Research: How AI Agents Broaden the Scope of Knowledge Work

Jeremy Yang, Kate Zyskowski, Noah Yonack, Jerry Ma (Perplexity) · article · 2026-07-29 Reader 🖥️

Agentic AI AI Unit Economics Enterprise AI Adoption & Readiness Human as Orchestrator of Agents Human-in-the-Loop Jevons Paradox Measuring AI Impact Vendor Research Conflict of Interest

AI agents don't just make existing knowledge work faster — by combining much longer autonomous execution with far broader tool and context integration, they lead people to attempt fundamentally different, more cross-disciplinary, and previously-impractical work, so leaders must redesign roles, review points and team boundaries around this expanded scope rather than just harvesting the time and cost savings.

practical · mixed — Opens as a theoretical/empirical field study (what changed when the same users moved from a conversational assistant to an autonomous agent) but is organized entirely in service of a practical conclusion — five explicit management moves for leaders. The data sections (autonomy, efficiency, scope) are evidentiary scaffolding for the closing prescriptive section.

Structure

Part	Role
Opening vignette (two tools, one competitive-landscape request) plus a three-generation taxonomy — conversational assistants, copilots, agents — defined along two axes: autonomy and context integration.	context — establishes the conceptual map and vocabulary the rest of the piece measures against
Autonomy findings: ~48x more autonomous machine-time per matched session, pause/clarify rates up sharply (38% vs <1%) with stop rates flat (3.7% vs 3.4%), lower dissatisfaction, and complementary (not substitutive) use with the assistant.	evidence — establishes the behavioral shift from operator to supervisor
Efficiency findings: 87% time reduction and 94% cost reduction for agent-plus-human vs. assistant-plus-human workflows, triangulated (per the authors) against an LLM-based estimate and user self-reports, plus the fixed-cost/marginal-cost explanation for when agents pay off.	evidence, explicitly flagged by the authors themselves as 'the easiest thing to over-index on' and not the main point
Scope findings: more higher-order/create-level/abstract queries, more knowledge domains per query, more cross-occupational requests, and 23-40% of agent tasks never attempted with the assistant at all.	evidence — the paper's most novel claim and its true center of gravity
Five management prescriptions (design for supervision, route by complexity/verifiability, build a human commit boundary, redesign roles around scope, measure scope not just efficiency).	conclusion / call to action
Limits paragraph (early-adopter sample skew, matched-query exclusion of complex agent-only tasks, session-as-task noise, LLM-classification measurement error, single-ecosystem observation).	qualification

Key terms

assistant / agent
Two operationally distinct product categories measured along two axes — autonomy (minutes of unsupervised planning/execution) and context integration (rate and count of external tool/connector calls) — not marketing labels. The assistant is the low-autonomy, low-integration conversational baseline; the agent is high on both.
scope
The range and kind of task a user attempts with a tool — operationalized via LLM-based classification of cognitive level ('higher-order', 'create' vs. 'remember/summarize'), count of knowledge domains touched, cross-occupational-boundary crossing, and incidence of tasks 'never even tried' with the assistant. Treated as a construct orthogonal to speed/cost efficiency.
matched sessions / matched queries
Pairs of near-identical opening requests the same user submitted to both the assistant and the agent, used to isolate a same-task, same-user comparison for the autonomy and efficiency numbers.

