KOMETA MEDIA

AI & Work6 min read

AI vs. Humans: Who Is Better at What?

Studies show where AI makes work faster and better, where it makes mistakes, and why the combination usually wins.

A human hand and a glass robot hand reaching for the same chess piece

Talk about AI and you usually hear two camps. One expects machines to do everything better soon. The other considers AI an expensive toy. Neither helps with a concrete decision in your business. The more useful question is: for which task does AI help, for which does it do harm, and where is the combination of human and machine strongest? There are now solid studies on this. Here is what they show, and what they do not.

The starting point: according to Eurostat, around 20 percent of EU enterprises with at least ten employees used AI technologies in 2025, up from 13.5 percent a year earlier [6]. Among small enterprises the share was 17 percent, among large ones 55 percent [7]. Many businesses are still asking where to begin.

01

Where AI is already better

The best-known study comes from researchers at Harvard Business School and other universities, carried out with the consulting firm BCG. 758 consultants worked on realistic tasks, some with access to GPT-4, some without. On 18 tasks that were within AI's capabilities, the AI users completed 12.2 percent more tasks on average, and did so 25.1 percent faster. Their results were rated more than 40 percent higher in quality [1].

25.1%

faster: how much quicker consultants completed tasks with AI support when the task was within AI's capabilities (GPT-4, 2023) Source [1]

A second example comes from customer service. Researchers around Erik Brynjolfsson analysed data on 5,179 support agents. With an AI assistant they resolved 14 percent more issues per hour on average. For novice and low-skilled workers the gain was 34 percent, while the effect on experienced workers was minimal [2]. AI mainly lifts the lower and middle of the range. In the BCG study, too, weaker participants improved by 43 percent, stronger ones by 17 percent [1].

Then there is time. Stanford's AI Index reports that with a short time budget of two hours, top AI systems score about four times higher than human experts [5]. For drafts, summaries and first-pass analysis, that means the quick result increasingly comes from the machine.

A jagged line of light between glowing and dark glass cubes
02

Where AI is worse

The same BCG study included a task deliberately placed outside AI's capabilities. Here, participants using AI solved it correctly less often: 19 percentage points less often than the group without AI [1]. The authors speak of a "jagged technological frontier": some tasks the AI handles easily, while others that seem similarly difficult are outside its capabilities [1]. According to the authors, those who performed worse tended to adopt the AI's answer without checking it [1].

19 percentage points

fewer correct solutions with AI than without, on a task outside AI's capabilities Source [1]

The 2025 AI Index names a basic weakness: these systems still cannot reliably solve problems for which provably correct solutions can be found. This affects their trustworthiness and their suitability for high-risk applications [5]. The picture also flips with time: with long budgets of 32 hours, human experts score twice as high as the AI in the same comparison [5].

Responsibility also stays with the company. In Moffatt v. Air Canada, the airline's chatbot gave a customer incorrect information about bereavement fares. Air Canada argued that the chatbot was a separate legal entity responsible for its own actions. The Canadian tribunal did not accept this, treated the chatbot as part of the website, and awarded the customer damages [8]. This is a decision under Canadian law, not a German ruling, but the signal is clear: what your chatbot says, you say.

03

Gut feeling misleads

How hard it is to judge the benefit in daily work is shown by a study from the organisation METR. 16 experienced developers worked on 246 real tasks in their own open-source projects, sometimes with AI, sometimes without. With AI, tasks took 19 percent longer. Beforehand the participants had expected a 24 percent speedup, and afterwards they still believed AI had made them 20 percent faster [3].

19%

longer task time with AI for experienced developers who believed they were 20% faster (small study, early 2025) Source [3]

Some context: the sample was small, and METR redesigned its study in February 2026. METR calls the data of a follow-up study an unreliable signal, among other reasons because many developers did not want to work without AI and therefore did not submit tasks. METR considers it likely that the benefit is larger now than in early 2025 [4]. So the lesson is not "AI slows you down" but: your own impression is no yardstick. If you want to know whether AI helps in your workflow, you have to measure it.

A person and a figure of light in front of the same screen
04

Best together

The BCG study showed two successful ways of working. "Centaurs" divide the work and decide which steps they take themselves and which they hand to the AI. "Cyborgs" interleave their workflow closely with the AI and work with it continuously [1]. What both have in common is that the human judges where the frontier lies. The authors stress that on tasks outside the frontier, professionals must question the AI and keep applying their own judgment [1].

The authors also note that on tasks inside the frontier, AI levels out performance differences between staff while raising the quality of the work. They add, however, that the best approaches to working with AI are not yet fully understood and need to be examined further [1]. For your business that means there is no ready-made recipe to copy. There are tasks where a trial is worthwhile, and a way to evaluate that trial properly.

05

What this means in practice

  • Break work into tasks and check, task by task, whether it suits AI. Drafts, summaries and structured routine are good candidates [1][2].
  • Require a human check of every AI output, especially for figures, legal matters and customer information [1][8].
  • Test small and measure time, quality and error rate before and after. Do not rely on gut feeling [3].
  • Involve less experienced staff in particular; the effect was largest for them in the studies [1][2].
Sources (8)
  1. [1]Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality (Working Paper 24-013)Harvard Business School with Boston Consulting Group (Dell'Acqua, McFowland, Mollick, Lifshitz-Assaf, Kellogg et al.) · September 22, 2023
  2. [2]Generative AI at Work (NBER Working Paper 31161)National Bureau of Economic Research (Brynjolfsson, Li, Raymond) · April 1, 2023
  3. [3]Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer ProductivityMETR (Becker, Rush, Barnes, Rein) · July 10, 2025
  4. [4]We are Changing our Developer Productivity Experiment DesignMETR · February 24, 2026
  5. [5]AI Index Report 2025, chapter Technical PerformanceStanford Institute for Human-Centered AI (HAI) · January 1, 2025
  6. [6]20% of EU enterprises use AI technologiesEurostat · December 11, 2025
  7. [7]Use of artificial intelligence in enterprisesEurostat, Statistics Explained · December 1, 2025
  8. [8]Moffatt v Air Canada: Bot Betrayal and Negligent Misrepresentation (on Moffatt v. Air Canada, 2024 BCCRT 149)WeirFoulds LLP · March 6, 2024

How much of this is in your day-to-day?

In a free initial consultation we look together at what AI can already do in your business today – and where it has no place.

Book an initial consultation

More posts