For years, firms have chased a familiar dream: automate the tedious, eliminate error, and let efficiency compound. The recent evidence points in a different direction. Automation, pursued as an end in itself, often yields brittle systems and meagre returns. The more durable gains arrive when AI amplifies human work—embedded in tasks, supervised by people, and judged by outcomes rather than technological purity.
The data are clear on two fronts. First, most corporate AI spending has yet to translate into business impact. MIT’s Project NANDA reports that only a sliver of pilots reach production with measurable results; early readouts put that share at roughly 5%, with most initiatives stalling before scale. At an enterprise level, “adoption without transformation” is the modal outcome.
Second, where AI does raise output, it does so by aiding people at work. In a large field experiment done in 2023 with more than 5,000 customer-support agents, access to a conversational-AI assistant raised issues resolved per hour by ~14–15% on average, with the largest gains for less-experienced staff—evidence of augmentation rather than replacement.
1. The Mirage of Automation vs. the Reality of Collaboration
The history of technology is littered with the allure of automation. From the industrial loom to robotic assembly lines, we have been told that the ultimate goal is to remove human beings from work. But the promise of AI as a general-purpose technology is different. It can reason, synthesise, and communicate in ways that make it less like a machine and more like a partner.
Automation promises tidy outcomes; reality delivers edge cases. The sensible aim is not to remove people, but to raise their throughput and judgment. In an article in the Atlantic in Aug 2025, David Autor and James Manyika make this case directly: AI’s promise is to enlarge the set of tasks people can do well, not to evacuate people from the loop. The historical pattern—ATMs redefining, not erasing, bank-teller work—is a caution against simplistic substitution.
2. Enterprises Are Buggy, Messy, and Political
Real-world organizations are messy: outdated systems, informal workarounds, and people invested in protecting their roles. AI deployed as rigid automation can stumble here—especially in environments with legacy processes and edge-case exceptions.
The evidence supports assistance over wholesale replacement. In the customer-support study, AI nudged performance up most for junior agents and helped them learn on the job—a pattern far more compatible with embedded “copilots” than with fully automated front lines.
What’s clear is that AI designed to assist—rather than replace—can bridge real-world complexities, adapt to human workflows, and deliver tangible value, especially where workers are already trying to do more with less.
3. Employees Are Leading Companies on AI adoption
One of the most striking insights from the MIT survey is that employees are adopting AI faster than their employers. Individual workers are already using ChatGPT, MidJourney, and other tools to draft documents, create prototypes, and analyse data. A 2024 LinkedIn Work Trend Index on the State of AI at work reported that by mid-2024, 75% of knowledge workers reported using AI at work; 78% said they were “bringing their own AI” rather than relying on sanctioned tools.
Why? Because AI makes them more productive. Workers don’t wait for perfect governance frameworks or centralised budgets; they experiment and adapt. Yet enterprises, in their zeal to “get AI right,” are tying themselves in knots with endless task forces, procurement hurdles, and compliance debates. The result: employees move ahead, companies lag behind.
This grassroots adoption is a powerful signal. It tells us employees see AI not as a threat but as a productivity amplifier. The organisations that embrace this bottom-up energy—by embedding AI in workflows rather than mandating automation from the top—will capture the upside.
4. The Generational Stakes
The sharpest labour-market effect so far is not mass displacement but the quiet disappearance of entry-level rungs in AI-exposed jobs. Using millions of ADP payroll records, Stanford’s “Canaries in the Coal Mine?” documents a ~6% decline in employment for 22–25-year-olds in the most AI-exposed occupations between late-2022 and July-2025, even as employment for older workers in the same roles rose 6–9% (see chart below). The impact is concentrated in roles where AI automates core tasks; where AI augments, employment holds up.
This matters because entry-level tasks are where expertise is built. If junior lawyers never draft contracts, or junior accountants never reconcile ledgers, how will they grow into trusted advisors? AI isn’t just reshaping today’s jobs; it’s reshaping tomorrow’s expertise pipeline.
The challenge is to think of AI as an “and,” not an “or”—a way to accelerate skill-building, not eliminate it. Otherwise, we risk producing a generation of workers with credentials but little practical judgment.
5. The Erosion of Human Judgment
Automation also corrodes expertise in the here-and-now. As Autor and Manyika warn in “A Better Way to Think about AI”, AI systems that fully replace tasks risk eroding the human judgment needed to intervene when it matters most.
History offers examples: airline pilots overly dependent on autopilot systems, traders deferring to algorithms that exacerbate flash crashes, hiring managers outsourcing judgment to biased models. When humans stop practicing judgment, they lose the ability—and confidence—to exercise it in moments of crisis.
6. The ROI Trap of Automation
Perhaps the most underappreciated point: automation demands perfection. A self-driving car cannot be “mostly safe”; a compliance engine cannot be “mostly correct.” The cost curve for AI accuracy is steep: Reliability engineering shows that each additional “nine” of availability (from 99.9% to 99.99%, and so on) requires order-of-magnitude increases in engineering effort, redundancy and validation—steep costs for diminishing customer benefit.
Collaboration, instead, thrives on 80–90% accuracy, enabling significant productivity at a fraction of the cost. A junior analyst with AI support can be 80% right and still deliver value, because the human fills the gap. A doctor with AI diagnostic assistance does not need the machine to be flawless, just helpful. This makes collaboration cheaper, faster, and ultimately more sustainable than chasing the elusive dream of error-free automation.
7. The Way Forward
So what should companies do? A few guiding principles emerge from the evidence:
- Embed AI in workflows, not outside them. Collaboration succeeds when AI meets employees where they already work.
- Encourage bottom-up adoption. Employees are showing the way; enterprises should follow.
- Protect the pipeline of expertise. Ensure that junior employees still get apprenticeship and exposure to learning foundational skills.
- Treat AI as an “and.” Pair human judgment with machine scale, rather than choosing one over the other.
- Aim for productivity, not perfection. Accept that AI will make mistakes—and design systems where those mistakes are tolerable.
Conclusion: Making AI an Ally
The automation dream has made AI look like the enemy—an unstoppable force poised to deskill workers, hollow out expertise, and impose unrealistic expectations of perfection. But a collaborative vision changes the story.
AI as an assistant, as a partner, as an “and”—that is where the real impact lies. It makes employees more productive today and sustains expertise for tomorrow. It preserves human judgment while extending human capacity. And it avoids the false trap of perfection by thriving on imperfection.
Automation may be seductive, but collaboration is sustainable. If businesses want to unlock AI’s promise, they must stop trying to replace their people and start helping them succeed.