# Where AI is Superhuman: The Right Jobs to Automate with LLMs

Andy Tadman, Theory Ventures | AI Engineer Summit 2025 | 11:53

Source: https://www.youtube.com/watch?v=R2VfIjuWhgw
Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks.
Page: https://aietalks.com/talks/where-ai-is-superhuman-the-right-jobs-to-automate-with-llms
Published: 2025-02-22
Tags: enterprise, product-strategy, team-adoption, workflows

## TL;DR
- LLMs are strongest at transforming information, synthesizing large amounts of data, and reasoning over patterns represented in their training data.
- High-volume, relatively low-complexity jobs are the most likely to be disrupted because AI can repeat the same task at a scale humans cannot match.
- Companies should break jobs into tasks, assess their volume and complexity, and account for the surrounding team structure before automating them.

## Summary
Andy Tadman presents Theory Ventures' framework for deciding where LLM automation can support startups and enterprise investment. He separates work by volume and complexity. Strategic work is complex and low volume, so LLMs will mostly act as copilots. Mid-range workflows can receive substantial automation while humans review or handle escalations. High-volume, relatively simple work is the strongest fit for end-to-end automation because AI can repeat tasks indefinitely and can outperform rules-based software. Tadman illustrates this with Dropzone AI investigating security alerts and Amplitude personalizing customer engagement. Automation changes jobs as people review outputs, maintain systems, and take on work that routine operations previously displaced. Organizations may shrink or move from pyramid structures toward inverted pyramids or diamonds, creating problems for entry-level hiring and training. The framework helps identify technology-workflow fit, although Tadman says business success also depends on pain severity, incentives, and market structure.

## Key ideas
### LLMs are useful because they transform, synthesize, and reason over information
[01:12](https://www.youtube.com/watch?v=R2VfIjuWhgw&t=72s)
Tadman identifies three properties that matter for workflow automation. Transformation converts information between formats, such as a PDF into a spreadsheet or an email into JSON. Synthesis distills large amounts of information or answers a question, as with deep research systems that summarize hundreds of websites. Reasoning approximates human decision-making from written reasoning data. This works well where the relevant patterns exist in training data, including common support questions, code, math, and logic. More specialized reasoning, such as how a lawyer or security analyst handles a case, needs more bespoke data collection.

### Automation analysis starts by breaking a job into small tasks
[02:41](https://www.youtube.com/watch?v=R2VfIjuWhgw&t=161s)
Tadman says a job should be analyzed at the task and subtask level. In security operations, an analyst might receive an alert, run a query, research the result, transform or synthesize information, analyze it, and repeat the process before reaching a conclusion. Even querying can be split into choosing a tool, writing a query, and debugging errors. A job can be performed by several people or by systems configured by people. The job also includes meetings and other interpersonal work, which affects how automation changes the organization.

### Complex, low-volume work will mostly use LLM copilots
[03:31](https://www.youtube.com/watch?v=R2VfIjuWhgw&t=211s)
Tadman places jobs on a spectrum from complex, low-volume work to high-volume, relatively simple work. Strategic planning sits at the complex end because it involves context, priorities, and coordination, much of which is unwritten. He expects humans to remain in charge of these workflows for the foreseeable future, using LLMs to delegate or speed up selected tasks. The models may help with parts of the work, but they will not generally drive the entire workflow.

### Mid-range workflows can be automated while humans handle exceptions
[04:39](https://www.youtube.com/watch?v=R2VfIjuWhgw&t=279s)
Core workflow automation sits between copilots and complete job replacement. In this category, LLMs can automate substantial or end-to-end workflows, but the systems require expert configuration. Humans still review individual tasks or take escalations. Tadman says the firms Theory Ventures has researched often find that 40 to 70 percent of day-to-day work can be automated in this category. The remaining human work depends on domain expertise and the ability to configure and supervise the system.

### High-volume, low-complexity work is where LLMs can be superhuman
[05:15](https://www.youtube.com/watch?v=R2VfIjuWhgw&t=315s)
Tadman expects the greatest disruption in jobs dominated by large queues of relatively simple tasks. These teams often struggle with volume and rely on rules-based automation. An LLM system does not become slower when it repeats a task 10, 100, or 1,000 times per day. That changes the comparison for new AI products, since they compete with older rules-based software as well as human labor. The system may not need to match a human on every difficult case. It can still be valuable if it handles the volume better than existing rules.

