LLMs as a Development Tool: An Honest Assessment
After using AI tools heavily in day-to-day engineering work, here is where they genuinely help, where they create more work than they save, and what we have changed our minds about.
The hype around AI coding tools has been loud enough that honest assessments are quite rare. People tend to either evangelise enthusiastically or dismiss everything as overblown. Having used these tools heavily in real projects for over a year, neither of those positions reflects what we have found.
Where They Genuinely Accelerate Things
Boilerplate is the obvious one. Generating the skeleton of a new API endpoint, a React component with standard props, a migration file with the correct syntax — LLMs handle this well and it is time you would rather not spend.
Less discussed but equally useful: exploring unfamiliar libraries. Asking an LLM to show you how to do something specific with a library you have not used before is often faster than reading documentation, particularly for libraries with sparse or inconsistent docs. You still need to verify the output, but verifying is faster than searching from scratch.
Test generation is another genuine win. Given a function with clearly defined behaviour, a good model will produce a solid set of unit tests that cover the happy path and a reasonable set of edge cases. The tests need review, but starting from something is faster than starting from nothing.
Where They Create More Work Than They Save
Complex debugging is the area where expectations most consistently outpace reality. Pasting a stack trace and a description of unexpected behaviour into a chat interface sometimes produces the answer immediately. More often, it produces a plausible-sounding suggestion that does not quite fit and takes time to evaluate and dismiss.
The problem is that debugging well requires deep context about the specific system, and that context is hard to convey in a prompt. A senior engineer who knows the codebase will outperform an LLM on non-trivial bugs almost every time.
Architecture decisions — where to put something, how to structure a new system, what tradeoffs to make — do not benefit much from AI assistance either. The model will generate a reasonable-sounding answer, but it lacks knowledge of your specific constraints, team, and codebase that makes architectural advice actually useful.
What We Have Changed Our Minds About
We expected code review assistance to be more useful than it has been. LLMs catch style issues and obvious mistakes but miss the subtle logic errors and architectural concerns that matter most in a review.
We did not expect documentation assistance to be as good as it is. Writing clear explanations of complex systems — for internal docs, README files, handover materials — is an area where these tools add real value, particularly for engineers who find technical writing tedious.
The honest position: these are good tools that are improving quickly. They raise the floor, making it easier to produce adequate code faster. They have not yet raised the ceiling for what excellent engineers can do.
Curious about how AI tooling fits into a proper engineering workflow? Get in touch and we can talk through what makes sense for your team.