Every AI demo works. Every single one. That's the trap.
Someone runs a prompt in a boardroom, the output's incredible, everyone gets excited. Six months later the project's dead. Not because the AI broke — because nobody thought about what sits between that demo and something that actually runs. In production. At scale. With real users who don't read instructions and real data that looks nothing like the test set.
I keep seeing this. Government departments. ASX companies. Same story, different logo on the slide deck.
Demos are controlled environments
Curated inputs. A human cherry-picking the best output. That's not a product. A product handles the worst input on the worst day from the least technical person in the building and still delivers something useful.
What actually kills these projects
Nobody owns it. AI sits between IT, the business unit, legal and procurement. Everyone's "interested." Nobody's accountable. It gets discussed in meetings for months and built by nobody.
The data's a mess. Model works great on clean test data. Production data is spread across five systems with three naming conventions and a CSV someone emailed in 2019 that's still the source of truth.
Compliance gets bolted on at the end. PII, SOC2, the AI Ethics Framework — these are architecture decisions, not a checklist you tick before launch. Bolt them on later and you're rebuilding.
What I tell clients
Start with the problem. If you can't describe it without the word "AI" you don't have a problem. You have a technology looking for a job.
Build the boring parts first. Auth. Logging. Data pipelines. Access controls. Then add the model. The model's the smallest piece. Everything around it is what makes it work.
Ship to five users before five thousand. Watch them. Don't ask them what they think — watch what they actually do. That's where the product is.
The organisations getting real value from AI right now aren't the ones with the best models. They're the ones that treated it like a product problem instead of a science experiment.