Propositions

explicitOn matched sessions, the agent spent about 26 minutes of autonomous planning/execution versus roughly 33 seconds for the assistant (about 48x more machine work on average; about 40x at the median).conf: high
explicitAlthough the agent paused for permissions/clarifying questions far more often than the assistant (about 38% of sessions vs. under 1%), the rate at which users actually stopped it was essentially unchanged (3.7% vs. 3.4%) — people treated the pauses as checkpoints, not red flags.conf: high
explicitUsers were, if anything, more satisfied with agent sessions than assistant sessions (any dissatisfaction 10.8% vs. 16.6%) — more autonomy did not produce more frustration.conf: high
explicitAdopting the agent increased daily assistant usage by about 1.05 searches per user — the two tools are complements, not substitutes, because offloading longer work to the agent freed capacity for the shorter tasks the assistant handles well.conf: high
explicitRelative to an assistant-plus-human workflow, an agent-plus-human workflow reduced average task time by 87% (269 to 36 minutes) and cost by 94%, with a second, LLM-based estimate reported as landing in a similar range (84% time saved, 93% cost saved) and users self-reporting even larger speedups (median ~25x).conf: high
explicitThe efficiency advantage follows from a cost-structure shift: agents carry a higher fixed cost (specifying the task, reviewing output) but a low marginal cost per execution step, so they pay off on long, complex, multi-tool tasks and lose to assistants on short, simple ones.conf: high
explicitcontestedWhat people asked the agent to do was categorically more cognitively demanding than what they asked the assistant: more higher-order cognition (76% vs. 55%), more "create"-level work (50% vs. 26%), more abstract-vs-routine tasks (71% vs. 53%), and more knowledge domains per query (2.40 vs. 1.74, with over half of agent queries spanning 3+ domains vs. 17% for the assistant).conf: high
contradictsArvind Narayanan on Why AI Isn't All That Revolutionary · p4 — Narayanan argues AI is compressing only the 'execute' layer of cognitive work while 'decide' and 'deliver' remain human-bound; this piece's finding that agent queries skew heavily toward 'create'-level, abstract, higher-order work (50% vs. 26% create-level) suggests agents are already reaching into the decide/deliver layers, at least within one product's real usage.
explicitAgent requests crossed occupational boundaries more often than assistant requests (59% vs. 50%), with the largest increases in cross-occupational task share concentrated in management/entrepreneurship (+19 points), digital technology (+13), and arts/design (+12).conf: high
explicitcontestedAbout 23% of what people asked the agent to do (up to 40%+ under a looser definition) were tasks they had never even tried with the assistant — work that only became realistic once an agent could do it, and these unlocked tasks were generally more complex than the rest.conf: high
supportsYou Don't Get to Create Anything (Still Burning: Beck + Randy Shoup on Jevons paradox and distributed systems under AI) · p1 — Beck/Shoup predict from Jevons paradox that near-free cognition creates more cognitive work by making previously-uneconomic tasks viable; this piece's field data — 23-40% of agent tasks were never attempted with the assistant, plus agent adoption increasing (not cannibalizing) assistant usage — is a concrete empirical instance of that mechanism.
contradictsHow People Are Really Using AI in 2026 · p8 — That piece, reading public social-media posts, found autonomous agentic operations still 'experimental and small-scale' in 2026; this piece's internal-telemetry field study finds agent use already reshaping 23-40% of task content within one company's product — a sharp contrast that may reflect a real maturity gap between consumer-facing agent features and a dedicated agent product, or the very different (and equally non-independent) methodologies each study relies on.
explicitcontestedA 'time and cost saved per task' dashboard is the most likely metric to mislead managers, because it rewards doing old work faster and is blind to the new, more ambitious work agents make possible — scope expansion, not raw efficiency, is where the larger returns and larger risks live.conf: high
contradictsManagers Are Struggling to Keep Up with the AI Productivity Boom · p4 — That piece frames the fix for AI-accelerated manager overload as a manager-behavior change that can 'fully resolve' the bottleneck; this piece argues the deeper problem is structural — roles, boundaries, and review points must be redesigned around expanded scope, not just adjusted individual behavior.
explicitdevelopedLeaders should adopt five specific moves: design workflows for supervision rather than constant operation, route work to agents by complexity/verifiability rather than reflex, build an explicit human commit boundary for consequential or irreversible actions, redesign roles and team boundaries around scope rather than headcount, and measure scope expansion alongside efficiency.conf: high
supportsWhy OpenAI and Anthropic Are Becoming Consultants · p1 — Both pieces converge on organizational and role redesign, not raw tool capability, as the actual bottleneck and lever for realizing AI's value — here framed as redesigning roles around scope, there as organizational readiness outpacing model capability.
extends ←Human Oversight Is Not a Human in the Loop · p11 — The HBR piece's prescriptive leader moves (an explicit human commit boundary for consequential/irreversible actions, routing by complexity/verifiability) are a practical instantiation of Bataller's event-triggered-oversight design principle.
implicitBy presenting the 'second, LLM-based estimate' that lands 'in the same neighborhood' as the primary human-equivalent-cost estimate as corroborating triangulation ('we didn't rely on a single method'), the authors imply independent cross-validation — but both estimates were generated and calibrated by the same research team using models and assumptions of their own choosing, so their agreement is much weaker evidence than the framing suggests, and does not rule out a shared bias running through both.conf: high
implicitThe authors' own prescription to route agents only tasks whose output is 'relatively easy to verify' sits in direct tension with their own scope finding that agent use is increasingly crossing outside a user's primary occupational expertise — precisely the condition under which a user is least equipped to verify the output — and the piece never revisits the verifiability recommendation in light of the scope-crossing finding.conf: medium
implicitcorroboratedThe piece treats the authors' institutional position — Perplexity staff measuring engagement with and the economic value of Perplexity's own commercial assistant and agent products — as a neutral research setting rather than a disclosed limitation: the affiliation is noted once in passing ('where the author team works') and never discussed as a source of bias, even though the limits paragraph itemizes several other caveats (sample skew, matched-query exclusion, measurement error, single-ecosystem observation) in explicit, itemized detail.conf: high
supports ←AI and IT Teams Often Clash. But They Don't Have To. · p16 — Same undisclosed-institutional-interest pattern in a different guise: HBR contributors with a direct commercial/advisory stake (Perplexity's own product; the authors' consulting practice) present their own account of outcomes as neutral case evidence without disclosing or examining the conflict.
implicitThe classification of tasks as 'higher-order,' 'create-level,' or 'abstract,' and the finding that unlocked tasks are 'generally more complex,' rests on LLM-based classification that the authors themselves acknowledge introduces 'measurement error,' yet no coding scheme, example tasks, or inter-rater/human-audit reliability check is given, so a reader cannot independently judge how much of the reported scope-expansion is real behavioral change versus an artifact of how the classifying model was prompted.conf: medium

Arguments

Agents change the content of work, not merely its speed, and therefore require redesigning roles and team boundaries around expanded scope rather than just capturing the time/cost savings. strength: medium

Explicit premises:

Agent queries skew toward higher-order, create-level, abstract, multi-domain work more than assistant queries.
Agent requests cross occupational boundaries more often than assistant requests.
23-40% of agent tasks were never attempted with the assistant at all, and these are generally more complex.

Implicit premises:

The scope-expansion pattern observed within Perplexity's own agent and user base generalizes to agentic AI tools broadly, rather than being an artifact of Perplexity's specific product design, feature set, or early-adopter user population.
Current job descriptions, team boundaries and 'who does what' assumptions are drawn along single-domain-expertise lines that autonomous, cross-domain-capable agents actually violate in practice.

Inferences:

If scope continues expanding as agents mature, occupational taxonomies (job titles, hiring criteria, licensing) built around single-domain expertise may eventually need structural revision beyond what any single company's internal role redesign can address.

The reported efficiency gains (87% time, 94% cost reduction) are a credible order-of-magnitude estimate of the value agents create when delegated action-taking work.