Field study
Reducing Friction in Knowledge Work
Before automating anyone's job, it is worth measuring what the job actually is. Mostly, it is friction.
Tomás Carvalho · Priya Raghavan
Ask a financial analyst, a hospital administrator, or a policy researcher what their job is, and they will describe their judgment work: the analysis, the decision, the document only they could write. Watch their working week, and a different job appears. With consent, we instrumented and diaried the weeks of 140 knowledge workers across nine organizations. On average, 43% of working time went to the work participants would describe as their job. The majority went to four kinds of overhead.
The four frictions
- Retrieval (≈18%). Finding the thing: the prior version, the precedent, the person who knows. The modal retrieval episode ended at a substitute for the sought artifact, not the artifact itself.
- Translation (≈14%). Reformatting knowledge that already exists for a new audience or system: the same facts as a slide, a memo, a form, an email. No new judgment, no new information — pure impedance matching between contexts.
- Reconciliation (≈12%). Resolving divergence between copies of what is notionally the same truth: two spreadsheets, two definitions of the same metric, two versions of the plan.
- Status (≈9%). Producing legibility for others: updates, check-ins, and meetings whose function was to redistribute state that systems already held but could not share.
Friction is a structure, not a failing
The instinctive response — better personal productivity tools — misreads the data. These frictions are not personal inefficiencies; they are properties of the seams between people, tools, and institutions. An individual cannot fix them, and an individual assistant largely relocates them: drafts produced faster upstream became reconciliation work downstream, visible in our data as a transfer of hours rather than a saving.
The frictions are connective failures, and the durable fixes are connective: shared canonical sources so retrieval has one destination; interoperable formats so translation is mechanical; single definitions of record so reconciliation has nothing to reconcile; systems that expose their own state so status reporting stops being a human job. This is infrastructure work — slower and less demonstrable than shipping an assistant, and in our judgment worth several times more hours back per person.
We are publishing the instrumentation protocol and the anonymized time-use data alongside this report. If the field is serious about reducing unnecessary work, it should start by agreeing on a map of where the work actually goes.
More research
All research →Machines That Return Time
AI is evaluated by whether it completes tasks. We argue for a complementary standard: whether it expands the time and capacity people have for judgment, learning, care, and shared work.
Amara Osei, Daniel Reyes, June Park
Interfaces for Collaborative Intelligence
Most AI interfaces assume one person and one model. Most meaningful work happens in groups. Design patterns from a year of building shared-context systems for small teams.
June Park, Tomás Carvalho
Measuring Human Capacity, Not Just Model Capability
A benchmark tells you what the model can do. It tells you nothing about what its users can still do. First results from paired capability–capacity evaluations.
Daniel Reyes, Priya Raghavan