies, measures actual telemetry: an average 3.6 hours saved per developer per week, 4.1 for daily users, with daily users shipping roughly 60% higher pull-request throughput.

And the effect is not evenly distributed. Brynjolfsson, Li and Raymond's study of 5,172 customer-support agents (Quarterly Journal of Economics, 2025) found a 15% average lift in issues resolved per hour — 34% for novices, near zero for top performers. AI compresses the experience curve; it does not multiply your best people.

What actually accelerates work

• Grounding in institutional data. It removes the 6.4-hour context tax rather than shifting it onto staff.
• Tool access and execution rights. The 8.3x gap belongs to firms whose AI can act, not just answer.
• Workflow placement over tool distribution. Two-thirds of measured impact traces to how work is organised around the system.
• Production telemetry over self-report. The 19%-slower study is the cautionary tale for every dashboard built on perceived time saved.

Deliberate partnerships, measurable outcomes. The evidence now says the same thing we do: proven against your numbers, not a demo script.

Sources

• OpenAI, From assistance to execution: How enterprises put AI to work, 12 August 2026 — openai.com
• Microsoft, 2026 Work Trend Index Annual Report, 5 May 2026 — microsoft.com/worklab
• Glean Work AI Institute, The Work AI Index 2026, 10 June 2026 — glean.com
• METR, Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, 10 July 2025 — metr.org
• DX, AI-assisted engineering: Q1 impact report, 2026 — getdx.com
• Brynjolfsson, Li & Raymond, Generative AI at Work, Quarterly Journal of Economics 140(2), 2025 — academic.oup.com