I get pitched to build AI into projects that don’t need it. It usually comes from a good place: someone read an article or saw a demo, and wants their business to “have AI” the way it might have a website or a CRM.
I understand the instinct. I also think it’s often the wrong call, and I’d rather tell a prospect that up front than take the project and build something overcomplicated.
Here’s where I steer people away from AI, and toward something simpler instead.
Low-volume, one-off tasks
AI shines when you’re doing something repeatedly, at a scale where a small time savings per instance adds up. Run a task twice a year and the setup and maintenance cost of an AI-driven solution usually outweighs the time it saves. A spreadsheet formula, a five-minute manual process, or a checklist gets you there faster and with less to maintain.
I see this most with reporting. A business wants an AI system to generate a monthly summary that takes fifteen minutes to write by hand.
The AI version needs a data source, a prompt, a review step, and someone to notice when the output drifts. Fifteen minutes a month doesn’t justify that.
Predictable workflows don’t need a model
If a process follows the same steps every time, with the same rules, AI is the expensive way to solve it.
A new order triggers a notification. A form submission creates a record in your CRM. A signed contract moves a deal to the next stage.
Each of those is a sequence of if-this-then-that steps, and tools built for exactly that have existed for over a decade.
Zapier and Make handle this kind of work reliably and cheaply, without the unpredictability that comes with a model in the loop. They’re also easier to audit: if a workflow breaks, you can look at the steps and see exactly where it failed. An AI-driven version of the same task adds cost and a layer of “why did it do that” that a fixed workflow never has.
That holds at low to moderate volume, where these tools bill per task. A high-volume workflow, say a contact form that fires thousands of times a month into a CRM, can cost more over a year than a one-time custom endpoint that does the same job. And if your CRM already accepts form submissions natively, you may need neither.
I built a loyalty program for a wellness brand this way: members earn points for purchases, subscription renewals, consultations, and completed assessments, all fixed rules they can track on their own dashboard. None of it needs a model deciding anything.
I’ll build the AI version when the task requires judgment, like reading an email and deciding what it’s about. When it doesn’t, I’ll say so and point you at the simpler tool.
Decisions that carry real consequences
I’m cautious about AI making calls that carry financial, legal, or safety weight: pricing decisions, contract terms, medical or legal guidance, anything where a wrong answer creates liability. AI can draft a first pass or summarize options, but the decision should stay with a person who understands the context and can be accountable for it.
I built identity verification into a financial services application this way: it screens applicants step by step and sends sensitive data straight to the verification service without storing it. The rules are fixed and auditable, and a person owns the outcome. No model deciding who gets through.
Why I bring this up
I use AI tools every day in my own work. I write code with them, lean on them for research, and build AI-powered features for clients when the task calls for it. None of that means AI belongs everywhere.
Part of doing this work honestly is telling a prospect when the simpler, cheaper option wins, even when it’s not the more exciting pitch. That might be a no-code tool, a few lines of code against a system they already run, or nothing new at all. I’d rather tell you that than sell you AI you don’t need.
