PHP Development
Intelligence Inside a System You Already Run
Last updated:
Three things that fit most applications
- Classification — routing, categorising, prioritising incoming work
- Extraction — turning documents and free text into structured fields
- Drafting — replies, summaries and descriptions for a person to review
All three produce output a person can check in seconds, which is what makes them safe additions to a system people already depend on.
Call it from a queue, not from a request
Model calls are slow and occasionally fail. Making one during a web request means a user waiting and a request that sometimes times out.
- Queue the work, process it, store the result
- The interface shows a pending state, then the result
- Retries handled by the queue, not by the user
- A provider outage delays rather than breaks
Store the output as data
The model's output should land in your database as structured fields with a confidence indicator, not be regenerated each time it is displayed.
That makes it queryable, auditable, correctable and cheap. Regenerating on every view is slow, expensive and produces different answers over time.
Always keep the human path
- A review queue for anything uncertain
- In-place correction, captured as data
- A switch that disables the AI step and falls back to manual
- The original input preserved alongside the output
Costs and controls
| Control | Why |
|---|---|
| Hard daily cap | A loop must not spend the quarter's budget |
| Per-user cap | Abuse and accidents |
| Cost per item tracked | Rising cost is the early warning |
| Prompts in configuration | Changeable without a deployment |
| Model version recorded | So output can be explained later |
Frequently asked questions
Does this need a rewrite?
No. It is usually a queued job, a few database columns and a review screen alongside what you already have.
What about the data leaving our systems?
That is a decision to make deliberately: which provider, which region, what retention, and whether to redact before sending.
How long does it take to add?
Four to eight weeks for a first capability including the evaluation set and the review interface.
What if the model gets it wrong?
It will sometimes. That is what confidence routing and the review queue are for, and why the human path must stay.
Have a system that could classify or extract automatically?
Tell us what people currently do by hand and we will tell you whether it fits.
Related services
What we build for problems like this one