Editorial
White Paper
The Future of Leadership Development: AI-Driven Manager Enablement
Beware the illusion of phantom productivity. Across the first wave of enterprise AI, employees are saying they feel faster, leaders are praising increased output and adoption dashboards are glowing green. However, in many cases the organizations aren’t more productive; they are just more active. The two are not the same, and the gap between them is at risk of growing if leaders don’t realign the way they think about AI integration.
In a 2025 study by METR, experienced developers expected AI to make them 24% faster and reported feeling about 20% faster afterward. But when they were measured against the clock, they were 19% slower. That gap doesn’t register as a financial signal anywhere in the enterprise.
The phenomenon is best described as a head-start bias. AI is excellent at creating the feeling of momentum. It makes the blank page disappear and produces a fluent draft in seconds, which delivers real psychological relief. But enterprise productivity is not measured by how fast work starts. It is measured by how fast high-quality work reaches a decision, a customer, a claim, a quote, a patient or a release. AI expedites the visible portion of the task, leaving the invisible work that follows untouched: checking, correcting, reconciling, governing and defending the output. While the worker feels faster due to the instant appearance of the draft, the organization experiences a slowdown as the draft enters a system of review and rework.
The Root Cause: AI Bolted Onto Broken Workflows
These taxes are not inherent to AI. They are the predictable result of adding new technology to workflows no one redesigned, the same mistake enterprises made with RPA, analytics and digital. An AI agent beside a fragmented process is not transformation. An algorithm calling an API is not an operating model. A prompt library is not institutional capability.
McKinsey tested 25 attributes against whether companies actually saw earnings impact from generative AI, and the single biggest differentiator was workflow redesign, yet only about a fifth of firms had done it. Automating a broken process does not fix anything. It lets the process fail faster and at greater volume.
Measure the Workflow, Not the Task
Leaders need to refocus their AI success metrics away from misleading user adoption and token consumption rates, and toward real-world productivity gains. A serious scorecard measures the workflow end to end: cycle time from intake to completion, first-pass yield (the share of output accepted without major rework), rework rate, decision velocity, expert review burden and the business outcome itself. The fastest way to expose theater is to replace each vanity metric with the outcome it is pretending to represent.
The table below outlines some examples of how the thinking around AI results measurement needs to shift across different industries and functions.
| Function | Stop measuring (feels productive) | Start measuring (is productive) |
|---|---|---|
| Insurance underwriting | Submissions summarized | Quote cycle time, bind ratio, underwriting leakage, risk selection |
| Banking and credit | Memos and KYC files drafted | Approval cycle time, exception rate, compliance defects, onboarding velocity |
| Healthcare operations | Notes and prior-authorizations generated | Denial reduction, clinician time returned, coding accuracy, care-gap closure |
| Software engineering | Lines of code generated | Project Completion time, change-failure rate, bugs in UAT, review burden |
| Consulting and knowledge work | Decks and summaries produced | Proposal conversion, sharper client decisions, solution quality |
A Real-World Playbook for a New Approach to Productivity Measurement
Recognizing the challenge and implementing a fundamentally new approach to evaluating tech-driven productivity improvements are two very different things. Ultimately, true enterprise AI transformation is as much an exercise in change management as it is in technology integration. The businesses that have successfully shifted the narrative on AI with demonstrable results have done so by following six steps.
- Audit AI work, not AI tools. Map the actual work AI touches and hunt the hidden labor: prompting, checking, copying, correcting, escalating, redoing. The count of platforms and seats tells you nothing about productivity.
- Segment every use case and act differently on each. Sort them into four buckets and fund them accordingly:
- Workflow acceleration. Real, repeatable, verifiable gains. Scale and industrialize.
- Personal productivity. Genuinely useful but local. Allow and govern lightly.
- Illusion of productivity. Feels fast, creates downstream friction. Redesign or stop.
- Net-negative work creation. Generates noise and rework. Kill it.
- Embed AI inside the workflow, not beside it. If people must leave the system, paste data into a model and paste the output back, you have built swivel-chair AI, not throughput. Put it in the claims system, the underwriting platform, the EHR, the SDLC pipeline, the work queue.
- Measure rework aggressively. Track how often output is corrected, by whom, at what cost and which prompts, models and users generate the most friction. Rework is where the productivity claim gets tested and where it usually fails.
- Protect expert capacity. Do not let AI turn senior people into a cleanup crew for junior AI output. Reserve expert review for exceptions, high-consequence calls and capability building. If AI raises expert review load, the operating model is broken.
- Keep a stop-doing list. Do not use AI where the task is rare, highly ambiguous, legally consequential or impossible to verify quickly. Bad use cases waste money and burn trust, which is far more expensive to rebuild.
What Good and Bad Look Like
The companies getting this right measure outcomes, not usage.
One good example is Commonwealth Bank of Australia, which rebuilt its fraud and service workflows around AI and reported scam losses cut roughly in half and call-center volume down 40% year over year.
The cautionary case is Klarna, which aggressively automated customer service and posted spectacular speed and volume numbers, then admitted in 2025 that it had cut human support too far, watched quality slip and began rehiring. Klarna optimized the metrics it could see and missed the ones that mattered. Measuring at the workflow level means measuring quality, not just throughput.
The next competitive advantage in AI will not come from giving everyone agentic capabilities. It will come from knowing where AI improves throughput, where it only improves the feeling of progress, and where it quietly makes the system slower.
Engaged, But Not Convinced: What's Eroding Beneath Your Strongest Engagement Numbers
Watch Now
Retention Is a Communication Problem: New Data from 6,000+ Employees
Watch Now
The winners will not have the most prompts, pilots or AI champions. They will be the ones who redesign the work, connect AI to enterprise context, measure accepted work instead of generated work, protect scarce judgment and govern at runtime.
The hard truth is your AI may be making people feel productive while slowing them down, and unless you measure the full workflow, you will not know the difference.
Editor's Note: What other challenges around AI agents are businesses tackling?
- From AI Being Done to Us to AI Working for Us — Businesses are falling into an old trap with AI: treating it like a hammer in search of a nail.
- AI Promised Efficiency, But Businesses Got a Management Problem Instead — AI is helping us swap tedious work for something new: reviewing outputs we didn't create and carrying the blame when they're wrong.
- Can You Fire an Agent? Why Businesses Need a New Framework for Governing Digital Talent — As AI agents gain autonomy, companies need new governance tools to define authority, control costs and prevent rogue behavior.
Learn how you can join our contributor community.