A chatbot stapled to the corner of the product because the board asked what the AI story is.
I get asked some version of the same question a lot lately: should we add AI? Usually the honest answer is that the thing they want does not need AI at all. They have a rule they follow in their heads, or a spreadsheet someone maintains by hand, and what they actually need is a programmer to codify that logic. A few hundred lines of plain code, tested and boring, will outperform a clever model that is right most of the time. Most of the time is not good enough when the cost of being wrong is real.
The hype cycle pushes the opposite instinct. Every vendor wants to sell you intelligence, so every problem starts to look like it needs some. But a lot of business value is just clear rules applied consistently, and clear rules are a solved problem. We have been able to do that for decades. Reaching for a model first is often a way to avoid the harder, duller work of writing the rule down.
So where does AI actually earn its keep? In my experience it is when the logic genuinely cannot be written down in advance — reading messy free-text, ranking things by fuzzy similarity, drafting a first pass a person then edits. The pattern that works is the same every time: put it inside the core process, and keep a human in the loop. The model does the part it is good at, a person owns the judgment and the outcome, and the system gets better because someone is accountable for what it produces.
The pattern that mostly does not work is the bolt-on. A chatbot stapled to the corner of the product because the board asked what the AI story is. It demos well and quietly goes unused, because it was never connected to anything anyone actually does on a Tuesday. AI delivers when it sits in the workflow people already rely on, not in a novelty box beside it.
What I tell teams is to start from the process, never from the technology. Map what you do today, find the step that is slow or inconsistent, and ask the plain question: is this an unwritten rule we have just never bothered to encode, or is it genuinely ambiguous every time? If it is the first, codify it. If it is the second, that is where a model belongs — with a person checking its work until you trust it, and honestly, often after that too.
The return on AI is real, but it is narrower and quieter than the marketing suggests. It shows up as a step that used to take an afternoon now taking ten minutes, with a person still in charge of the result. No magic, no replacing the team, no slide about transformation. Just the same engineering discipline we have always needed: understand the process, write down what can be written down, and reach for the model only where the problem is genuinely fuzzy.