Production: what a prototype skips and a real system can’t
A prototype assumes every system answers and every input is clean. Production plans for the day they don’t. What we build into every system so it keeps working, and keeps someone responsible for it.
Plenty of AI projects end as impressive prototypes that never reach daily use. The prototype assumed the happy path: every source answers, every input is clean, every user does what the demo did. The fifth stage of every project is production: making the system survive reality, and staying responsible for it.
Plan for the bad day
On the data layer we built for an AI platform for lawyers and clinicians, every external system was expected to fail eventually: throttling, timeouts, inconsistent responses, changes without notice. So every call has retries and timeouts, circuit breakers keep one failing source from taking the others down, and structured logs let anyone follow a single record from its source to the assistant.
Ship on the client’s own ground
The AI sales-coaching product we built ran on the client’s own cloud and AI accounts from week five, so there was nothing to migrate later. Its CRM connection only reads, never writes. When the AI model is slow or unavailable, a rule-based scorer takes over and is labelled as such, so a manager never sees a blank and never mistakes a fallback for the model’s judgment.
Build for the real user
SuperEscuela’s students use old phones, often without a connection. So it works offline, on the oldest phone, without passwords to forget. Production means the conditions people actually have, not the ones on the developer’s desk.
What production includes
- Automated tests and a release gate, so problems are found before customers find them.
- Monitoring and alerts on what matters to the business, not just server health.
- Security and data separation designed in, and tested.
- Failure handling for every external dependency.
- Feedback from real usage, so the system improves on evidence.
- Someone who owns it: one team responsible from the first question to the running system.
Back to the start
Production feeds the first stage again: real usage creates new data, which is ingested, normalized, checked and turned into the next improvement. Ingestion → Normalization → Trust → Value → Production isn’t a one-way pipeline. It is how a trusted system keeps getting better.
This article is drawn from a real engagement. Client details withheld; every figure comes from the client’s own data.
Read the full case study →