The more powerful AI becomes, the more it needs to "kill" its past self.

AI agents have performed as well as or better than humans in most tasks, and the proportion is still rising. In the face of this trend, Goldman Sachs executives suggest: Don't be obsessed with holding on to old skills, but have the courage to let go. Future competitiveness doesn't lie in competing with AI in skills, but in using judgment, intuition, and values to harness AI. Only by letting go of the past can you embrace a brand - new professional identity.

Recently, a senior banker asked me a question: "As AI is becoming more and more powerful at an astonishing speed, which 10% of my job should I hold on to - the part that AI can never replace and will keep me at an advantage?"

In many ways, this is a reasonable question. In the past few years, AI has evolved from "good at simple tasks and not good at complex tasks" to "good at simple tasks and also relatively good at complex tasks", which means it can indeed be very useful in real life.

A benchmark test called GDPval by OpenAI compared the performance of agent models with that of humans in 1,320 tasks across 44 occupations in the 9 industries that contribute the most to the US GDP. The results show that agents based on the most advanced models (as of the time of writing) perform as well as or better than humans in 80% of cases. Six months ago, this figure was about 50%. Looking to the future, it seems destined to continue rising.

This poses an obvious problem for both companies and employees: It means that many of the skills we've accumulated in our careers may soon be performed by AI agents. Given the uncertainty of the future, people naturally look for familiar ground to stand on and hold on to their most trusted habits. Since our experience and expertise are often the capital that allows us to climb in our respective fields, the thought that these might disappear overnight is truly terrifying.

What I told the banker surprised him: Let go of that 10%. Have the courage to let old habits die, so that you can be reborn professionally and embrace a brand - new 100%, even if it's completely different from everything you've learned in the past.

Not all industries will be disrupted in the same way or on the same timeline. But if you're in one of the occupations identified in the GDPval report as most likely to be disrupted - such as developers, lawyers, and property managers - then how to adapt is an urgent issue. It's time to stay curious, keep an open mind, and be willing to let go of even the most successful professional habits, while holding on to those human qualities that won't change, such as your intuition, your judgment, and your values.

AI has made great progress in the past year

In the past year, generative AI has evolved from a useful tool similar to Google Search - saving users time but not fundamentally changing the way we work - to a technology capable of mimicking human reasoning, formulating plans, and taking actions. Delegating tasks to AI agents with minimal human intervention is becoming increasingly feasible, such as conducting fundamental research on companies, creating discounted cash - flow models, filling out forms, or solving simple customer support cases. As AI agents are widely adopted in organizations, they are clearly changing the way people work.

In a corporate environment, these agents can continuously improve through interaction with humans and feedback from internal evaluation benchmarks. For example, a research and analysis agent can learn the most reliable information sources, how to weigh them in the overall context, how to apply any company abbreviations and jargon, and crucially, how to make micro - decisions independently in the face of conflicting information - very similar to what experienced employees do.

This evolution requires a shift in the mindset of human users. It requires them to trust these agents and selectively learn when to relinquish control, transforming from operators to supervisors. At the most basic level, this requires a profound reflection on one's own habits.

In this context, my advice to the banker is to resist the impulse to have direct control over every step, such as creating every line of content in a presentation. Now, his task is to focus on providing clear instructions so that the agent can operate more effectively towards the goal and ensure that appropriate control measures are applied systematically and consistently, so that he can let the agent safely perform tasks on his behalf. Transform from an individual contributor to a supervisor and a mentor.

This is the brand - new 100%.

Why you should focus on judgment

Think about the question of which skills people should retain for career survival from a different perspective: Imagine an experienced rider learning to drive a car. Can 10% of the skills learned from horse - riding be retained for driving a car? Probably not a single one. Then, what are the skills that need to be 100% mastered to become a good driver? Their reaction speed and intuition.

Bankers are used to receiving many questions from customers - usually very complex ones. For example: How will the recently announced tariffs affect the companies in my portfolio, and how can I hedge against this risk? Giving meaningful answers requires collecting information, verifying information, formulating strategies, and then discussing with the customer. This usually happens within a few hours or days after the customer asks the question - in the jargon of trade settlement, it's t + 1 or t + 2.

