When AI Increases Work Load Instead of Reducing It

What new research reveals—and why “work-type design” is the missing fix

Peter Stefanyi, Ph.D., MCC, Colaborix GmbH

February, 2026

In February 2026, researchers at UC Berkeley Haas reported a finding that should make every CEO and transformation leader pause: generative AI didn’t free up time at work—it intensified work. (Haas News | UC Berkeley Haas)

Their eight-month ethnographic study inside a ~200-person tech company observed a clear pattern: employees moved faster, expanded what they considered “their job,” and extended work into more hours of the day—often without being asked. (Haas News | UC Berkeley Haas) The accompanying Harvard Business Review piece sharpened the point: AI’s promise of time savings can backfire because it makes it easier to start tasks, continue tasks, and keep multiple threads running at once. (Harvard Business Review)

This is not a niche concern. It explains why many “Copilot for everyone” rollouts feel like this:

• More output, but also more cognitive load

• More momentum, but less closure

• More capability, but rising expectations and constant context switching

So what’s actually happening—and what should organizations do about it?

The core mechanism: AI expands the feasible action space (and dissolves stopping points)

The Berkeley team didn’t start with a hypothesis that AI makes work worse. They observed what people did in real work: AI made it easier to take on tasks that used to be someone else’s job, fill micro-gaps (lunch, “between meetings,” evenings), and run multiple parallel workflows. (Haas News | UC Berkeley Haas)

That’s an important shift in framing:

• AI doesn’t just speed up tasks

• It widens the scope of what seems doable

• And it erodes natural boundaries that used to end the day

The researchers propose building an “AI practice”: intentional pauses, batching, sequencing, and human grounding so speed doesn’t crowd out reflection. (Haas News | UC Berkeley Haas)

We agree—but we’ve found something crucial: practice alone is fragile when the structure of work is mismatched.

The missing variable: not all work is the same

Most AI strategies treat “work” as a single category. In reality, organizations run at least three distinct work logics—plus a fourth state that causes most of the pain.

Work Type 1: Independent (pooled) work

People work mostly separately; outputs add up.Examples: sales outreach, individual analysis, drafting, individual coding.

Typical failure mode: forced collaboration (meetings, alignment rituals, collective review) that adds overhead with little gain.

Work Type 2: Sequential work

Work flows through handoffs; throughput is set by the bottleneck.Examples: approval pipelines, document production flows, service operations.

Typical failure mode: optimizing everything except the constraint—creating motion without system-level improvement.

Work Type 3: Reciprocal work

Work quality emerges through interaction and integration.Examples: strategy, product design, complex problem-solving, cross-functional decisions.

Typical failure mode: treating it like independent work (“everyone do your part, we’ll integrate later”), which creates rework and coordination debt.

Work Type 4: Complex (mixed) work

This is the reality of leadership, innovation, and most AI-enabled knowledge work: a mixture of independent analysis, sequential handoffs, and reciprocal integration—often in the same week.

Typical failure mode: skipping decomposition. Everything gets treated as “collaboration,” and AI accelerates every thread at once—so humans become the integration bottleneck.

That fourth state—undecomposed complexity—is where the Berkeley Haas “intensification” pattern becomes especially likely.

Why AI intensifies work: mismatch between AI mode and work type

This is where a lot of organizations accidentally recreate what Raisch & Krakowski call the automation–augmentation paradox: leaders want AI to both automate work and augment human capability—often simultaneously. (journals.aom.org) The paradox becomes a practical problem when organizations deploy one collaboration pattern everywhere.

A simple translation:

• Independent work wants AI as a tool (fast, standardized, low coordination)

• Sequential work wants AI at the bottleneck (workflow-aware, selective, measurable)

• Reciprocal work wants AI as a partner (iterative, integrated, governed by shared team norms)

• Complex work wants decomposition first (then apply the right pattern to each component)

If you skip that logic, AI increases capability and coordination cost at the same time—exactly the intensification Berkeley observed. (Haas News | UC Berkeley Haas)

Why “generic AI training” underdelivers in reciprocal work

A major point business leaders miss: in reciprocal work, performance is often constrained less by individual skill than by coordination quality—shared understanding, integration routines, and calibrated trust.

That is consistent with decades of team research showing that “team cognition” and shared knowledge structures are strongly related to team process and performance. (PubMed)

In plain language: if the work depends on integration, then dropping AI in without building “how we use it together” creates more threads, more drafts, more options—and more friction.

The Colaborix approach: structure before speed

What we do differently is simple to state and hard to execute without a method:

Step 1: Diagnose work by type (fast)

We map where work is independent, sequential, reciprocal—and where it’s mixed.

Step 2: Decompose complex work (the critical move)

Complex work is not a fourth “kind of work” to optimize directly.It is a signal to split the work into components and assign each component the right coordination logic.

Step 3: Match team design + AI mode to each component

• Independent: give people AI for personal productivity + set clear scope/stopping rules

• Sequential: place AI at the constraint + tighten handoffs + measure throughput/cycle time

• Reciprocal: build shared norms for AI use (who uses it when, how outputs enter decisions, how validation happens)

• Complex: orchestrate transitions between the above

This is what we mean by flexible teamwork: teams don’t “collaborate more.” They collaborate where the work structure requires it, and they protect focus and closure everywhere else.

What leaders can do next week

Here are three practical moves that apply immediately:

1) Run a “work-type audit” on your top 10 AI use cases

For each use case, ask:

• Is the output additive (independent)?

• Is it a handoff chain (sequential)?

• Does quality emerge through integration (reciprocal)?

• Or is it mixed (complex)?

2) Redesign coordination before scaling AI

If your rollout increases meetings, Slack traffic, and review loops, you’re probably forcing reciprocal coordination onto independent work—or accelerating non-bottlenecks in sequential work.

3) Define stopping rules

The Berkeley study is, at its heart, a story of disappearing stopping points. (Haas News | UC Berkeley Haas)Stopping rules are not “wellness tips.” They are operational controls—especially for independent and sequential work.

Bottom line

The Berkeley Haas findings should not be read as “AI makes work worse.” They show something more useful:

AI accelerates whatever work design you already have.If your work is structurally ambiguous, AI expands scope, dissolves boundaries, and intensifies workload by default. (Haas News | UC Berkeley Haas)

The solution is not to slow down innovation. It’s to design the operating system of work:

• differentiate work types

• decompose complexity

• match coordination mechanisms

• then deploy AI accordingly

That’s how you turn AI from a workload amplifier into a sustainable advantage.

References and suggested citations

1. Ye, X. M., & Ranganathan, A. (2026). AI promised to free up workers’ time… UC Berkeley Haas researchers found the opposite. Haas News (Feb 18, 2026). (Haas News | UC Berkeley Haas)

2. Ye, X. M. (2026). AI Doesn’t Reduce Work—It Intensifies It. Harvard Business Review (Feb 9, 2026). (Harvard Business Review)

3. Raisch, S., & Krakowski, S. (2021). Artificial Intelligence and Management: The Automation–Augmentation Paradox. Academy of Management Review, 46(1), 192–210. (journals.aom.org)

4. DeChurch, L. A., & Mesmer-Magnus, J. R. (2010). The cognitive underpinnings of effective teamwork. Journal of Applied Psychology, 95(1), 32–53. (PubMed)

5. Thompson, J. D. (1967). Organizations in Action: Social Science Bases of Administrative Theory. McGraw-Hill. (Foundational typology of pooled/sequential/reciprocal interdependence.)

6. Research on types of work and Hierarchical intervention model for team performance is in articles down this blog.