AI and the division of labour: autonomy remains the exception

A new global survey from research and advisory firm ISG offers an unusually concrete picture of how work between humans and AI is actually divided in large enterprises today, and how that split is expected to shift over the coming years. It's a more sober picture than the discussion that often settles for assumptions about AI "taking over" tasks at scale.

How the work is split today

ISG's report "The AI Value Gap", published in September 2026 and based on a global survey of 400 senior decision-makers at large enterprises, splits all AI-enabled work into four categories based on how much a human is involved. 55 percent of work is human-led today. Nearly another quarter is reviewed by a human before it proceeds. Just under 14 percent involves humans only for handling exceptions. Less than 7 percent is performed fully autonomously by AI, with no human involvement in the execution itself.

The expected shift to 2027

Companies in the survey expect the split to keep shifting, but not in the direction an "AI takes over everything" narrative would suggest. By the end of 2027, the human-led share is expected to fall below 40 percent, while the autonomous share is expected to nearly double, to 13 percent. Even two years out, among companies already among the most AI-mature globally, more than 85 percent of work remains human-led or human-reviewed.

Where the value already shows up

That doesn't mean AI lacks value today. More than 40 percent of companies in the survey said AI generated value over the past year through task automation and workflow execution, data analysis, and process optimisation. Confidence holds up too: 69 percent say they're willing to keep increasing their AI budgets, as long as the value keeps coming in at the current rate. The value is already there – it just rarely comes from letting AI act entirely on its own.

What it means for how you build

This matches how we already think about AI agents: let a system read, compile and suggest before giving it the right to act on its own. The report's numbers show that this order isn't a temporary caution on the way to full autonomy. According to the most ambitious companies' own expectations, it's still the reality two years from now. Building a workflow for review rather than autonomy isn't building for less – it's building for how the work will actually look.

Where Nodal comes in

We see the same pattern in the projects we build ourselves: the real gain almost always comes from removing the manual work and making what happens visible and understandable, not from letting a system act unsupervised. A workflow a person can quickly review and trust often beats one that tries to be fully self-running too soon.

Wondering where in your workflows AI should genuinely act on its own, and where it should stay at suggestion and review? Get in touch with us at Nodal, and we'll go through it together.