When to use AI vs deterministic approach in software engineering

You don't use a hammer to drive in a screw. Stop using AI to run your deterministic processes. The AI labs want you to use AI for everything. To them, no token is a wasted token. But engineers look for the optimal solution, not the sexiest one. AI is an extraordinarily useful tool. LLMs in particular are built for reasoning through ambiguity: unstructured problems, creative synthesis, decisions that require judgment. That's where they shine. But when a workflow is deterministic, AI is overkill. Or worse, inconsistent in production. Here's a pattern I'm seeing: someone discovers they can build with AI and writes a Python script they couldn't have written themselves. Then, instead of running it on a cron job, they wrap the whole thing in an agent and call it through Claude every time. The script doesn't change. The workflow is deterministic. But the AI is still in the loop, burning tokens on execution. I catch myself doing this too. Working in Claude Code, I'll type "commit this" instead of just running the git command myself. One commit is nothing. Scaled across hundreds of developers in an org, it adds up to real money. But the cost isn't even the main problem. It's a signal. It means we're defaulting to AI for everything, including the things we already know how to do. AI excels at design, analysis, and building. Writing that Python script. Spotting patterns in data a loop would miss. But executing a repeatable, deterministic workflow? A script on a schedule does that cheaper and more reliably. The same logic applies to MCP. MCP is the right tool when the API calls that need to be made are non-deterministic. When the LLM needs to decide what data to pull or what actions to take, use MCP. But if you always pull the same data the same way, script it. And if you need AI to analyze the result, that's fine — script the retrieval, feed the output to the LLM. You get the analysis without paying for schema overhead and model reasoning on every data fetch. You still have all the other tools in the toolbox. Use them. What's your heuristic for deciding when to use AI versus a simpler, deterministic approach? I'd genuinely like to hear how others are drawing that line. #AI #SoftwareEngineering #Tokenomics #AIStrategy #Automation | 55 comments on LinkedIn