AI Isn't the Team

What it takes to create complex things with AI

AI makes it possible for one person to do things that once required a team.

You can write a book, build an application, produce a video, edit audio, and automate much of the work that connects them.

I've spent the past several months using AI to publish two books and build software to support my creative work. I found that producing the individual pieces was often easier than bringing them together into a coherent product.

AI can perform the roles. That doesn't automatically make it a team.

Completing tasks isn't the same as building a product

Consider the difference between completing a set of tasks and building something from them.

TASK COMPLETION               SYSTEM COHERENCE
    ┌─────────┐               ┌──────────────────────────┐
    │ Task A ✓│               │         PRODUCT          │
    └─────────┘               │                          │
                              │   ┌─────────┐            │
    ┌─────────┐               │   │ Task A  │──────┐     │
    │ Task B ✓│               │   └────┬────┘      │     │
    └─────────┘               │        │           ▼     │
                              │        ▼      ┌─────────┐│
    ┌─────────┐               │   ┌─────────┐ │ Task C  ││
    │ Task C ✓│               │   │ Task B  │◄┤         ││
    └─────────┘               │   └─────────┘ └─────────┘│
                              │                          │
                              └──────────────────────────┘

In the first case, every task succeeds independently. In the second, the parts work together toward a shared purpose.

Publishing a book involves more than writing. There's editing, visual design, formatting, audio production, and distribution. Each has its own requirements, and decisions in one area can affect the others.

While working on my books, I had to track details across the manuscript to maintain consistency. A change that improved one chapter could create a contradiction somewhere else. Fixing individual passages wasn't enough. I had to consider the book as a whole.

I encountered a similar problem building software.

I started with scripts that automated specific tasks. As I added functionality, they grew into a larger system. I eventually organized capabilities behind an MCP server and command-line interface, then added a web interface to interact with items such as caption position and color.

The interface made those adjustments more accessible, but it also introduced another layer to maintain. A feature that made sense on its own didn't automatically make the overall tool simpler or more reliable.

The quality of individual parts doesn't guarantee the quality of the whole.

Why AI doesn't automatically close the gap

AI can implement a software feature from a specification or improve an individual chapter of a book. It may do both successfully without improving the product as a whole.

The problem isn't necessarily the execution. It may be what the task failed to account for.

A software feature might work perfectly while making the architecture harder to maintain. A revised chapter might read better while weakening the book's overall argument.

In both cases, the AI may have done exactly what was requested. The request simply didn't account for everything that mattered.

Complex work requires someone to define success, understand dependencies, evaluate trade-offs, and check whether the finished product accomplishes its purpose.

As execution becomes cheaper, these responsibilities become more important to the person directing the work.

Directing the whole

Directing complex work with AI comes down to three responsibilities:

Responsibility	What it means
Define	Establish the product's purpose, requirements, and standards for success.
Coordinate	Manage dependencies and ensure individual tasks contribute to the larger goal.
Evaluate	Verify results, examine how the parts interact, and reconsider decisions when necessary.

These responsibilities apply whether you're building software, publishing a book, or producing a video. They don't require separate people or AI agents. One person may oversee the entire process, and one agent may perform several roles.

The responsibilities describe what you need to manage. A practical workflow turns them into repeatable actions.

A practical method for building complex things with AI

Step	What to do	Key question
1. Identify the work	Map the responsibilities and dependencies involved in delivering the product.	What might otherwise be overlooked?
2. Define success	Set constraints and quality standards beyond task completion.	Will this improve the product, not just the task?
3. Delegate	Give AI bounded tasks with clear context and specifications.	Can it work independently without guessing at important decisions?
4. Evaluate	Verify the output and check how it fits into the whole.	Does it work, and does it work with everything else?
5. Reassess	Review the structure as complexity and dependencies grow.	Are isolated fixes still solving the problem?

The process is iterative:

 DEFINE → PLAN → BUILD → EVALUATE
    ▲                       │
    │                       ▼
    └──── REVISE ◄──── Meets goals?
                         │
                         └── Yes → DELIVER

Consider adding a caption-adjustment feature to a video tool.

An AI agent could implement the controls and make them work in isolation. But the work isn't finished if the controls introduce inconsistent settings, complicate the interface, or break existing caption workflows.

A useful task specification would establish what the feature should do, where it belongs, and what existing behavior must remain intact. After implementation, you'd verify the controls themselves, test the complete workflow, and check whether the added complexity is justified.

The point isn't to anticipate every detail. It's to give the AI enough context to make good decisions within a defined scope, then evaluate the result against the needs of the product.

The amount of oversight should match the complexity of the work. A small, well-defined task may need little supervision. A decision that affects the architecture or direction of an entire product deserves more scrutiny.

If every new feature requires more workarounds, or every revision introduces problems elsewhere, another isolated fix may not be the answer. It may be time to reconsider the structure.

Learn enough to direct the work

Directing work across several disciplines raises another question: how much do you need to know about each one?

AI makes it easier to explore unfamiliar areas. While publishing my books, I had to make decisions about editing, visual design, audio production, and distribution. While building software, I had to think about architecture, testing, and maintainability.

AI helped me learn about these areas and participate in work I might otherwise have struggled to undertake.

That didn't make me an expert in all of them. It gave me a better understanding of what each contributes and what questions I needed to ask.

Exploring a discipline is not the same as mastering it.

You still need enough understanding to recognize uncertainty, evaluate the consequences of decisions, and know when a problem requires deeper expertise.

The opportunity isn't to make every skill interchangeable. It's to become capable across more of the production process while understanding where your knowledge ends.

The real advantage

AI will continue to improve at planning, coordination, implementation, and evaluation. How much oversight a project requires will change as those capabilities become more reliable.

The advantage isn't simply producing more pieces at lower cost. It's knowing which pieces are needed, how they depend on one another, and whether they work together.

AI isn't the team. It's part of how you build one.