Pricing the Two Costs of AI Delegation
The Authority You Grant and the Capability You Surrender
In the adoption of almost every useful technology, there comes a quiet turning point when the fundamental question shifts from whether the system can do the work to whether it should be allowed to do it. Early on, you ask basic capability questions: can the AI draft the document, analyze the records, calculate the risk, or navigate the workflow? But as demonstrations improve and error rates drop, people naturally grow comfortable and want to give the technology more room to run. That is precisely where enterprise leadership needs to slow the conversation down, not because the technology is failing, but because it is succeeding.
A system’s ability to perform a task and its authority to act on behalf of your organization are two different things. Capability is not authority. Agency is a capability; autonomy is an allocation of authority.
Knowing what good looks like does not answer the next question: how much power should the system receive to act on that judgment? An AI system that produces a good contract is demonstrating capability. Giving it authority to send that contract, bind the company, change a database, move money, advise a client, or trigger the next automated action is something else. You have crossed from asking what the technology can do into deciding what it is allowed to do.
Authority Has to Be Earned
I think about warranted delegation through a chain of questions:
Capability → Validation → Decision Rights → Authority → Accountability → Adaptation
This is not a project sequence, because accountability does not suddenly appear after authority has been granted. It is a way of separating questions that are too easily collapsed into one another. Can the system do the work? Have we validated that capability in our environment? What decisions may it make? What operational authority will those decisions carry? Who remains accountable for the consequences? How will the arrangement adapt when the evidence or environment changes?
The mistake is jumping from capability straight to authority.
A model that performs routine work correctly 95 percent of the time has demonstrated impressive capability, but that number tells you remarkably little about how much authority it should receive. What happens in the other 5 percent? Are the failures detectable and reversible? Do they occur randomly, or in the unusual situations where judgment matters most?
The more authority you grant, the more consequential the same underlying error becomes. A hallucination in a draft is a text problem. The same hallucination attached to write access, commit authority, external communications, or an agentic workflow becomes a state-changing event. Autonomous execution is not autonomous reliability.
The Work Has Two Outputs
Delegation carries another cost that is easy to miss in professional work: automating routine execution can quietly destroy the learning system that produces future experts.
Think about how professional judgment actually develops. Junior professionals encounter ordinary cases, then unusual ones. They make mistakes and get corrected. They discover that a simple rule has exceptions, and eventually, a subtle pattern catches their attention before they can fully explain why. Repetition exposes them to the mistakes, corrections, exceptions, and consequences from which practical judgment develops.
Professional work therefore has two outputs:
- The immediate deliverable: the memo, analysis, contract, recommendation, or calculation.
- The increased capability of the person doing the work.
Our productivity accounting tends to measure the first output and ignore the second. When AI produces the deliverable faster and cheaper, automating the work product can also automate away the activity through which people learn to recognize when that work product is subtly wrong.
This is not an argument for preserving clerical drudgery. Moving information between boxes does not make anyone a better professional. The important distinction is between clerical friction and formative repetition.
If repeated performance teaches nothing we will need later, automate it aggressively. But if repetition is how people learn to recognize exceptions, understand consequences, or acquire judgment needed to supervise the system, its disappearance carries a steep price. If you automate formative repetition, you must design another way to produce the judgment it used to generate.
Reversibility Is Measured in Reconstitution Time
This dynamic creates a quiet atrophy gap: execution capacity goes up while independent recognition capacity goes down.
For a while, an organization has the best of both worlds. AI performs routine execution while experienced people who learned under the old system remain available to supervise it. But if junior people no longer accumulate formative repetitions, the organization is spending down a stock of expertise without replenishing it. The danger is not simply that current experts will forget how to do the work, but that the organization interrupts the pipeline by which the next generation becomes competent enough to challenge, supervise, and replace the system.
That dynamic directly changes how we must evaluate reversibility.
It is tempting to call an AI delegation reversible because, technically, you can turn the system off. But if a workflow has been automated for five years, senior experts have moved on, junior staff never learned the work, and procedures are stale, turning off the machine leaves you stranded.
Reversibility is measured in reconstitution time, which is how long it takes to become competent again after the machine stops. Reconstitution time requires rebuilding judgment, institutional knowledge, training pipelines, and practical experience.
Some capabilities can disappear without concern because they are cheap and easy to reacquire. Others take years to build and are difficult to reconstruct under pressure. Treating both as equally reversible because someone can flip a switch misses the operational reality.
How Far Can It Run?
Authority also changes the speed and propagation of failure.
An agent authorized to prepare fifty proposed transactions for human review has one kind of risk profile. Give that same system authority to execute those fifty transactions before anyone looks again, and nothing about its underlying capability has changed, but its delegated authority has.
Agentic systems can observe, decide, act, observe the result, and act again many times before a human re-establishes orientation. The useful control is not simply human oversight; it is deciding where authority stops. Read is different from propose, propose is different from write, and write is different from commit.
The greater the propagation horizon, the stronger the evidence should be before the system receives the authority.
The Warranted Delegation Test
Before granting consequential write, commit, or operational execution rights to an AI workflow, leadership should answer five questions:
- What specific authority are we granting: read, propose, write, commit, or something else?
- What empirical evidence warrants moving from validated capability to those decision rights?
- What human or institutional capability will be displaced, cease practicing, or cease being produced?
- If the tool fails or becomes uneconomic, what is our actual reconstitution time?
- How many unmonitored machine cycles can occur before human judgment is required to re-establish orientation?
Every consequential delegation has two prices: the risk of the authority you grant and the cost of the capability you surrender.
That second price includes the learning system through which future people acquire the judgment needed to supervise that work. The goal is not to preserve old habits, but to ensure that if AI removes formative work, the organization builds new mechanisms (such as simulations, deliberate practice, or structured exception reviews) to produce the judgment the old system produced incidentally.
Prudent AI is not a prescription for moving slowly. It governs consequential AI commitments under uncertainty while preserving the capacity to learn, adapt, and recover. Move fastest where experiments are cheap, failures are reversible, and learning is valuable. Consequential authority must be earned by evidence rather than inferred from impressive capability, pricing both the authority granted today and the human judgment surrendered tomorrow.
Accelerate reversible learning. Pace irreversible commitment.
Dennis Kennedy – CC BY 4.0 license
[Originally posted on DennisKennedy.Blog (https://www.denniskennedy.com/blog/)]
DennisKennedy.com is the home of the Kennedy Idea Propulsion Laboratory
DennisKennedy.Blog is part of the LexBlog network.
Recent Comments