TL;DR. Delegability isn't filtered by intuition but in two stages — first diagnose the bottleneck, then filter the bottlenecks by fit. Fit is decided by the asymmetry between one benefit (frequency × standardization) and three costs (verification, irreversibility, context). So delegation isn't "is the benefit large" but "is the risk controlled." And what you hand off isn't a copy of the work — it's a redefinition.
As the cost of execution collapses, anything can be delegated. So skill has moved from "what you delegate" to "what you keep." Yet most of us do this sorting by intuition. That's why the same technology yields outcomes at the extremes — what to delegate is decided not by how impressive the work is, but by whether it's the bottleneck and whether it's reversible.
The Bottleneck Is Rarely Where You Guessed
In an era where execution is free, the bottleneck is no longer the model. MIT's analysis of GenAI adoption, The GenAI Divide (2025), reports that 95% of pilots produced no measurable P&L impact — and pins the cause not on model quality but on failure to integrate with the workflow. The same report flags a sharper mismatch: more than half the budget goes to sales and marketing tools, while the actual ROI is highest in back-office automation (document processing, compliance, internal workflows). The money pooled where the bottleneck isn't.
The trouble is that even that bottleneck is often misdiagnosed. One insurer believed its slow claims processing was due to a shortage of adjusters, but process mining revealed the real constraint was the manual reconciliation forced by system fragmentation. Hiring more adjusters wouldn't have solved the problem — it would only have scaled it. The constraint hid not in team size but in the invisible prep work that eats weeks before the real work even begins.
Why this is fatal for delegation is simple. Automate a misdiagnosed bottleneck and you're just running the wrong work faster. A diagnostic error gets multiplied by execution amplification. The core principle is the one the Theory of Constraints names: improving a non-bottleneck doesn't raise total throughput. So the real first question of "what should I delegate" is "where is the constraint on the whole flow right now."
Among Diagnosed Bottlenecks, Which Can Be Handed Off
Finding the bottleneck doesn't make it all delegable. What actually gets handed off in the field is invoice matching, contract review, compliance documentation, lead preprocessing, code review — all standardized, mid-stream clerical and review tasks. Not flashy autonomous agents. Why exactly these splits along four variables.
Benefit — frequency × standardization. The more often it repeats and the more its procedure is standardized, the more the one-time definition cost is amortized into benefit. A structured task with fixed inputs is reusable once specified, but a one-off task that takes a different shape each time has a setup cost that exceeds the execution benefit. The cost gap is dramatic — when a human handles one customer-service case at $4.18, an agent finishes it at $0.46 (9×); when a senior engineer writes one routine PR at $48, a code-review agent does it at $0.72 (66×). The more standardized and repetitive, the more that gap becomes pure benefit.
Cost one — verification. Can you determine deterministically and quickly whether it's done and correct? If verification demands human judgment every time, the bottleneck just moves from execution to verification rather than disappearing. The data shows it — only 41% of agent rollouts clear positive ROI within 12 months, and 19% never reach payback. The cause is almost entirely evaluation drift, governance gaps, and unmeasured rework — not agent capability.
Cost two — irreversibility. If a mistake is cheap and reversible, you can hand it off (the argument that the axis is reversibility cost, not scale carries straight over from coding to delegation). Moving money, contracts that go out the door, posts that can't be unpublished — without guards their expected value is negative. The industry rule of thumb splits here too: high-volume, low-risk goes fully autonomous; high-volume, high-complexity goes to human-reviewed assistance; low-volume, high-risk goes to strict human-in-the-loop. In fact, in domains like legal (1.4×) and clinical (1.2×), where governance review eats almost all the speed advantage, the productivity multiple hits bottom.
Cost three — context. Is the context already structured, or do you have to gather scattered tacit knowledge every time? If the latter, the cost of filling in context erodes the execution benefit, and fit drops. Look at where delegation fails and it's clear — the bottleneck is everything between the frontier model and a measurable outcome. The bulk of the work isn't the flashy part; it's the part that makes context cheap.
The asymmetry here is the crux. There's one benefit but three costs. The delegation decision is governed less by "is the benefit large" than by "are the costs and risks controlled." Hand off not the impressive work, but the work where verification is cheap, the action is reversible, and the context is already organized.
Not a Copy, but a Redefinition
The most common mistake is automating a broken workflow as-is. If the manual version is slow because the underlying design is bad, automation won't fix it — it'll just run the bad design faster and at greater scale.
So delegation has to be a redefinition of the work, not a copy of it. Not transcribing the order a human used to follow, but redefining "what the success condition is, how to verify it, and where to stop." The product of that redefinition is exactly the verification apparatus and the guards. Measurement and guard infrastructure aren't a byproduct of delegation but a precondition. The figure of 171% average ROI circulates, but it's dangerous precisely because it's an average — with the same technology, some recover 171% while Gartner expects over 40% of agent projects to be canceled by 2027. What separated the winners from the failures wasn't the model but the deployment method. The ones who recovered started with a few high-value tasks, attached them to clean data, and measured cost-per-resolution rather than "adoption rate."
To Sort What to Keep, You First Have to Measure
In an era where execution is free, the bottleneck isn't execution but judgment. Decide what to delegate without measurement, and the agent will do the wrong thing perfectly, fast, and at scale. Skill in delegation is, in the end, the ability to sort out what to keep — and to sort it, you first have to measure: where the bottleneck is, and whether that bottleneck is reversible.
References
The bottleneck isn't the model
- The GenAI Divide: State of AI in Business 2025 (MIT Project NANDA, Jul 2025) — 95% of pilots with zero P&L impact; the failure cause is integration and the learning gap, not the model (mirror)
- Theory of Constraints — improving a non-bottleneck doesn't raise throughput
- Process-mining case of a misdiagnosed bottleneck (insurance claims processing)
The numbers that decide fit
- AI Agent Productivity Statistics 2026 (Digital Applied aggregation; Forrester TEI · Anthropic · Gartner sources) — cost-per-task 9–66×, positive ROI within a year 41% · never recovered 19%
- The autonomous / assisted / human-in-the-loop 2×2 rule of thumb
The 171% trap
- Deloitte State of AI in the Enterprise 2026 — average ROI 171% (US 192%), but an average
- Gartner — over 40% of agent projects forecast to be canceled by 2027
Related