Quick Answer
Our experience has been that AI is often excellent at a well-defined task and less dependable when expected to preserve the intent of an evolving project across many decisions without a clear external source of truth.
That's not a contradiction.
Writing a section, fixing code or analyzing a document is a bounded task.
Managing a project requires maintaining priorities, history, dependencies, changing requirements and the distinction between temporary experiments and permanent decisions.
Those are different problems.
The Individual Task Can Be Excellent
We've had AI produce useful articles, working code, strong visual concepts, research summaries, SEO analysis, and practical troubleshooting.
That's why we kept using it.
The surprise was that many excellent individual outputs did not automatically produce an excellent overall workflow.
Projects Have State
A project contains things that aren't obvious from the immediate request.
Why did we choose this architecture?
Which earlier rule was abandoned?
Was this experiment temporary?
What are we optimizing for?
Which tradeoff did we already accept?
What must remain compatible with earlier work?
Those questions involve state.
If the relevant state isn't clearly available, a perfectly competent task response can move the project in the wrong direction.
Local Optimization Can Hurt the Whole
Imagine asking AI to improve one article according to a validator.
It may improve the score.
But if the validator is measuring the wrong proxy, the article may become worse for readers.
The AI completed the assigned task. The project still lost ground.
That distinction matters.
We Initially Expected Continuity to Solve More Than It Did
Long conversations preserve useful context.
But they also preserve abandoned ideas, old instructions, failed experiments, and conflicting versions of the project.
More history isn't necessarily cleaner state.
We eventually learned to distinguish conversation history from current project truth.
Humans Need Structure Too
This isn't uniquely an AI problem.
A human team without current requirements, ownership or project documentation can drift too.
AI simply makes execution fast enough that drift can produce a lot of output before someone notices.
Our Better Division of Labor
AI can do large amounts of execution.
The project should still have an explicit external definition of goal, current standard, constraints, locked decisions, and acceptance criteria.
Then individual AI tasks operate inside that framework.
Worthy Trying Verdict
AI task execution: Often excellent
Autonomous long-project continuity: Requires structure
Main risk: Locally correct work that moves the overall project incorrectly
Human oversight: High for project direction
Would we use AI on another large project? Absolutely
The lesson: Don't confuse being excellent at the next task with knowing which task the project should do next.