Anthropic Mapped 207,049 Research Sessions. AI Is Doing the Wrong Work.
📖 Glossary
AI Box (also known as Agent Computer / Agent PC), is a dedicated local hardware device that runs AI Agents. Pre-installed with an AI agent management system, plug-and-play, running 24/7. Users can remotely command AI to work via Discord, Slack, Telegram, WhatsApp, and more.
On November 5, Anthropic's economics team published a 35-page report with a deliberately plain title: AI for scientific discovery: How researchers use Claude.
The method is straightforward, and the fact that nobody had done it systematically before is the point. Anthropic took 207,049 research-related sessions from consumer plans, covering August 2 to September 30, 2026, labeled every one of them using the SciNet taxonomy — 7,262 tasks across 34 fields and 318 subfields — and layered on top a survey of 528 working scientists.
Some of the headline numbers are arresting. If current capabilities were adopted broadly, the report estimates AI could take over 27% to 52% of the researcher time a new study requires, with a median of 38%. Converted into research output under the same budget, that implies 2.02x in statistics, 1.96x in computer science, and 1.41x in chemistry.
If the report stopped there, it would not deserve a long read. The valuable part is in the comparison between one column of numbers and another.
A division of labor that doesn't line up
Set the researcher's actual time against the work AI actually picks up:
Data analysis: 27.6% of researcher time, 33.7% of Claude sessions, 48.1% of Claude Code sessions.
Writing and communication: 16.0% of researcher time, 32.0% of Claude sessions.
Data gathering: 28.6% of researcher time, but only 11.5% of Claude sessions and 11.9% of Claude Code sessions.
Ideation and hypothesis generation: 7.1% of researcher time, 4.2% of Claude sessions, 2.1% of Claude Code sessions.
The two things scientists themselves rank as most important for producing meaningful work — generating ideas and gathering data — are precisely the two where AI is used least. And the work AI takes most of, analysis and writing, is exactly the part that was already digital.
The 528-person survey agrees. Asked what matters most for producing important research, scientists put ideation first and data collection second. Anthropic offers two explanations. Gathering data mostly requires hands on a physical sample, and AI cannot reach those. Ideation requires no hands at all, but researchers may simply prefer to think for themselves, and may not trust AI to pick the right problem. There is also a measurement artifact: ideas get woven into other tasks, which makes them easy to undercount in classification.
So the report is not really saying "AI has become strong." It is saying "here is where AI's capability boundary currently sits." Those sound similar and are not the same claim.
Scientists voted with their usage

Anthropic also scored every research task for how well AI could plausibly perform it, using something close to an Elo-style head-to-head: take two research steps, see which one AI handles more comfortably.
The result is unusually clean. Tasks in the top 20% by capability — data visualisation, statistical tests, code debugging — see usage roughly 2.1 times what their share of researcher time would predict on Claude, and 3.9 times on Claude Code. Tasks in the bottom 20% — purifying mixtures in a lab, fieldwork, animal studies — sit at a usage ratio of about 0.2.
The correlations line up with the mechanism: task capability correlates 0.78 with suitability for reinforcement learning and 0.69 with cognitive work, and negatively with physical work at -0.57.
Translated: the better a task can be broken into a drill with rapid feedback, the better AI does. The more it requires hands and results that are hard to verify, the worse AI does. Scientists are voting with their usage, and they are voting along the capability boundary rather than along preference.
The gap between disciplines
By field, the distribution skews hard.
Computer science is 4.3% of the world's researchers but 32.7% of Claude Code research sessions. Artificial intelligence and machine learning is 1.2% of researchers and 15.2% of Claude Code sessions. Chemists and clinical doctors, who live in labs and clinics, use it much less.
The other half of that story is physical work. Agricultural, biomedical, chemical and materials sciences spend 38% to 41% of researcher time on steps that require hands, against a median of 4% across fields.
Which is why the three scenarios diverge so sharply.
