Last month I was helping a founder think through the AI side of their user research project — twenty-something one-on-one interviews with Singapore consumers about a personal-use app. The kind of interview where, if you can't read what someone's not saying, you walk away with nothing useful.
Watching how the work was running, I came away with the clearest sense I've had in months of what's still irreducibly human in this kind of work — and what's not.
The interviews themselves: irreducible. You can't AI your way into someone's hesitation, the moment they almost say something but don't, the contradiction between what they said two minutes ago and what they're saying now. Those require a human in the room reading three things at once.
The workflow around the interviews was a completely different story. Finding venues across the island. Coordinating twelve different calendars. Drafting follow-up notes. Pulling specific moments out of hours of video into workshop-ready clips. Work that used to require a coordinator, an admin, and probably a junior video editor. We mapped together what was already doable by agents, and almost none of it survived as something the founder needed to do.
The interesting thing wasn't how much AI did. It was the contrast — the part of the work that suddenly mattered more sitting right next to the part that suddenly mattered less.
That contrast is the central professional question of this decade.
1 Shift — From "What can AI do?" to "What am I choosing not to delegate?"
For two years, the AI conversation has been mostly about capability. What can it do now that it couldn't do six months ago? Where's the frontier? Which tools should I learn?
Those questions still matter. But they've stopped being the question that determines whether you're valuable to your organization in 2026.
The new question is more uncomfortable and more honest: what am I choosing not to delegate, and why?
Because almost everything is now delegable. And what you keep — what you refuse to hand off even when AI could do it competently — is now the closest thing you have to a professional identity.
That refusal isn't sentimental. It's strategic. It's the answer to: what do I contribute that AI can't, that the people who pay me actually care about, that I would feel diminished to give up?
If you can't answer that quickly and specifically, you're not slow on AI adoption. You're slow on something more fundamental.
Insight 1 — The bottleneck has moved from generation to verification
Through 2024, the bottleneck for most knowledge work was generating output — writing the deck, drafting the analysis, producing the document. AI compressed that bottleneck dramatically. A senior consultant who used to spend three days on a strategy memo can now produce a competent first draft in twenty minutes.
But here's what happened next: the bottleneck moved.
It's now verification. Whose judgment is good enough to know if the AI output is correct, strategic, on-brand, defensible? Whose taste is sharp enough to tell a good answer from a competent-looking one?
Verification is invisible. It doesn't generate impressive output. But it's the place value is now stored.
This is why senior people with deep domain expertise are seeing their leverage increase in 2026, even as junior knowledge workers are seeing it compress. The senior person can verify; the junior person can mostly only generate. And generation is cheap.
A Stanford labor market study released this year found something most people missed: in AI-exposed occupations, workers aged 22-25 saw employment decline by 16%. Workers aged 30 and over saw gains of 6-12%. The difference wasn't AI fluency. The difference was accumulated judgment — having been wrong enough times in enough specific contexts to know, by instinct, when a clean-looking answer isn't right.
If you're early-career, this is the most important career strategy of your decade: invest in becoming someone whose verification is trusted, not just someone who generates well. Generation will continue to commoditize. Verification will continue to compound.
Insight 2 — Four levels of AI work in 2026 (and where the gap actually is)
Most discussions of "how to use AI" still default to a 2023 mental model: prompting a chatbot for individual tasks. That's now Level 1 of a much bigger picture.
Here's what the levels actually look like in non-technical enterprise work in 2026:
Level 1 — AI as a faster typist. You use a chatbot for atomic tasks: drafting an email, summarizing a document, generating bullet points. Time savings: real but bounded, 10-20% on existing tasks. Most people are still here.
Level 2 — AI as a thinking partner. You iterate with AI on structured work — pressure-testing a strategy, generating options, drafting and rewriting until the thinking sharpens. The output isn't just faster; it's better than what you'd produce alone. Senior knowledge workers tend to land here.
