# The AI Function Clock Source: https://www.bobzhu.tech/the-ai-function-clock/ Markdown: https://www.bobzhu.tech/assets/agents/the-ai-function-clock.md Tags: Essays, AI, Systems Summary: The common claim is that AI cannot replace your job. That may be true. But jobs are bundles of functions, and the better question is which functions are already moving closer to reliable AI handoff. Feature image: https://www.bobzhu.tech/content/images/2026/04/ig_01bae110f69970670169ecd004362c8191a1683b9f3d935305.png Feature image alt: Minimal hand-drawn illustration of an open notebook with a symbolic clock gauge and small markers for video, support, code, and medical admin functions. Feature image caption: The clock tracks functions before job titles visibly change. "AI can't replace me." That's probably true. At the job level, the claim holds up more often than not. A job is a bundle: context, taste, judgement, relationships, accountability, edge cases. Most AI systems can't hold all of that together. But the job isn't the right unit to watch. Generative video series. Customer support triage. Code review. Medical admin documentation. These aren't jobs. They're functions: recurring slices of work that can be specified, produced, checked, and integrated. As those functions get cheaper, more reliable, and easier to hand off, the role around them changes. Not overnight. Not cleanly. But steadily. I wanted a way to watch that movement without pretending the future is a single date. ## A pressure gauge, not a prophecy The [Bulletin of the Atomic Scientists](https://thebulletin.org/doomsday-clock/2026-statement/) sets its Doomsday Clock at 85 seconds to midnight in 2026. It is a symbolic warning system that tracks how much pressure is building, and moves the hands when the evidence shifts. The AI Function Clock borrows that structure, not the apocalypse. It tracks individual work functions on a scale from 0 to 100, where 100 represents broadly reliable handoff to AI. Different functions move at different speeds based on capability, cost, reliability, integration, oversight burden, regulation, and social acceptance. It's not a countdown to replacement. It's a way to see which parts of work are changing faster than the job titles suggest. ## The function-level evidence is already substantial Several major research efforts now operate at the task or function level rather than the job level. That's worth paying attention to. [Anthropic's analysis of labour market impacts](https://www.anthropic.com/research/labor-market-impacts) is explicitly task-based before it aggregates to occupations. Their observed exposure measure combines theoretical LLM capability with real-world Claude usage, and finds that computer programmers, customer service representatives, and financial analysts are among the most exposed occupations. But actual AI coverage remains a fraction of what's theoretically possible. In practice, the gap between "a model could do this" and "an organisation trusts it to do this" is still large. [Microsoft's analysis of 200,000 Bing Copilot conversations](https://www.microsoft.com/en-us/research/publication/working-with-ai-measuring-the-occupational-implications-of-generative-ai/) found the most common user requests involved gathering information and writing. Knowledge work, administrative support, and sales-related communication showed higher applicability. Microsoft [explicitly warns](https://www.microsoft.com/en-us/research/blog/applicability-vs-job-displacement-further-notes-on-our-recent-research-on-ai-and-occupations/) that applicability scores do not mean elimination. That distinction is one of the strongest arguments for the functions-not-jobs frame. A function can be highly applicable to AI and still require human judgement, oversight, and accountability to complete in practice. [OpenAI's GDPval benchmark](https://openai.com/index/gdpval/) measures model performance on 1,320 real-world knowledge-work tasks across 44 occupations: legal briefs, engineering blueprints, customer support conversations, nursing care plans, slides, diagrams, spreadsheets. Frontier models are approaching expert quality on some tasks and can complete them much faster and cheaper in pure inference terms. But the benchmark is one-shot. It does not capture workplace iteration, ambiguity, client context, or oversight loops. Treating a task benchmark as a job-replacement forecast overstates what it actually measures. The [Stanford AI Index 2026](https://hai.stanford.edu/ai-index/2026-ai-index-report/economy) puts the adoption picture in proportion. Generative AI appears in at least one business function at 70% of surveyed organisations. Overall AI use reaches 88%. But AI agent deployment sits in single digits across nearly all business functions. The gap between "we use AI somewhere" and "we trust an agent to handle this function end to end" is where most organisations currently sit. Productivity gains concentrate in structured, measurable work where outputs are easy to monitor. That's not a coincidence. Those are exactly the functions where the path to reliable handoff is shortest. ## The labour effects are real but concentrated Anthropic's data shows no systematic increase in unemployment for highly exposed workers since late 2022. Mass displacement is not visible in aggregate data. But there are signals in the pipeline. Anthropic finds tentative evidence that hiring into exposed occupations has slowed for workers aged 22 to 25. The [World Economic Forum's Future of Jobs Report 2025](https://www.weforum.org/reports/the-future-of-jobs-report-2025) projects 170 million roles created and 92 million displaced by 2030, a net increase of 78 million. At the same time, [41% of employers](https://www.weforum.org/press/2025/01/future-of-jobs-report-2025-78-million-new-job-opportunities-by-2030-but-urgent-upskilling-needed-to-prepare-workforces/) plan workforce reductions where AI automates tasks, while 77% plan to upskill existing workers. Work may grow in aggregate while specific functions and entry points get squeezed. That's a harder pattern to see from the outside. By the time a job title disappears from an org chart, the functions inside it have usually been shifting for years. [Anthropic's survey of 81,000 Claude users](https://www.anthropic.com/research/81k-economics?xs=1) picks up something the aggregate data misses. People in more exposed occupations reported more concern about displacement. Early-career respondents showed higher concern. And the people experiencing the largest AI speedups also expressed the most job-displacement concern. The people closest to the tools can feel both productivity gain and threat at the same time. ## A