Scenario one, current capabilities adopted broadly: AI takes 27% to 52% of research time, and output acceleration runs from 1.35x in the arts, 1.36x in anthropology and 1.41x in chemistry up to 2.02x in statistics and 1.96x in computer science.
Scenario two, 75% of non-physical steps automated: computer science reaches roughly 8x, statistics and library science around 7x, economics around 6x, while lab-based and medical sciences manage 2x to 3x.
Scenario three, every non-physical step automated: statistics, computer science, economics, mathematics, psychology and most humanities and social sciences reach 15x to 16x — while biology, chemistry, materials, agriculture and biomedical sciences stay stuck at 3x to 4x.
Why does the ceiling sit at 16x? The report's cost assumption: AI agents at $30 per hour, researchers at $95 per hour. Fully automated tasks cost about one-sixteenth of a human, so acceleration stops there. Halve the cost of AI and the ceiling doubles.
Anthropic states repeatedly that these are illustrative scenarios, not forecasts.
What the report does not tell you
Read the sample boundary carefully. All 207,049 sessions come from consumer plans; enterprise and API usage is absent from every figure. The 528 survey respondents were recruited mainly from applicants to Anthropic's own AI for Science grant programmes, and Anthropic concedes they are likely heavier AI users than typical scientists. Classification labels come from a model, and the SciNet time benchmarks are themselves estimates.
So this table describes, quite precisely, how individuals and small teams use consumer AI tooling for research. Which happens to be the layer closest to most people.
What follows from that

Three conclusions follow if you push this one step further.
First, the first wave of AI in research is not about replacing scientific thinking. It is about keeping the already-digital parts of the pipeline running continuously. Analysis, writing, code reproduction, literature handling — these are where AI is most reliable and least fussy.
Second, the bottleneck was never the model score. It is compute, data, and time spent present. The memory-architecture work that had to design a dedicated consolidation mechanism for 350,000 files and 420 million tokens did so precisely because that pipeline needs idle time and sustained compute. And the 48.1% data-analysis share on Claude Code holds only if someone keeps it running. Data works the same way: if the data isn't yours, AI can only process what you feed it by hand, and the 28.6% spent on data gathering never connects.
Third, whether "AI takes half your time" materialises depends on whether the work runs on your machine or waits in someone else's queue.
Cloud APIs carry three bills you cannot avoid: you pay per call, you wait per request, and you make a compliance judgement every time data leaves. For individual researchers and small teams these costs scale linearly with usage — and research is exactly the kind of work that, once started, does not stop.
KAIHE AIBOX takes the other path. It is a self-hosted agent computer: the compute is yours, the data stays local, it runs 24/7, and OpenClaw and Hermes are preinstalled and waiting. When multiple agents collaborate, one analysis task splits across several roles in parallel, and the results settle into a local knowledge base so the next run picks up where the last one stopped. It replaces the default assumption that AI is a one-shot reply with something closer to a colleague that keeps running, keeps track, and can be woken up at any moment.
What it cannot do
Being straight about this matters more than the marketing.
The hardest ceiling in the report — the 3x to 4x ceiling for lab sciences — has nothing to do with which hardware you use. The flask still needs heating, the pipette still needs a human hand, and that 40% of physical work exists for everyone. KAIHE AIBOX does not conjure a mass spectrometer, and it does not make an AI-generated hypothesis true.
What it can address is the other 60%: whether the parts that are already digital can keep running without interruption, at no marginal cost, with nothing leaving your premises.
One detail in the report deserves rereading. Tasks in the bottom 20% see a usage ratio of roughly 0.2. That suggests the smarter approach is not to insert AI into every step, but to hand part of the judgement about which steps belong to AI: state the goal, state the boundaries, and let it decide how to remember and how to organise.
That is the point where the two Anthropic reports meet. One describes how memory should be designed. The other describes where capability ends. Both land on the same conclusion: what decides whether AI can keep working is not how accurate a single answer is, but whether it has a stable place to be present, with enough compute, and with your data in hand.
That place can be the cloud. It can also be a machine you own.