A small portion of the email we get is high-stakes — a client question, a vendor escalation, a partner ask. The vast majority is logistics, FYIs, status updates, newsletter content, and meeting confirmations. Sorting through that takes time. Time that, before we built this, was coming out of either client work or sleep.
So we built an AI layer on top of our own inbox. We call it part of "Ash" — our internal executive assistant system. This post is about that one layer: email triage.
What it actually does
Every incoming email gets read by an AI agent. The agent:
- Categorizes the email by intent and urgency. (Client question → high. Newsletter → low. Vendor invoice → medium-routine.)
- Routes it into the right folder or queue — no more inbox-as-backlog.
- Drafts a response for anything that warrants one, using context from previous threads with that sender and from our CRM.
- Queues the draft in the drafts folder. A human reviews, edits if needed, sends.
We never click send on a draft we haven't read. That's the design — humans approve everything outbound, especially to clients. The AI doesn't get to speak for us.
What it costs
About $50/month in API costs. That's what we measured over the first three months of production.
We use a mix of models: cheap local models for the categorize-and-route step (which doesn't need a frontier model), and a more capable cloud model for drafting responses to higher-stakes emails. Smart routing per task keeps the bill flat.
Compared to hiring a part-time EA for the same workflow, the cost difference is roughly a 50× multiple in favor of the AI layer. We're not pretending an AI replaces what a human EA does at the relationship-management level — but for the email triage and drafting work specifically, it's not close.
What we measured
Per a before-and-after time-tracking comparison: roughly 8 to 10 hours per week saved on email management alone, for one person.
That's measured by tracking time spent in the email app and the email-related calendar entries (writing back about things we said we'd handle), not by guessing.
This is one inbox, one workflow. Your numbers may be different — but if your inbound volume is similar (50–150 messages a day, half requiring some kind of response), you'll save in the same range.
Where it doesn't help
Client relationship moments. When a client emails about something sensitive — a project that slipped, a budget question, a personal note — the AI doesn't get to draft that. A human writes it, and the system knows to leave it alone (we've tuned for those signals).
Long, complex threads. When a thread has 20 messages and multiple decisions, the AI summary is useful but the response draft is usually wrong. Those we write manually.
Anything regulated. We don't let the AI draft anything that touches client PII, billing, or anything that needs to be HIPAA-aware. Those go to local-only models, and only for categorization — not drafting.
What you'd need to deploy this
This isn't a product we sell as a checkout-cart SKU. It's a custom build because everyone's email patterns, escalation rules, and integration needs are different. A typical client deployment runs:
- Discovery: 1–2 weeks. We map your inbound, your sender categories, your escalation rules, and your integrations (CRM, PSA, calendar).
- Build: 2–4 weeks. We deploy the agent on your tenant, hook it into your tools, and tune the categorization model on your actual emails.
- Validation: 1–2 weeks. Side-by-side with your current workflow before we cut over.
Total: about 6–8 weeks. Cost varies based on the integration surface, but most mid-market deployments land in the $15K–$35K range for the build, plus ~$200–500/month in API + monitoring.
If you're losing 8–10 hours a week per person on email triage, that math works pretty quickly.