inbound
Every inbound email gets classified, prioritised, assigned and logged automatically, with a reply drafted and waiting. You approve it before it goes out — every time, not just when the AI is unsure.
Four things go wrong in every inbox we've been asked to look at. None of them are about writing replies.
Every email has to be read just to find out what it is
Sorting is not the work, but it is the part that cannot be skipped — so it happens first, by hand, on every message that arrives.
Urgent issues sit behind newsletters and invoices
An outage and a receipt arrive in the same list, in the order they were sent. Priority is whatever the sender happened to choose.
Nothing reaches the CRM until someone types it in
The record of an inquiry is created by hand, days later, from memory — which is why reporting never quite matches what happened.
The backlog is invisible until someone complains
Without a queue and a timestamp on arrival, a missed response window is only discovered when the customer raises it for you.
One table, no rounding in your favour. Substitute your own volume and rate — the shape of the calculation doesn't change.
| Line item | Basis | Amount |
|---|---|---|
| Manual triage today | 2.5 h/day × 21 days | 52.5 h / mo |
| Triage after classification + drafting | same volume, review only | 17.5 h / mo |
| Time saved | 52.5 − 17.5 | 35 h / mo |
| Value of that time | 35 h × $45/h loaded | $1,575 / mo |
| Build | one-off, fixed scope | $10,000 |
| Payback | $10,000 ÷ $1,575 | 6.3 months |
Pick an inquiry and press run. This is the same classification output the production pipeline writes to your CRM and posts to Slack.
Sample scenarios — not live data.
takes about eight seconds
The approval gate is not a limitation we're working around, and not a setting that gets loosened once you trust the numbers. It is the reason the rest of the pipeline is safe to run.
Confidence is not correctness
A 97% score is the model's estimate of its own reliability, not a guarantee about this particular email. The remaining 3% doesn't announce itself — it goes out under your name, to a customer, in your tone of voice.
The costs are not symmetric
A reply held for ten minutes costs you ten minutes. A confidently wrong reply sent in four seconds costs an apology, a correction, and sometimes the account. Speed is only worth having on the side of the trade where mistakes are cheap.
Approval is where the system stays observable
Every edit a reviewer makes is a visible signal about where the drafts are weak. An auto-send path deletes that signal precisely for the messages nobody ever reads again.
Accountability has to land on a person
When a customer asks who told them something, "our classifier was 96% sure" is not an answer. Someone on your team clicked approve, and can say why.
What stays automatic: reading, classifying, prioritising, assigning, logging and drafting — the work that is repetitive and reversible. What doesn't: the send. That is one click, by a person, on a draft that is already written.
Each inquiry is read once and tagged with category, priority, intent and owner — tuned to the way your business actually sorts work.
Every classification carries a confidence score, and an uncertain one says so plainly. It tells you how closely to read the draft — it never decides whether the draft is sent.
Nothing reaches a customer until a person approves it. Approval is the control that lets you hand over the reading, sorting and drafting without handing over your voice.
The human stays in the loop by design. Nothing leaves your business unread. Every decision the AI makes is logged with its confidence score, so you can see exactly what it got right — and the AI's remit stops at the draft.
Built around your inquiry types, your tools and your approval rules. Not a subscription to a generic inbox tool — a system that fits how your team already works.
starting at
$10,000
One-off build. Scope and timeline confirmed before anything starts.
Replace with your real contact address before publishing.
Trained on your real categories, not a generic taxonomy.
Results land where your team already works.
Every reply is queued for a person to approve. No auto-send path, at any confidence level.
Volume, response times and where the AI needed help.
Set up in your environment, with a handover your team can run.