The pattern has become predictable
A new AI capability becomes available. Within weeks, it is integrated into every product that can plausibly touch it. Customer service gets an AI chatbot. The email tool gets an AI writer. The analytics platform gets "AI insights." The search bar gets "AI-powered suggestions." Most of it makes the product marginally worse to use while allowing it to say "AI-powered" in the marketing materials.
This is not cynicism about AI. It is a description of what happens when the adoption question becomes "how can we add this" rather than "should we add this, and where."
The useful question
Before adding AI to any workflow, product, or process, the question worth asking is: what does the person using this actually need to do, and does AI make that easier or does it add a step?
An AI chatbot on a support page that can answer common questions instantly is useful. The same chatbot that makes users interact with a bot for five minutes before reaching a human — when what they needed was a phone number — is not. The technology is identical. The outcome is completely different because the first one was designed around the user's actual need and the second one was designed around deflecting support volume.
Where AI integration genuinely adds value
AI adds genuine value where it handles work that is high-volume, repetitive, and where the cost of occasional errors is low. Classification, summarisation, first-draft generation, pattern detection in large datasets — these are tasks where the economics of AI are compelling and the downside of occasional imprecision is manageable.
It adds genuine value in tools where the user has already demonstrated intent and is asking for help completing a task they understand. A code autocomplete that surfaces the right pattern. A writing tool that restructures a paragraph on request. An analytics tool that flags unusual patterns worth investigating. These are cases where AI reduces friction on a task the user was already doing.
Where it usually does not
AI integration tends to produce poor outcomes when it is added to a process that requires judgment, domain expertise, or accountability. An AI that writes the first draft of a legal document is useful if a lawyer reviews it carefully. It is a liability risk if the assumption is that the AI is reliable enough to reduce the review time significantly.
It also tends to produce poor outcomes when it is added between the user and what they actually want. Every AI-generated summary layer that sits between a user and the source information it summarised degrades the user's ability to evaluate what they are being told. The summary is convenient until it is wrong, and then it is misleading.
The integration principle worth following
AI should reduce work, not add a step. If implementing AI makes a workflow longer, more complex to maintain, or harder to evaluate — even if it looks impressive — that is a signal the integration is solving the wrong problem.
The organisations that will use AI most effectively are not the ones who move fastest to add it everywhere. They are the ones who are deliberate about where it genuinely removes friction for real users, and patient enough to implement it properly in those places rather than superficially everywhere.
A practical test
Before adding AI to a product feature or internal workflow, ask three questions: What specific task does this make faster or better? Who is the person doing that task, and does this actually help them? What does failure look like, and is the downside of that failure acceptable? If the answers are clear and reasonable, proceed. If they are vague, it is probably not ready.