The Amplification Thesis

"Where can we cut headcount?"

That is a legitimate budget question. It is a poor specification for a product. Removing a task from someone's day might save money, or it might free them to do work that was previously out of reach. Those are different outcomes, and a headcount target cannot tell you which one happened.

Start with the work you want done better. A faster first draft is useful only if getting to an acceptable final result also gets easier.

Automate, or stay manual

There is a third option: change how the work is done. A person might frame the problem, use AI to explore possibilities, and return to information the model cannot supply. That is a candidate for amplification. It still has to earn its place.

Researchers distinguish human augmentation, beating the human-only baseline, from beating the stronger of the human-only and AI-only baselines. The latter is the harder test. Passing one tells you nothing certain about passing the other.

Better than a person. Worse than the best option.

Michelle Vaccaro, Abdullah Almaatouq, and Thomas Malone analyzed 106 experiments reported in 74 papers, with 370 effect sizes. On average, human-AI combinations improved on humans alone. But they performed worse than the better standalone option. Both results came from the same review.

The plotted values are standardized effects, not percentage improvements. The papers covered January 2020 through June 2023, so this is evidence about those systems and experiments, not a verdict on every current model.

Decision tasks showed losses against the stronger baseline. Content-creation tasks were more promising, but their pooled advantage over that baseline was not statistically distinguishable from zero. "Creative work always wins" would overstate this result too.

A narrower case for creative work

The May 2025 version of Niklas Holzner, Sebastian Maier, and Stefan Feuerriegel's meta-analysis included 28 studies and 8,214 participants overall. Its human-with-AI versus human-without-AI comparison used 21 observations and found a small positive effect on creative performance.

Results varied substantially; removing some individual studies made the confidence interval cross zero. A separate comparison found no statistically significant difference between AI alone and humans alone. That does not establish equivalence, and stitching the comparisons together does not establish that the combination beats both.

The review also found lower idea diversity with AI assistance, based on six observations from four studies. Better-rated individual outputs and a narrower pool of ideas can coexist. If you want exploration, measure variety as well as the quality of the selected answer.

Give the person something useful to contribute

Complementarity is a design hypothesis: the person brings information or a capability that improves what the model can do, or vice versa. A designer may know why customers abandoned a feature. A model may produce variants the designer had not considered.

Neither contribution is guaranteed. People can miss errors; a fluent model can invent context. Asking someone to approve an answer does not, by itself, add expertise to it. The useful question is specific: what changes because this person is involved?

An old ambition, not an old proof

In 1960, J.C.R. Licklider proposed a partnership in which people set goals and evaluate results while computers do routinizable work. Douglas Engelbart's 1962 framework treated improving human problem-solving as a system-design problem, including methods and tools.

That is a useful tradition to work in. It is not six decades of experimental proof that any human-AI pairing will win.

Ask the people doing the work

Stanford's WORKBank study collected responses from 1,500 workers across 104 occupations. It found varying preferences for human involvement, alongside interest in automating low-value, repetitive tasks.

Preferences are design input, not evidence that a preferred arrangement performs better. Ask which parts of the work people want help with, and which knowledge they need to keep using. Then test the proposed division of work with them.

A designer working on a sign-up form

Imagine a designer trying to improve a public library's sign-up form. This is a hypothetical example, not a production case or a disguised set of operational results. The designer brings observations from user interviews and sets constraints: accessibility, plain language, and the information the library actually needs.

AI proposes alternative wording and layouts. The designer rejects options that misunderstand the service, tests the remaining options with users, and revises the form. Generating variants is only one part of the job. Discovering that a question should not be on the form at all may matter more.

Count the whole job

Compare the designer's existing workflow with the assisted version on comparable briefs. Where safe and feasible, include an AI-only output given the same documented constraints. Decide the quality bar before looking at the results. Assess usability and accessibility, and count time spent prompting, checking, and reworking.

Use more than a favorite demo. Rotate task assignments to reduce practice effects, have outputs evaluated without revealing how they were made where possible, and inspect costly errors separately from average quality. This is a proposed evaluation, not a claim that the hypothetical workflow succeeds.

Better output is not proof of learning

A good form does not tell you whether the designer learned anything. If skill growth is part of the investment case, test that separately later. Can the person explain the design choices, recognize a flawed suggestion, and handle a comparable brief without assistance?

Those are checks I would build into the rollout, not findings established by the short-task studies above. Keep opportunities for independent practice if that capability matters. Do not book "compounding expertise" as a return before measuring it.

The business case has to survive the costs

Domain knowledge and a well-tested workflow can be valuable. Calling them an uncopyable moat goes too far. Collaboration costs time, and automation can also improve as systems and processes improve. Neither strategy comes with a predetermined growth curve.

A useful investment case says what got better, what it cost, and which alternative it beat. If the benefit disappears once review and rework are included, redesign the workflow or stop using it for that task.

Some things should be automated

If automation meets the required quality and risk constraints at lower cost, use it. In the library example, checking whether required fields are present may need ordinary code rather than a generative model or a reviewer.

And if an AI-only option scores well but is not legally or operationally feasible, say so. The best benchmark score and the best deployable workflow are different questions. Human involvement may be required for accountability even when it does not improve the measured score.

Design for amplification. Make it prove itself.

I would start by looking for work people could do better with AI, rather than assuming the only return is a smaller team. That opens possibilities a replacement-only plan misses. It does not excuse a weak comparison.

Keep the person involved for a reason you can name. Keep the workflow because the results justify it.