Every founder has an AI line on the roadmap now. Most of what sits under it is theatre: a chatbot bolted to the corner of the screen, a strategy deck, a feature shipped so the announcement can say "now with AI." Almost none of it changes a number anyone cares about.
The useful part of AI is real, but it is smaller and more specific than the noise suggests. Here is how to find it.
Start with a job, not a model.
The failure pattern is always the same. It begins with "we should add AI" and then goes looking for somewhere to put it. That is backwards, and it reliably produces features nobody uses.
The fix is to start from a job that is currently slow, manual or expensive, and ask whether a model does it better. Not "where can we add AI" but "what task is eating my team's week, and can this do it." If you cannot name the task, you are not ready to build anything yet.
Three shapes of AI that earn their place.
1. A feature inside your product
The model does one job your users already need: summarise a long document, draft a first reply, classify an incoming item, pull an answer out of a pile of text, or make search understand meaning rather than keywords. Narrow, measurable, and it makes the product visibly better. This is where most of the real value lives.
2. An agentic workflow in the back office
A task that used to need a person, now done end to end by a system you can trust: triage every inbound enquiry and route it, read documents and pull the structured fields out, watch a queue and act on what lands in it. The win is plain. Hours removed from someone's week, every week.
3. AI-accelerated delivery
This one is invisible to your users and matters anyway. The team building your software uses AI to ship faster, which means you get more for the same budget. You should not see it in the product. You should see it in the pace.
The boring parts that decide whether it works.
A demo takes an afternoon. Production is the actual work, and it lives in the parts nobody films:
- Evaluation. How you know the output is good often enough. Measured, not felt.
- Guardrails. What happens when the model is wrong, because sometimes it will be.
- Data. The retrieval and context that make answers specific to you, not generic.
- Fallbacks. The path when the model is unsure or unavailable.
- Cost. What each call actually costs at the volume you really run.
A team that only talks about the demo and never about this list is selling you theatre. The boring list is exactly where the value is won or lost.
What to ignore.
Most of the AI folder can go straight in the bin:
- A chatbot added because every site has one now, answering questions nobody asked.
- An "AI strategy" that is a deck rather than a build.
- Anything whose pitch cannot name the hours or the money it removes.
The test is simple. If you cannot say, in one sentence, what measurable thing this AI makes faster, cheaper or better, it is theatre. Real AI in a product always has a number attached to it.
We build the useful part and leave the theatre alone: AI features inside the software we build, agentic workflows that remove real hours, and a delivery practice that ships faster because of it. If you have an AI line on your roadmap and want to find the part that is real, that is a good first conversation. Drop me a line, or see how we build it.