The nine percent it refuses to answer.
Connecting a model to an inbox takes an afternoon. The decision that matters is one number per branch: how sure it has to be before it answers on its own. At 0.97, almost nothing sends. One wholesale supplier settled near 0.92 after a month of watching the log, which left about nine percent going to a person, flagged, with the records the model was reading when it stopped.
The model scored f1 0.94 on a week it had never seen before auto-send was turned on. The 166 it won’t decide go to a person, flagged, with the records it was reading when it stopped.
Six things, and what each one gets you.
Machine learning development
A small classifier trained on your own resolved history. Eighteen months of tickets teaches it your SKUs, your abbreviations and the way your customers write, none of which a general model knows.
GPT and LLM integration
Claude for the reasoning and a small local classifier for the routing, and you can see which one is doing what. Sending every message to the biggest model costs more and answers slower.
AI-driven automation
The model reads each message and suggests what to do. Fixed rules decide what actually happens, so a bad guess lands in a review queue instead of becoming an order.
Custom AI software development
The parts with no off-the-shelf version: the piece that reconciles two systems that disagree, the queue a person works through, and the screen that shows why a decision was made.
Data analytics and visualization
What the model is doing, in numbers: volume by branch, how often it agreed with the person who checked, and the cost per message. After a bad week, you can see exactly what happened.
Evaluation, and the drift watch after
Every prompt change is retested on a week of messages the model has never seen, the same way a code change would be. After launch, a weekly report flags it if agreement drops below 0.90, before a customer notices.
$12,000 to $22,000 for a first router, and you own it.
| Build | $12,000 to $22,000 for a first router. Yours varies with the project, and the price is fixed once quoted |
|---|---|
| Running | $0.004 a message at current volumes, on your own cloud account |
| Us, after | $0. No retainer and no licence. Your data stays in your tenant and isn’t used for training |
| Not included | Labelling, if your history isn’t labelled yet. You’ll know how much there is before anyone quotes you |
We read your history
Eighteen months of resolved tickets, before a price exists. If there isn’t enough labelled history for a model to learn from, that’s the finding, and we stop there.
It answers nothing, and we keep score
The model runs on live messages and sends nothing. Wherever its answer differs from your team’s, both are kept and compared. Auto-send stays off until it scores 0.94 (f1) against your team on a week of messages it has never seen.
Live at 0.97, and the slider is yours
It starts so cautious it barely sends anything. You watch the log for a month and lower the threshold yourself, one branch at a time, without having to ask us.
What people ask before they start.
What happens the first time it gets one wrong?
Is our data training someone else’s model?
We already pay for a chatbot. Why is this different?
What if our volume is too small for this to pay?
Start with where the repeated work is.
The free audit asks where your team’s hours go and prices the bottleneck we’d fix first, whether that turns out to be AI or something simpler. If you already know what you want a model to do, tell us directly.