Before wiring a model into a business process, three things are worth settling.
1. Know what you are sending
A request assembled from record data sends that data to a third party. Be clear about what personal or commercially sensitive information can end up in a request, whether your provider's terms permit it, whether it may be used for training, and what you tell your customers.
If your privacy notice does not mention it, it needs to — before you go live, not afterwards.
2. Do not trust what comes back
Model output is not reliable by construction. It can be confidently wrong, and it can contain anything at all — including text designed to manipulate whatever processes it next.
- Never display model output as raw formatted content without escaping it.
- Never use it to build a query, a command or a link without checking.
- Where the output drives a decision that matters, keep a person in the loop. Use it to draft, suggest and prioritise — not to decide unsupervised.
3. Expect it to be slow, and sometimes unavailable
A request can take many seconds. Do not put one anywhere that has to respond quickly. For anything customer-facing, either show a clear progress indicator or do the work in the background and present the result when it is ready. Providers also have outages and rate limits, so whatever uses the gateway should cope with no answer at all.
Worked Examples
- Good: suggest a category for an enquiry, with the agent able to override it.
- Good: summarise a long case history for a human reading it.
- Risky: automatically closing cases the model judges resolved.
- Avoid: making eligibility or pricing decisions with no human review.