One thing I’ve learned recently: Adding an LLM API is the easy part But building an AI feature that people can actually trust inside a real product is much harder.
In one of the SaaS platforms I’m working on, we built a Smart Upload workflow for energy and compliance data.
The basic idea sounds simple:
Upload CSV/Excel → AI reads it → structured data is created.
But the real workflow became much more interesting:
Upload → Analyse → Review → Correct → Re-analyse → Validate → Import
A few lessons stood out.
AI output should not immediately become application data.
We treat the AI result as a proposal. The user can review detected assets, energy types, units, dates and consumption values before anything is finally imported.
Human corrections are valuable context.
If the AI misunderstands something, the user can give an instruction such as: “The unit is kWh.”
or
“The reporting period is January to March.”
The system can then analyse the file again while preserving corrections the user has already made.
That sounds like a small UX detail, but it changes the experience from:
“AI got it wrong, start again”
to:
“AI and the user are progressively improving the result together.”
Context matters more than a clever prompt.
We’re also working with chatbot-style functionality inside the platform.
The useful version of a chatbot isn’t one that knows everything.
It’s one that understands:
AI needs normal software engineering around it.
Validation, permissions, audit history, structured schemas, error handling and deterministic business rules are still essential.
The LLM is one component of the system—not the system itself.
My biggest takeaway so far:
**Good AI products are less about generating answers and more about designing a reliable collaboration between AI, application data and the user. **