87% productivity gain<\/strong>.<\/p>\nWhat the dashboard doesn’t capture: the analyst spends 20 minutes specifying parameters and providing sample formatting. Another 10 minutes fixing mid-run errors \u2014 pulled the wrong data source, misinterpreted a chart instruction. And 15 minutes reviewing and correcting the output.<\/p>\n
Total human time: 45 minutes. Total displaced manual work: 90 minutes. Actual leverage ratio: 90 \/ 45 = 2x<\/strong>. Positive ROI \u2014 but far from the advertised 87%.<\/p>\n\n
Worse scenario:<\/strong> A junior associate drafts a routine contract with AI. Dashboard shows 8 minutes vs 60 minutes manual. But the associate spends 30 minutes writing a detailed prompt, 10 minutes re-routing through a missed compliance check, and 25 minutes reviewing for jurisdictional accuracy. Total human time: 65 minutes. Leverage ratio: 60\/65 = 0.92x<\/strong>. Negative ROI, counted as 87% improvement.<\/p>\n<\/p><\/div>\nThe Three Hidden Channels<\/h2>\n\n
Specification Time (T_spec)<\/h3>\n
The cost of translating human intent into AI-understandable instructions. Detailed prompts, examples, boundary conditions, policy constraints, fallback instructions.<\/p>\n
Optimization insight:<\/strong> Agent memory matters as much as capability. An agent that retains context across tasks reduces re-specification time. An agent needing the same instructions repeated inflates T_spec without increasing output.<\/p>\n<\/p><\/div>\n\n
Interrupt Resolution Time (T_int)<\/h3>\n
The cost of handling deviations. The agent goes off course, misunderstands an instruction, or hits an edge case its training didn’t cover.<\/p>\n
Optimization insight:<\/strong> Better capability reduces interrupt frequency. But the relationship is nonlinear \u2014 remaining interrupts become harder as easy problems get solved first.<\/p>\n<\/p><\/div>\n\n
Review Time (T_rev)<\/h3>\n
The cost of verifying output quality. Even correct execution must be validated before use.<\/p>\n
Optimization insight:<\/strong> Trust calibration. As an agent demonstrates reliability on specific task types, review time decreases. But for novel or high-stakes tasks, review must remain high. Trust is task-specific, not agent-general.<\/p>\n<\/p><\/div>\nPer-Task vs. Windowed Leverage<\/h2>\n
This is the paper’s most important strategic insight<\/strong>.<\/p>\nPer-task leverage<\/strong> measures a single execution. It is bounded by task novelty \u2014 novel work always requires human specification and review.<\/p>\nWindowed leverage<\/strong> measures across recurring tasks. Upfront specification, agent configuration, and workflow design get amortized across the window.<\/p>\nAn AI deployment for customer support ticket triage: per-task leverage on the first ticket might be 0.5x<\/strong> (setup outweighs savings). By the 100th ticket, with refined templates and calibrated agent, it might be 8x<\/strong>. Windowed leverage captures both.<\/p>\nThis changes investment strategy. Low per-task leverage on a new deployment is not failure \u2014 it may be upfront investment amortized across hundreds of tasks.<\/p>\n
The Task Novelty Floor<\/h2>\n
The paper identifies a fundamental constraint: truly novel tasks always require human time regardless of AI capability<\/strong>.<\/p>\nA novel task cannot be fully specified in advance \u2014 you don’t know what the output should look like until you see it. A novel task generates unexpected problems. And a novel task requires judgment about correctness that cannot be delegated.<\/p>\n
The strategic implication:<\/strong> Organizations should not aim to eliminate human involvement. They should understand where the novelty floor sits for different task types. High-novelty tasks need human-in-the-loop. Low-novelty tasks are candidates for automation.<\/p>\nThis guides workforce strategy. Routine, well-specified roles face the highest automation pressure. Novel problem-solving roles face the lowest.<\/p>\n
What Business Leaders Should Do Next<\/h2>\n\n- Audit your current AI tooling<\/strong> \u2014 For the top 10 AI-augmented workflows, estimate T_spec, T_int, and T_rev. Compute actual leverage ratios.<\/li>\n
- Identify quick wins<\/strong> \u2014 Tasks above 3x are scaling candidates. Below 1x need redesign.<\/li>\n
- Track the three channels<\/strong> \u2014 Add specification, interrupt, and review time to dashboards.<\/li>\n
- Model amortization curves<\/strong> \u2014 How many recurring executions until upfront investment pays back?<\/li>\n
- Classify tasks by novelty<\/strong> \u2014 Map roles to novelty levels. Guide reskilling toward high-judgment work.<\/li>\n
- Invest in agent memory<\/strong> \u2014 Context retention amplifies leverage on recurring tasks.<\/li>\n
- Balance the portfolio<\/strong> \u2014 Low per-task leverage today might be high windowed leverage tomorrow.<\/li>\n<\/ol>\n
Conclusion<\/h2>\n
Stop asking if AI is productive. Start asking what your leverage ratio is per task type.<\/p>\n
\n
The Leverage Ratio framework exposes hidden human costs standard dashboards miss, distinguishes investment-phase from genuinely unproductive deployments, and provides a clear framework for prioritizing AI investment. Organizations that implement it will make better decisions. Organizations that don’t will systematically overestimate their AI productivity.<\/p>\n<\/p><\/div>\n