Research

The numbers everyone quotes about AI at work, traced back to what they actually measured. Every claim here carries its source, its year and its sample size, and where a figure turns out to be unsupportable it says so.

Topic Evidence

The HR Assistant question

Human Resources Assistants have the second largest "needs software built first" band of any People occupation, at 68 per cent against a Job Zone 3 average of 30. Only 11 per cent of the role is reachable with a chat window today, which means the advice being sold to these people, learn prompting, is aimed at two tasks out of nineteen. It is also one of only two People occupations you can enter without a degree, and an independent Tufts model puts it at the highest projected job loss of the nine. The risk is not a redundancy round. It is a role that quietly stops being hired, which removes the way in rather than the people already through it.

4 sources 2026-08-26 9 min

Topic Evidence

The 70% statistic

The 70% figure traces back to Hammer and Champy in 1993, who described it as an unscientific estimate and later disowned the fixed-rate reading. A peer-reviewed study in 2011 followed the number across five published sources and found no empirical basis in any of them. McKinsey states it without a source, the Harvard Business Review version traces to a magazine column, and BCG's own breakdown shows 30% full success and 44% partial value, which is not a failure rate. If you need a defensible number, McKinsey's 2018 finding that 16% of digital transformations both improved performance and sustained it is the better one.

3 sources 2026-08-23 6 min

Topic Evidence

The 95% AI pilot claim

MIT Project NANDA's State of AI in Business 2025 found that around 5% of custom enterprise AI tools reached production with sustained productivity or P&L impact. That is not the same claim as 95% of GenAI pilots failing, which is how it was reported. The evidence base is 52 executive interviews, 153 survey responses collected at four conferences, and a review of 300 or so publicly disclosed initiatives. The report is not peer reviewed and its method has been publicly challenged. The finding worth your attention is a different one: externally sourced tools reached deployment around 67% of the time against 33% for tools built in house.

4 sources 2026-08-23 7 min

Topic Method

Measuring an AI pilot

A pilot that started without a baseline cannot prove anything, because there is nothing to subtract from. UK government benefits management doctrine, published in the Teal Book, requires that current performance is captured before work starts, that every benefit has a named owner, and that monetisable benefits are separated into cash-releasing and non-cash. Applied to an AI pilot, that means hours released and cash converted are different figures that must never be added together, and anything a vendor claims stays in a third column until a baseline makes it measurable.

3 sources 2026-08-23 8 min

Evidence

What a widely quoted number actually measured, what its sample was, and whether it survives being read properly.

Method

How an established practice works and where it came from. Mostly older than AI, which is the point.

Regulation

What the law requires of People teams using AI, and by when. Dates change and get reported wrongly.

Why this exists

Most writing about AI in HR repeats a handful of statistics that nobody has read the source of. Two of the most quoted turn out to say something narrower than reported, and one cannot be sourced at all. Checking takes an afternoon each, so the checking is published here rather than kept.