Faster busywork is still busywork

Over 85% of employees at major companies now use AI tools regularly. Adoption is not the bottleneck. It never was. BCG's 2025 Build for the Future report found that only 5% of companies globally qualify as future-built, actually generating substantial business value from AI. Those 5% see five times the revenue increases and three times the cost reductions compared to their peers. Meanwhile, 60% report minimal gains despite significant spend.

The difference is not which tools they bought. Nearly 90% of future-built companies attribute their value to reshaping and inventing business processes, not to efficiency gains on existing workflows. They redesigned roles, decision structures, and end-to-end flows around what AI makes possible. Everyone else bolted powerful tools onto legacy operations and called it transformation.

Speed without direction

The instinct is understandable. You have a process. It's slow. AI makes it faster. Ship it.

But most organisations are using AI to summarise meetings that shouldn't exist and generate drafts of documents nobody reads. Powerful tools doing trivial work. BCG's separate AI Adoption Puzzle report asks the obvious question: if everyone is using AI, why hasn't value exploded? Their answer is that most adoption is stuck at individual productivity tools. Almost nobody is progressing to structural redesign.

The return on AI investment for these companies is what you'd expect from making busywork faster: negligible.

The intensification trap

A February 2026 study in Harvard Business Review reveals the mechanism. Berkeley Haas researchers Aruna Ranganathan and Xingqi Maggie Ye tracked 200 employees at a US technology company from April to December 2025. Their finding was counterintuitive: AI doesn't reduce work. It intensifies it.

Three work intensification patterns emerged. Task expansion: employees managing multiple AI-generated threads simultaneously, doing more because the tools made more possible. Cognitive overload: constant switching between checking, correcting, and directing AI outputs. Temporal creep: work bleeding into evenings and weekends because AI makes it feel low-effort. "I'll just ask Claude to draft this" at 9pm becomes routine. Always-on capability meets always-on culture, and the boundary between work and rest dissolves.

The initial productivity spike gave way to burnout, lower quality, and turnover. Give people AI tools without changing the work itself and they don't do less. They do more, faster, across more hours, until something breaks.

Automating broken things

Klarna is the highest-profile version of this failure mode. In early 2024, CEO Sebastian Siemiatkowski announced that an OpenAI-powered chatbot was handling the work equivalent of 700 customer service agents, resolving queries in under two minutes versus the previous eleven. Headlines celebrated. By mid-2025, Klarna was rehiring humans. Customer satisfaction had dropped 22%.

Siemiatkowski admitted publicly: "We focused too much on efficiency and cost. The result was lower quality, and that's not sustainable."

The chatbot handled volume. It couldn't handle dispute resolution or the emotional nuance of real customer problems. The humans had been compensating for gaps in an incomplete service model, and automation exposed those gaps rather than filling them. Klarna automated the surface without redesigning what sat underneath.

An Orgvue and Forrester survey found the pattern is widespread: 55% of companies that rushed to replace human workers with AI now regret the decision. The problem was never capability. Automating a broken process makes it worse faster.

A line often attributed to Bill Gates captures it: automation applied to an efficient operation magnifies the efficiency; automation applied to an inefficient operation magnifies the inefficiency. Most companies skipped the step where they figure out which one they are.

The coordination tax

A MarTech analysis documented what I'd call the broken workflow trap. A marketing team's newsletter process involved five handoffs among four people. Even with AI tools bolted on, it still took four days. Not because the writing was slow, but because of wait states: time spent sitting in someone's queue between steps.

Only when they redesigned the end-to-end flow did the timeline collapse. An AI agent detecting new posts, drafting content, generating graphics, and staging for human review. The bottleneck was never the work itself. It was the coordination overhead around it.

McKinsey's research on the agentic organisation reinforces this. They estimate AI-powered workflow redesign could unlock $2.9 trillion in annual economic value in the US alone by 2030, but only if organisations redesign work around human-AI partnerships rather than automating tasks in isolation. High-performing companies in their data are nearly three times as likely to have fundamentally redesigned workflows. Amazon, Moderna, and McKinsey itself are eliminating layers of middle management, building flat networks where two or three people supervise 50 to 100 AI agents.

The legibility problem

There is a deeper reason organisational redesign is hard. Processes encode decades of accumulated constraint. Every handoff, approval step, and review cycle exists because someone once decided it was necessary. Many of those decisions were responses to problems that no longer exist, or workarounds for capabilities that humans lacked but AI now provides.

Redesigning work requires making implicit knowledge explicit. The judgement calls the senior person makes without thinking. The unwritten rules that keep operations running. None of this is documented because it never needed to be. AI forces a legibility audit: you can only automate what you can articulate, and most organisations cannot articulate how they actually work.

The anti-patterns persist for this reason. Your team uses AI to summarise meetings because questioning whether the meeting should exist means confronting who called it and why. You draft emails faster because questioning whether the email needs sending means examining the communication structure. You speed up approvals because questioning whether the step adds value means challenging someone's authority. Process redesign is political, not technical.

What the 5% actually do

The companies generating real value start from outcomes and work backwards to the minimum viable process. They assign human versus AI responsibilities based on where judgement is required, not on historical role boundaries. They identify which handoffs exist for coordination rather than value, and eliminate them.

As a product engineer, this maps to every bad automation project I've seen. Teams automate CI/CD without fixing the test suite and deploy faster but break more. Teams add AI code review without addressing PR size conventions and get more noise, not signal. Speed amplifies whatever is already there, good or bad. The 5% treat AI as a forcing function to examine which work should exist at all. The 95% treat it as a way to do existing work faster.

The real curve

The adoption curve is a red herring. Everyone has adopted. The technology is not the constraint.

The curve worth measuring is redesign depth: how far down the organisational stack you are willing to go. At the surface, individuals use AI for personal productivity. One layer deeper, teams redesign workflows and strip out coordination overhead. At the foundation, organisations rethink which work should exist, which roles should be recombined, and which processes only persisted because humans couldn't structure them differently before.

Most companies are stuck at the surface. The Berkeley Haas research shows what happens there: people work more, not better, and the gains evaporate into burnout. The 5% that have gone deeper are pulling away, and the gap compounds with every quarter. As the cost of machine intelligence continues to fall, the returns to structural redesign will steepen while the returns to surface-level automation approach zero. The 95% will eventually face a choice: redesign the work, or watch the gap become permanent.