### Dropzone AI applies agentic investigation to security alerts
[06:07](https://www.youtube.com/watch?v=R2VfIjuWhgw&t=367s)
Tadman describes Dropzone AI as a system for security operations. As companies add security products, they receive more alerts and need analysts to decide whether each one is real, requires remediation, or is a false positive. Analysts may inspect only 1 percent of alerts, while rules-based systems filter the rest and become difficult to maintain across edge cases. Dropzone AI performs end-to-end investigations, writes and checks queries, provides continuous coverage, and shares learnings across customers. Tadman says its advantage is that it only needs to outperform the existing rules-based process, since analysts cannot review the full alert volume.

### Agentic customer engagement can test personalisation at scale
[07:33](https://www.youtube.com/watch?v=R2VfIjuWhgw&t=453s)
Tadman uses Amplitude to illustrate customer engagement automation. Marketing teams can create rules for new users or abandoned carts, but different preferences, channels, timing, and interests create too many combinations to manage manually. Amplitude's system tests what to send, to whom, and when, while measuring effects over time. Marketers become experimentalists who create message variants and let the system distribute them. Tadman gives the example of a food delivery company that discovered a group of users responding to 11 p.m. or midnight messages, revealing a late-night-snacker cohort the company had not previously identified.

### Automation may shrink teams and remove many entry-level roles
[09:32](https://www.youtube.com/watch?v=R2VfIjuWhgw&t=572s)
When high-volume work is automated, employees spend more time reviewing outputs, maintaining workflows and data systems, and doing strategy. Those activities require experience and expertise, while many current organizations have pyramid structures with large numbers of junior individual contributors doing routine work. Tadman expects some teams to shrink because fewer people are needed. The remaining structure may look more like an inverted pyramid or a diamond, with more advanced and managerial positions. This creates a practical problem for hiring and training when entry-level roles become less common.

## Notable quotes
- "High volume relatively low complexity jobs will be the most transformed by llms because that's where they're already superhuman." (05:15)
- "Your competition is no longer AI versus human it's AI versus previous generation of rules-based software." (05:32)
- "They just need to be better than the rules-based systems which is pretty easy." (06:58)
- "Organizations will instead look like inverted pyramids or diamonds." (10:30)
- "You need to break jobs down to their fundamental tasks, understand that jobs exist on a spectrum of complexity and volume, and that AI will be most disruptive in high volume low complexity tasks." (11:07)

## Tools & references mentioned
- Theory Ventures
- Dropzone AI
- Amplitude
- Gemini
- OpenAI
- Stack Overflow

## Who should watch
- Founders choosing which workflow to build an AI company around will get a way to compare task volume, complexity, and existing rules-based software.
- Executives planning automation for a business function will find a framework for deciding where humans remain in control and where teams may shrink.
- Engineering and operations leaders dealing with large queues of alerts, cases, or messages can compare their work with the examples from security and customer engagement.

## Related talks

- [The LLM Triangle: Engineering Principles for Robust AI Applications](https://aietalks.com/talks/the-llm-triangle-engineering-principles-for-robust-ai-applications) (Almog Baku, 26:19)
- [Real ROI: Lessons from Enterprises That Have Already Succeeded with LLMs at Scale](https://aietalks.com/talks/real-roi-lessons-from-enterprises-that-have-already-succeeded-with-llms-at-scale) (Raza Habib, Humanloop, 20:01)
- [12-Factor Agents: Patterns of Reliable LLM Applications](https://aietalks.com/talks/12-factor-agents-patterns-of-reliable-llm-applications) (Dex Horthy, HumanLayer, 17:06)
- [Building AI Agents that Actually Automate Knowledge Work](https://aietalks.com/talks/building-ai-agents-that-actually-automate-knowledge-work) (Jerry Liu, LlamaIndex, 17:57)
- [Make Your LLM App a Domain Expert: How to Build an Expert System](https://aietalks.com/talks/make-your-llm-app-a-domain-expert-how-to-build-an-expert-system) (Christopher Lovejoy, Anterior, 19:18)