Augmentation Beats Replacement

If you remember one finding from the research on AI in the workplace, make it this: the biggest, most reliable productivity gains from AI go to your least-experienced people. That single fact should reshape how a leader thinks about where to deploy AI and what it is for.

The intuition runs the other way for most executives. We assume the expert — the person who already knows the most — will squeeze the most out of a powerful new tool. The field evidence says the opposite. A novice paired with AI starts to perform like a mid-level employee; the expert, who already operates near the top of the task, has little headroom to gain.

This reframes the entire conversation. AI is not primarily a tool for replacing people. It is, on the best current evidence, a tool for raising the floor — pulling your weakest performers up toward your average, and your average up toward your best. That is an augmentation story, and it is a far more defensible bet than a replacement story, because it grows your existing capacity rather than betting that the technology can do a whole job alone.

The leadership question shifts from "how many roles can AI replace?" to "how much can AI raise the capability of the people I already have?"

The cleanest, largest piece of evidence comes from a study by Brynjolfsson, Li & Raymond, Generative AI at Work (NBER, 2023), run across 5,179 customer-support agents at a software firm. The agents were given a generative-AI assistant that suggested responses in real time, and their performance was measured against agents working without it.

The results (as of the 2023 study — verify the live source before quoting):

Three things make this study unusually trustworthy for executives. First, it is large and real — over five thousand agents doing actual paid work, not a lab simulation. Second, it measured a hard business outcome — issues resolved per hour — not a soft proxy like "satisfaction with the tool." Third, it found two bonus effects beyond raw speed: customer sentiment improved, and employee retention rose, because the job got less frustrating and newer agents ramped faster.

The shape of the result — big gains for novices, near-zero for experts — is the augmentation thesis in one chart. AI did not replace the support team. It made the less-experienced half of the team dramatically more productive.

A second study from an entirely different domain — software engineering — found the same pattern, which is what makes the augmentation thesis credible rather than a one-off.

In a controlled experiment, GitHub had developers complete a coding task (building an HTTP server) with and without its Copilot assistant. The developers using Copilot finished ~55.8% faster (the study reported a wide confidence interval, roughly 21–89%, as of 2023 — verify live), and a higher share completed the task at all (78% vs. 70%). As with the support study, the largest gains went to the less-experienced developers.

Two different functions, two different research teams, two different kinds of work — and the same directional finding: AI augments the less-experienced most, and the gains are real and measurable on tasks that sit inside the technology's frontier. When independent studies in unrelated domains converge, a leader can treat the pattern as a planning assumption rather than a vendor anecdote.

Why do novices gain most? The answer is the strategically useful part, because it tells you what AI is actually doing inside your organization.

The AI assistant in the support study had effectively absorbed the patterns of the best agents' answers and was surfacing them, in the moment, to everyone. A new hire who would have taken months to learn the right phrasing, the right troubleshooting path, and the right tone was handed that tacit knowledge instantly. In effect, AI compressed the skill gap — it moved know-how that normally lives in your top performers' heads out to the rest of the team.

This is the deepest reason augmentation beats replacement as a frame: the value is not just "a tool that types faster." It is a mechanism that diffuses your best people's expertise across your whole workforce. Top performers' know-how is usually your most valuable, least transferable asset — locked in individuals, lost when they leave. AI turns some of that tacit knowledge into something the whole team can draw on.

Reframed for the board: AI's clearest near-term value is spreading the know-how you already have, not acquiring know-how you don't.

Here is where most AI business cases quietly fall apart, and where a sharp leader earns their seat: a task-level speedup is not the same as organizational throughput.

The studies measured one task in near-ideal conditions. A 55.8% faster coding task or a 14% lift in issues-per-hour is a true result — for that task. But the value your organization actually banks is the work that flows all the way through to a shipped, correct, integrated outcome. Between the fast task and the realized value sit several things the studies did not count:

• Review. Faster drafts still need human review — and AI output, which never says "I don't know," often needs more scrutiny, not less.
• Integration. A function that's quicker in isolation can stall at the handoff if the next step in the workflow wasn't redesigned around it.
• Rework. Output that looks done but is subtly wrong creates downstream rework that can erase the upfront time saved.
• The bottleneck may be elsewhere. Speeding up a step that wasn't the constraint produces no net throughput gain at all.

This is why so many organizations report local speedups that never show up in the financials. The honest executive position is: treat task-level numbers as evidence the capability is real, then ask the harder question — does this speedup survive review, integration, and rework to become value the business can actually count?

Put the evidence together and a clear strategy choice emerges. The augmentation case rests on three durable findings: AI lifts the less-experienced most, it works by transferring expert know-how across the team, and its task-level gains only become value when you redesign the work around them. None of those findings supports a wholesale replacement play.

So the strategic posture is to design roles around human + AI, not to swap one for the other:

Replacement is the riskier bet on every axis, and — as the canonical Klarna case showed — it can reverse: Klarna replaced roughly 700 customer-service roles with AI in 2023, looked excellent on volume metrics, then saw customer-satisfaction quality deteriorate and rehired humans by mid-2025, with its CEO conceding the company had focused too much on cost at the expense of quality. The augment-and-blend approach the evidence supports would likely have avoided that round trip.

The practical implication: aim AI at lifting people, redesign the role and the workflow around the human-plus-AI pairing, keep a human accountable for judgment and quality, and measure the outcome, not the task.