AI inbox & ops triage
Somebody spends the first
ninety minutes of every day sorting email.
It is usually your most capable admin. It shows up on no report, and it is why enquiries sit for four days. We build the system that does the sorting and writes the first draft of every reply. Your people still press send.
What it costs you now
The arithmetic nobody runs.
Triage does not appear on a budget line, so it never gets costed. It is worth doing once.
- People touching the shared inbox
- 3
- Minutes each, every day
- 90
- Loaded hourly cost
- $30
- Annually
- $35,000
Put your own numbers in and the shape rarely changes. The cost is real, it recurs every month, and it buys nothing except messages arriving in the right place eventually.
That is the visible half. The expensive half is the enquiry that sat for four days, the review that mentions it, and the customer who went somewhere else.
What we build
Read, sorted, drafted.
Not a chatbot bolted to your website. A system that works on the inbox you already have, in the tools your team already opens.
01
It reads everything, first
Every message that lands, before a person opens it. New enquiries, complaints, invoices, suppliers, applicants, and noise, separated into the categories your business already uses and named the way your team names them.
02
It pulls out what matters
Addresses, order numbers, unit references, dates, whatever the message type calls for. Structured, so it can be routed and searched rather than re-read.
03
It writes the first draft
In your team's voice, using your existing answers. A reply that took four minutes to compose takes four seconds to check.
04
A human presses send
Always. Nothing reaches a customer without one of your people approving it, and that is enforced in the code rather than promised in a contract.
05
Anything urgent jumps the queue
Escalations land in Teams or Slack with a summary attached, inside minutes. The things that cost you money when they wait stop waiting.
06
One digest, every morning
What arrived, what was handled, what needs a person. Thirty seconds to know the state of the inbox without opening it.
The part that matters
It tells you when it does not know.
Everyone has met an AI that is confidently wrong. The safeguard here is not that the system is clever. It is that the system knows the edge of what it can tell. When a message is ambiguous, it declines to draft anything, routes it to a person, and attaches its reasoning.
A system that guesses here is worse than no system. That refusal is the feature.
A real example from the demo
“There's water everywhere and I'm honestly not happy about it. This was supposed to have been dealt with.”
52% confident. Not enough to draft a reply.
Active leak, or a complaint about previous work? No address, no job reference. A human needs to call.
Being straight with you
What we deliberately do not automate.
Every honest version of this has a boundary. Ours:
- Anything priced. Quotes stay with a person who can be held to them.
- Your largest accounts, unless you ask. Too much history to hand to a first draft.
- Anything the system is not confident about. It stops rather than guesses.
- Sending. Every message is approved by a named human, without exception.
If a workflow does not have enough volume to be worth building, we will tell you that during the Blueprint rather than build it anyway.
How it runs
Three steps, no surprises.
01
Blueprint
$2,500
We map your inboxes, design the system, and model the return on your real numbers. Credited in full against the build.
02
Build
From $15,000
Ten business days for a typical inbox. Live in your own cloud accounts, documented, with your team trained on it. Full refund if it misses the acceptance criteria we agreed in writing.
03
Keep it current
$2,000 / month
Monitoring, prompt tuning as your business changes, and 4 hours of changes a month. Optional, cancel any time, and the system stays yours either way.
Does this actually pay
Most of this does not work. Here is what does.
You have read the stories about AI projects going nowhere. They are largely true, and the research says something useful about which ones do not.
95%
of AI pilots produced no measurable change to profit.
MIT looked at 300 deployments. The ones that worked were narrow, sat inside a workflow people already used, and came from an outside vendor rather than being assembled in-house. That is a description of what we sell, which is why we lead with the failure rate rather than a success story.
MIT, The GenAI Divide, 2025 →+14%
more issues resolved per hour, and +34% for the newest staff.
A study of 5,179 support agents with a real control group. Someone two months into the job worked at the level of someone six months in. The gain lands on your least experienced people, which is usually who is watching the shared inbox.
Brynjolfsson, Li and Raymond, NBER 31161 →On revenue, the industry data on response speed all points one way. Roughly three quarters of enquiries go to whoever answers first, and the median business takes closer to two days than two hours. Treat the exact figures as marketing and the direction as real.
Your numbers will differ from all of these. On the call we use yours: how many messages a day, how long before someone answers, what an average job is worth, and what share of enquiries turn into work. If that arithmetic makes the system look unnecessary, we will say so.
We run this ourselves
Not a thing we imagined.
This system reads our own inboxes every morning before we do. We built it because we needed it, then found that every operations person we described it to wanted one. That is the whole origin story. There was no market research.
Which also means when something about it is annoying, we find out before you do.
The questions we get
- Will it email our customers on its own?
- No. A named person approves every message, enforced in the send path rather than in a policy document. There is no configuration flag that turns the check off.
- Is our customer data training an AI?
- No. Message content goes to the model provider under enterprise terms, which contractually excludes training, and is not retained afterwards. The system runs inside your own Microsoft 365 or Google Workspace on a service account scoped to the mailboxes we agree on.
- What happens when it gets something wrong?
- A wrong draft is caught by the person approving it, which costs seconds. The failure worth designing against is silent confidence, and the answer to that is the confidence floor: below it, the system refuses to draft and asks for a human.
- How long until it is running?
- Ten business days from the point we have credentials and twenty sample messages. The clock pauses if we are waiting on you, which is the only way a fixed date can be honest.
- What if we want to stop?
- Cancel the monthly fee and keep the system. The code is in your repository, the data is in your accounts, and the documentation is written for somebody else to pick up.
Find out what your inbox
is costing you.
Tell us how the work moves today. We come back with a scope, a price, and a date. Or we tell you it is not worth building.