Level 3 — AI as a junior team member. You delegate multi-step workflows to agents. Finding venues. Scheduling. Researching. Drafting. Pulling moments out of video. The kind of operational work that used to require a small team you no longer need to staff. This is the 2026 frontier for non-technical operators. It's where the bulk of new leverage is being created — and it's also where most companies haven't reorganized yet.
Level 4 — AI as infrastructure. You've designed systems where AI runs ongoing operations and you intervene only at decision points. Custom GPTs trained on your domain. Agentic workflows you've architected. Internal AI tools that capture your team's tacit knowledge. The ones who are have a disproportionate advantage, because they've moved from using AI to owning AI as a layer of their work.
The interesting thing about these levels isn't that one is better. It's that most people are operating at Level 1 while their organizations are starting to expect Level 3. The gap between expectation and reality is where career risk is concentrating fastest in 2026.
If you're a manager, the strategic question is no longer "how do I get my team using AI?" It's: "what's the highest level any individual on my team is operating at, and what does it take to raise the floor?"
If you're an IC, the strategic question is: "what's the level of AI work I'm currently doing, and what's the next level look like for my specific job?"
What I came away from that project with
The founder I was working with is now operating at Level 3 for most of the operational layer of their business, while pouring their own attention into the Level 0 work — the work no AI can touch. The interviews. The pattern recognition across thirty conversations. The strategic call about what the product actually needs to become.
AI didn't replace any of the work they value. It cleared the space for them to do that work better.
That's the model worth aiming for. Not "AI as a productivity tool." AI as a way to get more of yourself onto the work that only you can do.
1 Reflection
If AI handled 80% of your daily work tomorrow, what would you refuse to delegate — not because AI couldn't do it, but because handing it off would mean losing something essential about why you do this work?
That refusal is your professional identity. It's worth knowing what it is, before someone else figures out you don't.
— Joe Peng, Founder, Zenotal Consulting
