Answer
What happens to your skills when AI does the work?
The production skills fade and the judgement skills only develop through production, which is the problem worth designing around.
The ability to produce declines with disuse, and the ability to judge declines with it, because judgement is built by producing. That is the connection worth designing around rather than the skill loss itself.
The straightforward part is that a skill not used declines. That is ordinary, it applies to arithmetic, navigation, spelling and every other capability that has been delegated to a tool, and by itself it is not alarming — nobody laments their lost ability to use a card index. The reason this case is different is the second-order effect, and it is the one worth thinking about.
Evaluating work requires knowing what good looks like, and that knowledge is built by doing the work badly and then better. A reviewer who has never produced the thing they are reviewing can assess whether it reads well and not whether the approach was right, which specific is wrong, or what was left out. So the capability being delegated is also the training route for the capability being retained.
This is most acute for people early in a career, and it is a management problem rather than an individual one. The tasks traditionally given to a junior — the drafting, the first pass, the tedious research — are exactly the ones now most easily delegated, and they were how judgement was acquired. A firm that removes them has improved its throughput and removed its training pipeline, and the consequence arrives several years later when nobody has the judgement the review depends on.
For an experienced person the risk is different and slower: the judgement was already built, and it decays with disuse and drifts out of date as the field moves. Someone who stopped producing five years ago is reviewing against a standard from five years ago, and nothing in the process surfaces that. Periodically doing the work yourself is the only correction, and it is easy to defer indefinitely because the output is fine.
The design response is to keep some production deliberately rather than to resist delegation. Doing a share of the work yourself, choosing the cases where it matters most, treating some tasks as practice rather than as output. This is a cost and it should be recognised as one — you are paying in efficiency for retained capability, which is a reasonable trade to make consciously and a poor one to make by accident in the other direction.
There is a specific version worth naming for anyone reviewing generated work regularly: the checking habit erodes when it usually finds nothing. Skill at reviewing is maintained by finding things, and a stream of adequate output provides no practice. Periodically checking something you know contains an error, or reviewing against a source rather than by reading, keeps the capability real rather than nominal.
You cannot keep the ability to tell whether it is right while giving away the practice that taught you.
Siddharth Sharma, Context Theory
Related questions
Is this an argument against using it?
No, and framing it that way loses the actual decision. The argument is for deciding which capabilities you intend to retain and paying for them deliberately. Most skills are fine to lose, some are load-bearing for judgement you still need, and the difference is worth thinking about once rather than being resolved by default.
How should a business train people now?
By having them do work whose output could have been generated, deliberately, as training rather than as production. That is expensive and there is no obvious substitute: the alternative is people who can operate the tools and cannot tell when the output is wrong, which is a worse position for the business than the training cost.
METHOD
Every figure below carries its source and the date it was verified. Nothing on this page is asserted.
The numbers on this page.
| What | Value | Specific to |
|---|---|---|
| Close rate — response under 5 minutes vs over 24 hours | 32% vs 12% | Category-wide |
| Sub-15-minute compliance — automated routing vs manual only | 62.5% vs 39.1% | Category-wide |
Optifai speed-to-lead benchmark · n=939 companies · Q2 2025–Q1 2026 · verified
2026 speed-to-lead benchmark · verified
What is specific to this page.
| Kind | Claim | Check it against |
|---|---|---|
| Workflow | Judgement about quality is built by producing work, so delegating production removes the route by which the retained evaluative capability was acquired, which is the second-order effect rather than the skill loss itself. | Asking whether a reviewer who has never performed a task can identify what was omitted from an output of it. |
| Constraint | The tasks traditionally assigned to junior staff are the ones most easily delegated and were the mechanism by which judgement was acquired, so removing them improves throughput and removes the training route, with the consequence arriving years later. | Listing the tasks currently delegated and comparing against the tasks previously given to new staff. |
| Response | An experienced reviewer's standard decays with disuse and drifts as the field moves, and nothing in the review process surfaces that the standard is out of date. | Comparing a reviewer's stated standard against current practice in the field. |
| Software | Reviewing skill is maintained by finding defects, so a stream of adequate output provides no practice and the checking capability becomes nominal, which periodic review against a source rather than by reading counteracts. | Inserting a known error into a reviewed output and observing whether it is caught. |
Each row would be wrong on another industry's page. Where a sourced figure exists it is in the table above instead; these are the constraints that shape the work and do not happen to be numbers.
Start with the measurement.
Reading about a benchmark is not the same as knowing your own number. The audit produces yours, measured rather than estimated.
$497 · delivered in 5 business days · credited against month one