The figure
58%
24 company cases on this site report a result from using AI, not a failure. In 14 of them, every figure comes from the company itself. In the other 10, researchers or an official record measured the result.
A company’s own figure can be true. The evidence gap says who produced the figure, not whether it is wrong. It shows how much of what you read about AI success nobody outside the company has checked.
How it is computed
- Take every company case on this site that reports a result from using AI: a measured result, a company-reported result, or a result the company later partly walked back. These are the success stories, as their sources present them. Today that is 24 cases.
- Count the cases whose figures come only from the company, with no study, court or official record behind them. Today that is 14.
- Divide the second number by the first and round to a whole percentage: 14 ÷ 24 = 58%.
Cases where something went wrong are left out, because they report a failure, not a result. Today that is 15 of the 39 cases.
Each case page states how the case turned out and who produced its figures (the four labels). The figure on this page is recalculated from those labels every time the site is updated.
The cases behind it
Figures from the company itself: 14 cases
- Ashurst: first drafts 45–80% faster in timed trials, with lawyers still ahead on accuracy (Company-reported result)
- Commonwealth Bank of Australia cut 45 roles for a voice bot, then said they were not redundant (Partly walked back)
- Duolingo: 148 new language courses in about a year, after 12 years for the first 100 (Company-reported result)
- EIST: a six-person metalworking shop cut quoting time from two hours to ten minutes (Company-reported result)
- Gilberts: a smaller audit firm started with the AI in its existing tools and two firm rules (Company-reported result)
- Klarna: AI cut spending on outside marketing suppliers by a quarter (Company-reported result)
- Klarna: an AI assistant took two-thirds of support chats, then the company hired people again (Partly walked back)
- META-Regalbau: sensors and AI analysis of order picking paid back in about a year (Company-reported result)
- Milkforbubs: a small online shop put a ready-made chatbot on customer questions (Company-reported result)
- Ministry of Clean: a cleaning company moved its customers from spreadsheets to a CRM with AI built in (Company-reported result)
- SAM: a two-person advisory firm cut the paperwork on each client file from six hours to two and a half (Company-reported result)
- Temple & Webster: an online furniture retailer with about 200 staff put AI into customer care first (Company-reported result)
- Walmart: generative AI used to fix 850 million pieces of product data (Company-reported result)
- WPP: £300 million a year on an AI platform, and revenue still fell (Company-reported result)
Figures checked outside the company: 10 cases
- eBay: machine translation raised cross-border sales by about 11% (Measured by researchers)
- Garfield.Law: a regulator authorised an AI-run law firm for small debt claims, with limits (Official record)
- Herges: an AI model estimated production times five times more accurately than the firm’s experts (Official record)
- Hitachi Energy: an AI model pointed to the two production steps that were spoiling chips (Measured by researchers)
- HMRC gave 3,000 Copilot licences at random: about an hour a week saved, by staff estimate (Official record)
- Procter & Gamble: one person with AI matched a two-person team without it (Measured by researchers)
- PSG Global Solutions: AI ran the job interviews, and people made the hiring decisions (Measured by researchers)
- Rue La La: machine-learning price recommendations raised revenue by about 10% in a test (Measured by researchers)
- Siemens: inspectors caught more defects when the AI showed what it had seen (Measured by researchers)
- UK Department for Business and Trade: 1,000 Copilot licences, satisfied users, no proven productivity gain (Official record)
Limits
- The set is small: 39 cases, of which 24 report a result. One case more or less moves the figure by several percentage points.
- The editor chose the cases. Each one is a named company with a public source. They are not a random sample.
- The figure describes this set of cases, not all companies that use AI. Do not quote it as the share for any industry, country or company size.
- Most published cases come from large companies. Small firms rarely publish numbers.
- A success story here is a case that reports a result, not a failure. Some of those results are small: researchers or an official record sometimes found no gain at all.
- The figure changes when cases are added or relabelled. Quote it with its date.
How to cite
Quote the definition and the value with its date, and link to this page:
The evidence gap is the share of AI success stories whose figures come only from the company that reports them. In the ai-first.company case set it is 58% (October 2026).
Szczęsny, P. “Evidence gap”. ai-first.company, 18 October 2026. Value: 58%. https://www.ai-first.company/evidence/evidence-gap/
The figure and the case list behind it are open data under CC BY 4.0. You may reuse them if you credit the source: “Data: CC BY 4.0, Paweł Szczęsny, ai-first.company”. The author is Dr Paweł Szczęsny. The case list is also a download: the columns outcome and figures_from give the figure.
History of the value
The value each time the case list changed it.
| Date | Cases reporting a result | From the company only | Evidence gap |
|---|---|---|---|
| 24 | 14 | 58% | |
| 21 | 13 | 62% |