上个月我协助一位创始人在她的用户调研项目里（听起来很传统吧？）如何更多应用 AI。那是一场在新加坡进行的、关于一款偏私人使用的 APP 的二十多场一对一深度访谈。这种访谈，如果你读不出对方"没说"的那一部分，就什么洞察都拿不到。
观察和支持她工作的过程，让我更清楚当下这种工作里仍然不可被取代的"人"的部分到底是什么，以及可被取代的"流程"部分到底是什么。
访谈本身是不可还原的。一个人迟疑那半秒、一个差点说出口又咽回去的句子、两分钟前说过的话和现在这一句之间的矛盾：这些都需要读懂"字里行间"，AI 替代还有点难度。
而访谈周围的工作流，完全是另一回事：在新加坡找适合的场地、协调十几个不同的日程、起草跟进笔记、之后从几个小时的视频里挑出适合放进 workshop 的片段。过去，这可能需要一个协调员、一个行政、加一个初级剪辑师才能完成的工作。我们一起 mapping 出来，几乎没有一项最后还需要她亲自做。
1 Shift — 从"AI 能做什么"到"我选择不交出去的是什么"
两年来，AI 对话的焦点基本是关于能力。它六个月前不能做的事现在能做了吗？前沿在哪？我该学什么工具？
这些问题依然重要。但它们不再是决定你在 2026 年对你所在组织是否有价值的那个问题。
新的问题更不舒服，也更诚实：我选择不交出去的是什么，为什么？
因为几乎所有事情现在都可以交出去。而你留下的那部分——你拒绝交给 AI 即使它能做得过得去的那部分——是你现在最接近"职业身份"的东西。
这种拒绝不是情感的，是战略的。它回答的是：我贡献的、AI 做不到的、付我钱的人真正在意的、我要放弃会觉得自己被掏空的——是什么？
如果你不能很快、很具体地回答这个，你不是 AI 应用得慢，而是在更根本的事情上节奏慢了。
Insight 1 — 瓶颈从"生成"移到了"验证"
直到 2024 年，大多数知识工作的瓶颈在生成产出——写 deck、起草分析、生产文档。AI 极大压缩了这个瓶颈。一位高级顾问过去写一份战略 memo 要三天，现在二十分钟就能出一份过得去的初稿。
接下来发生的事是：瓶颈移位了。
新的瓶颈是验证。谁的判断力够好，能判断 AI 输出对不对、是否战略、是否符合品牌调性、是否经得起推敲考证？谁的品味够"锋利"，能在一个"看起来合格的答案"和"真的好的答案"之间分得清？
验证是隐形的，但价值就在那里。
这就是为什么在 2026 年，拥有深厚领域专长的资深人士反而看到自己的杠杆在增大，与此同时初级知识工作者的杠杆在塌缩。
斯坦福今年的一项劳动力市场研究指出一个大多数人没注意到的细节：在 AI 高暴露的职业里，22-25 岁的工作者就业下降 16%，而 30 岁以上的反而增长 6-12%。差距不是 AI 熟练度，是积累的判断力——在足够多具体情境里实操过足够多次，进而能凭直觉知道一个看起来干净的答案哪里不对。
如果你处在职业早期，这是这一代最重要的策略：让自己变成那种"验证被信任"的人，而不只是那种"生成做得好"的人。生成会继续商品化。验证会继续复利。
Insight 2 — 2026 年 AI 工作的四个层级（以及真正的 gap 在哪里）
大多数"如何使用 AI"的讨论还停留在 2023 年的心智模型：用聊天机器人做单任务。那现在只是更大画面里的 Level 1。
在 2026 年非技术性企业岗位里，四个层级的实际样子是：
Level 1 — AI 作为"打字更快的你"。用聊天机器人做原子任务：起草邮件、总结文档、生成要点。效率提升真实但有限，10-20%。大多数人还在这里。
Level 2 — AI 作为思考伙伴。你和 AI 在结构化工作上来回——压力测试一个战略、生成选项、起草再改写直到思路变锐。产出不仅更快，而且比你独自做出来更好。资深知识工作者通常落在这里。
Level 3 — AI 作为初级团队成员。你把多步骤工作流交给 agent。那种过去需要一个小团队但你现在不再需要去配置的运营工作。这是 2026 年非技术性 operator 的前沿。新的杠杆大量在这里被创造——同时这也是大多数公司还没有重新组织的地方。
Level 4 — AI 作为基础设施。你设计了让 AI 跑日常运营、你只在决策点介入的系统。不论是在你擅长领域上训练和定制化 GPT、架构 agentic 工作流，还是有了一些在你的团队里捕捉隐性知识的 AI 工具（说得不太好听是"蒸馏"）。这个层级的优势就更加显著了，因为他们已经从"使用 AI"走到了"让 AI 成为工作组成部分，并掌控它"。
这些层级有意思的不是哪一级更好。是大多数人在 Level 1 工作，而他们的组织已经开始期待 Level 3。这个 gap 是 2026 年职业风险集中得最快的地方。
如果你是经理，战略问题不再是"我怎么让团队用上 AI"，而是："我团队里任何一个人在运行的最高层级是什么？把底线抬上去需要什么？"
如果你是团队中的个体贡献者，战略问题是："我现在做的是哪一级的 AI 工作？下一级对我这个具体岗位长什么样？"
我从这个项目里带回的
我合作的那位创始人现在基本在 Level 3，把自己的注意力全部聚焦在 "Level 0"：AI 还搞不定的那种工作。譬如上面的那个访谈本身：三十场对话里的模式识别，关于产品到底应该变成什么的战略决定。
AI 没有取代任何她在乎的工作，反倒给她腾出了空间把那部分工作做得更好。
不是"AI 作为生产力工具"。AI 作为一种让你能把更多的自己投入到只有你能做的工作上的方式。
1 Reflection
如果明天 AI 接管了你 80% 的日常工作，你拒绝交出去的是什么——不是因为 AI 做不到，而是因为把那一项交出去意味着失去你做这工作的某个本质？
那个拒绝就是你的职业身份。值得在别人替你想明白"原来你没有"之前先知道它是什么。
— Joe Peng，Zenotal Consulting 创始人