function moving in real time Generative entertainment shows what function-level movement looks like in practice. [ByteDance's Seedance 2.0](https://seed.bytedance.com/en/blog/seedance-2-0-official-launch) shows how quickly the video function is moving: text, image, audio, and video inputs; multi-shot 15-second audio-video generation; and editing or extension workflows. [Kuaishou's Kling AI 2.0](https://ir.kuaishou.com/news-releases/news-release-details/kling-ai-advances-20-era-empowering-everyone-tell-great-stories) adds multimodal visual-language inputs and controllable edits like adding, removing, or replacing elements in a generated video. [Higgsfield](https://openai.com/index/higgsfield/) points to the production layer around the models: link-to-video workflows, multi-minute outputs, and high-volume social video creation. That's not AI replacing entertainment. It's a function moving through stages: short-form episodes, trailers, product stories, social-first series, ad creative, and previz. Each of those has its own clock position. Fully generated long-form shows still have harder bottlenecks around taste, rights, continuity, and audience trust. The function lens is what separates practical assessment from panic. ## Why the sceptics are partly right The sceptics are right about something important. Jobs contain judgement, liability, trust, and human context that do not transfer cleanly to an AI system. A customer support agent navigates frustration, exceptions, and the gap between policy and what actually needs to happen. Ticket resolution is the measurable function. The rest is harder to hand off. A financial analyst interprets ambiguity, manages client relationships, and carries accountability for the recommendation. The modelling is one function inside that. But function-level automation compounds anyway. Each function that gets cheaper and more reliable changes the economics of the role around it. You do not need to replace the whole job. You need to change enough functions that the role gets restructured, the headcount gets revised, or the entry path narrows. That's what the clock is trying to make visible. ## What moves the hands Several things push a function closer to reliable handoff: - **Capability.** Can the model do the work at a useful quality? - **Cost.** Is it cheap enough to run at scale? - **Reliability.** Does it fail rarely enough to trust without constant checking? - **Integration.** Can the output slot into existing workflows without manual rework? - **Oversight burden.** How much human review does each output still need? - **Regulation.** Do legal or compliance requirements slow adoption? - **Social acceptance.** Will the people affected trust and use the output? These move at different speeds for different functions. Customer support resolution is closer because the work is structured, the quality bar is measurable, and the integration path is well understood. Medical admin documentation is also technically close, but accountability, privacy, and local policy keep human oversight central. A function can score well on capability and still sit back on the clock if integration is painful or the regulatory environment is not ready. ## The companion app I've built a companion app that starts with four functions across this spectrum. The [AI Function Clock](https://www.bobzhu.tech/ai-function-clock/) scores each function from 0 to 100 based on how close it is to broadly reliable AI handoff. The month-level dates are my estimates based on current evidence, not objective forecasts. The confidence levels reflect how much support each position has. The first seed set is deliberately small: generative video series, customer support resolution, code review and fix loops, and medical admin documentation. The point is to make the surface easy to update as the evidence changes. The app is a living surface. Some of these functions will move faster than I expect. Some will stall on integration, regulation, or trust problems that are not visible yet. As the evidence shifts, the scores move. ## Where this leaves the question The clock is not there to make people panic. It's there to stop people from waiting until a whole job title disappears before they notice the work has already changed. "AI can't replace me" is a comforting sentence. At the job level, it's often still defensible. At the function level, the pressure is already building. Some functions inside your job are probably closer to midnight than you think. The roles will follow the functions. They usually do. ## Try this prompt Break one job into functions, not job titles. For each function, rate how close it is to AI handoff today, what evidence supports that rating, what human judgment still matters, and what would move the hand forward. Then summarize which functions are already changing and which are mostly protected for now. ## Related on this site - [AI Does Not Automatically Give You Time Back](https://www.bobzhu.tech/ai-does-not-automatically-give-you-time-back/) looks at why AI-driven speed gains do not automatically turn into reclaimed time or lighter workloads. - [AI Function Clock](https://www.bobzhu.tech/ai-function-clock/) is the companion app tracking work functions from useful tool to reliable handoff. ### Sources - [Doomsday Clock 2026 Statement](https://thebulletin.org/doomsday-clock/2026-statement/) - [Anthropic: Labour Market Impacts](https://www.anthropic.com/research/labor-market-impacts) - [Anthropic: 81,000 User Survey](https://www.anthropic.com/research/81k-economics?xs=1) - [OpenAI: GDPval](https://openai.com/index/gdpval/) - [Stanford AI Index 2026: Economy](https://hai.stanford.edu/ai-index/2026-ai-index-report/economy) - [World Economic Forum: Future of Jobs Report 2025](https://www.weforum.org/reports/the-future-of-jobs-report-2025) - [WEF: 78 Million New Job Opportunities](https://www.weforum.org/press/2025/01/future-of-jobs-report-2025-78-million-new-job-opportunities-by-2030-but-urgent-upskilling-needed-to-prepare-workforces/) - [Microsoft Research: Working with AI](https://www.microsoft.com/en-us/research/publication/working-with-ai-measuring-the-occupational-implications-of-generative-ai/) - [Microsoft Research: Applicability vs Job Displacement](https://www.microsoft.com/en-us/research/blog/applicability-vs-job-displacement-further-notes-on-our-recent-research-on-ai-and-occupations/) - [ByteDance: Seedance 2.0](https://seed.bytedance.com/en/blog/seedance-2-0-official-launch) - [Kuaishou: Kling AI 2.0](https://ir.kuaishou.com/news-releases/news-release-details/kling-ai-advances-20-era-empowering-everyone-tell-great-stories) - [OpenAI: Higgsfield](https://openai.com/index/higgsfield/)