Every product roadmap has an "add AI" line item now, and the first real decision isn't which model to use — it's whether to build anything at all. A surprising number of AI initiatives should end with a ChatGPT Team subscription and a well-written prompt, not a development project.
When an off-the-shelf tool is genuinely the right call
If the task is a person doing something on a computer — drafting emails, summarizing documents, writing first-pass copy, researching a topic — a general assistant like ChatGPT or Claude, used well by a trained team, usually beats a custom-built internal tool on cost and speed to value. Building software to replace a task a $20 subscription already handles is solving a problem that doesn't exist.
When it's actually worth building
Custom AI development earns its cost when the AI needs to act inside your existing product and data — answering support questions from your actual documentation and order history, extracting structured data from documents your customers upload, or automating a workflow that spans multiple internal systems no general assistant has access to. The signal is integration: does this need to plug into things a chat window can't reach?
The middle ground most teams miss
Between "use ChatGPT manually" and "build a custom AI product" sits automation: wiring an existing AI model into your existing tools with scripts and integrations, without building a bespoke application around it. This is often where the real, fast return lives — a document that used to take an hour to process now takes two minutes, without a multi-month build.
The question that actually decides it
Does this need to run automatically, at scale, integrated with data a general chat tool can't see? If yes, build. If a skilled person with the right prompt and twenty minutes could do this today, you don't have a build problem — you have a training and workflow problem, and it's far cheaper to solve.
Where to start
- List the actual repetitive task before naming a technology — "we want AI" is not a task.
- Try it manually with an off-the-shelf tool first, even for a week, before committing to a build.
- Build only the parts that genuinely require integration with your own data or systems.
- Measure time saved before scaling the automation further.