In the future, with agents working in the background, we may be able to give answers at the t - 1 moment - even before the customer asks the question. Imagine a banker receiving a briefing email in the morning: These are the key events that occurred overnight, how they may affect the following customers, and here are some possible strategies and discussion points.

The banker's value here lies in evaluating the suggestions, using judgment, discussing with the team and the agent, and finally calling the customer before the customer contacts you with a question. Just like an experienced driver making full use of traction control and auxiliary braking systems when driving on a mountain road in bad weather.

The situation in practical applications

The new challenge is not just about optimization, but about rethinking our roles and the company. Don't just relearn skills; reimagine skills and build new habits. Consider the hybrid workforce composed of agents and humans as the new normal, and restructure your company around this assumption.

This requires several basic elements:

1. Leadership

Letting go of old habits doesn't happen naturally. It requires strong leadership and a top - down approach to hold people accountable for change.

Based on my experience, this is the most challenging of all tasks. You're driving a fundamental change - a transformation. Applying AI to simplify old processes, doing more of the same things faster can bring temporary relief, but in the long run, it will ultimately deviate seriously from the goal. In fact, such large - scale change management requires top - level leaders to be committed to a deep - seated change, which is impossible to achieve without a complete change in the way of working. If you want your developers to change their habits, ask them to triple their productivity, not just increase it by 20%. If you want to prevent candidates from cheating in interviews with the help of AI, assign tasks that are extremely difficult and can only be completed by those proficient in AI. For example: Create a working Excel clone in three hours.

If you want to simplify the procurement - to - payment process, the goal should be to reduce the points of human intervention by 90%, not 20%. Even if you only achieve half of it, you know that your team has at least gone through a thorough rethinking, not just optimization.

2. Clear goals and results

If we don't know what good performance is, neither humans nor AI will know how to take the right steps towards success. We must attach great importance to evaluation and benchmarking.

Most companies view tasks as a series of step - by - step actions. They codify them into standard operating procedures. Then they build control measures on top of these procedures. In real life, organizational processes and decision - making are more like the "Garbage Can Model" - to some extent, chaotic, accidental, and non - linear.

At Goldman Sachs, when applying AI to established company - wide processes (such as customer onboarding), we first focus on clarifying what good performance is, drawing on process quality indicators and the decisions of experienced operators, and then creating a set of evaluations to compare the output of agent AI with the expected results.

By establishing an appropriate feedback loop, AI will improve itself until the output meets your results. Just like you would tell a map application to guide you to your destination by the fastest route and avoid bridges, rather than telling it how many left or right turns to make, and provide feedback at the end of the journey. Move from step - by - step, rigid rule - based process execution to a results - based agent system that can make small decisions autonomously under human supervision.

3. Master your own data

Agents cannot operate without context. They will degenerate into chatbots. Data is the lifeblood of context - the real situation of your organization - and also the guiding map for humans and autonomous operating programs. Without this real situation, there can be no clear direction.

My experience in this area is that AI transformation follows data transformation, not the other way around. In many companies, data is scattered, mapped to multiple unrelated ontologies (imagine books in a library, some arranged by author, some by subject, some by ISBN number, all randomly mixed on the same bookshelf), and this data is repetitive and outdated. The most typical problem with AI is "garbage in, garbage out" - and it can make the garbage output look quite convincing.

Therefore, leaders may want to postpone (which is a very unpopular concept these days) the implementation of large - scale AI projects until their data is in order. This may take months or years, and the data preparation situation is very useful for determining the priority use cases for AI transformation.

What does this mean in terms of changing habits?

Resist the temptation to directly accept the output of AI. Check the sources, supervise, and verify the output, or if you've only relied on the results of your own work until now, learn to do so. A future driven by agents requires everyone to become a manager in a sense.

After all, that's what it boils down to. Personal change is even more difficult. Having the courage to let go of our trusted habits and embrace a brand - new, complete professional identity to thrive in is one of the greatest challenges for every practitioner today.

This article is from the WeChat official account “Harvard Business Review” (ID: hbrchinese). Author: HBR - China. Republished by 36Kr with permission.