AI inbox triage for email and chat

Nothing is missing. It is just spread across four places.

A customer emails the shared mailbox. Another writes to your personal address. A third fills in the contact form. A supplier sends a chat message. Every one of those is real work. None of them is in the same place.

So someone sorts it by hand. Usually the owner. Usually in the evening. They read everything, decide what matters, and forward each item to whoever should handle it. That person is the routing layer. When they take a week off, requests sit for a week.

The reading is not the expensive part. The deciding is. Genuinely urgent, or only written that way. Complaint, or question. Five-minute reply, or two-week project. Already answered, or still waiting.

One inbox for every mailbox and chat channel — read before you open it

Provostal connects to your mailboxes — Gmail, Microsoft 365, IMAP or Fastmail, as many as you have — and to Slack and Telegram. Every message is threaded into one inbox. Then an agent reads each one, searches your own knowledge base, labels it, sets its status and identifies the sender.

You no longer start the day by reading everything. You start with it already sorted.

Unified inbox: Every mailbox, plus Slack and Telegram, threaded into one inbox and labelled as messages arrive.
Work items: Actionable conversations become work items with the original message attached, deduplicated against work already open.
Contacts: One record per person and organisation, linked to the conversations they appear in.
Projects: Work grouped into projects, kept in two-way sync with Todoist and Linear if you already use them.
Knowledge base: Your documents, policies and notes — the agents search them before they decide, and cite what they used.
Ask directly: Ask an agent a question in plain language and get an answer with its sources listed.

How AI inbox triage works, in four steps

Four stages. Every one of them is recorded, and every one can be turned off.

Connect. Point Provostal at a mailbox — Gmail, Microsoft 365, IMAP or Fastmail — or at Slack or Telegram. Connect as many as you like; each one feeds the inbox only once you switch it on.
Triage. An agent reads the conversation, searches your knowledge base, applies labels, sets the status and matches the sender to a contact. Every tool call it makes is recorded.
Decompose. Conversations worth acting on become work items — checked against open work first, so the same request twice does not become two tasks. A second agent rewrites the body into a clear description.
Assign. Work items land on a board with a status, a type and a priority. You assign them, to a person or to an agent you have allowed to pick work up.

Triage travels back to your mailbox. Archive a conversation in Provostal and the mail moves to Archive where it came from; mark it spam and it moves to Junk. Nothing is ever deleted at the source.

Most AI automation fails in the same two ways

Pointing a general-purpose model at your mailbox and letting it reply is not triage. It is a faster way to be wrong.

It does not know anything about you.

A general model knows the internet. It does not know your refund window, which supplier is on notice, or what your team told this customer last month. So it produces answers that are fluent and wrong. Provostal’s agents search your own knowledge base before they act — your documents, your policies, your Notion or your markdown vault — and every answer lists the sources it used, so you can check it against the file it came from.
You cannot see why it did that.

When an automation gets something wrong, the useful question is not what happened but why it decided that. Every agent run is a record: which agent, which tools it called, what it passed in, what came back, what it changed. That is the AI agent audit trail, and you can open a run a month later and read it. Nothing is a black box.

Not a chatbot. Not a rules engine you have to hand-maintain. An operations layer that reads what arrives and turns the real work into work items.

Built for whoever currently does the sorting

The job is the same everywhere: read what came in, decide what it is, get it to the right person. What differs is who is stuck doing it.

Small businesses and owner-operators

There is no ops team, so sorting lands on whoever is most senior — late, in batches, held in one person’s head. Provostal reads and labels each message as it arrives, so what you open in the morning is already sorted and the actionable parts are written up as work.
Operations and CX managers

Triage quality varies by who is on shift. Provostal reads each message the same way every time, and the deterministic rules you write in the app — label this, archive that, on these channels — run alongside the agents, so the boring half never depends on anyone’s attention.
Agencies and service businesses

Client messages arrive over whatever channel the client feels like using, and the quick ones get lost. Provostal attaches each conversation to the right contact and turns the actionable ones into work items, so a passing message lands on the same board as everything else you owe that client.

Inspectable by design. Hostable on your own infrastructure.

The question is not whether the model is good. It is what happens when it is wrong, and whether you can see it.

Every meaningful action is published as a typed event to an append-only log, and that log — not a module’s own tables — is the source of truth. The audit trail is a by-product of how the system runs, not a log someone remembered to write. Agent runs record each tool call individually.

The log is queryable over the API and replayable from the command line. It is also pushable: register an endpoint with the event patterns you care about, and every matching event is POSTed to you with an HMAC signature you can verify, with retries and a dead-letter queue behind it. Your systems do not have to poll Provostal to know what it did.

The stack is deliberately boring: an API-first Laravel backend, a Vue 3 and TypeScript front end, a Python/FastAPI service for ML workloads, embeddings in PostgreSQL with pgvector, Redis and Horizon for queues, Reverb for realtime. It runs as a self-hosted AI platform via Docker, with database backup and restore built in.

Read this before you evaluate self-hosting. Model calls go to a hosted provider: the registry ships OpenAI and Anthropic models, and OpenAI is the default and also generates the embeddings. So self-hosting moves the application, the database and the event log onto your infrastructure — it does not move the model, and message content goes to whichever provider you configure. If content must never leave your network, that is a provider question, and we will answer it straight before you connect anything.

Questions worth asking before you connect anything

What is AI inbox triage?

AI inbox triage is the automatic first pass over everything a business receives — the mail in every mailbox, and the messages in every chat channel. Instead of a person reading each one to decide what it is and whether it needs doing, an AI agent reads it, labels it, sets its status and identifies the sender, against your own documented context.

Is this just another inbox I have to check?

It is an inbox, and it is one instead of five. The difference is what you open: conversations arrive labelled and triaged, and the ones worth acting on are already written up as work.

Does the AI act without me?

It acts inside a fixed set of moves: label a conversation, set its status, link it to an existing contact, open a work item. It cannot reply to anyone — Provostal sends no mail and no chat message to your customers, because outbound is not built. Every agent can be switched off individually.

What happens when it gets something wrong?

You change it, and your change is recorded on the same log as the agent’s. Open the run and you see which tools it called and what each returned. The best correction is usually to fix what it read: the agents search your knowledge base, so an out-of-date policy in there is the thing to edit.

How is this different from a shared inbox tool?

A shared inbox gives several people one mailbox and a rule builder you maintain by hand — every new case is another rule. Provostal keeps deterministic rules for the boring work, but the judgement calls are read by an agent grounded in your knowledge base, and the output is work items rather than a tidier inbox. It is not a helpdesk: no tickets, no customer portal, no outbound replies.

What does it connect to?

Mailboxes over Gmail, Microsoft 365, IMAP or JMAP/Fastmail — as many as you have. Chat over Slack and Telegram. Tasks and projects over Todoist and Linear, in two-way sync. Knowledge from Notion or a markdown vault. A website contact form reaches the inbox the same way it reaches you now: by sending mail to a mailbox Provostal is watching.

What does Provostal cost?

Pricing is not published yet, and that is not a sales tactic — it is not decided. Email us and we will tell you where it stands, including whether it fits the size of operation you are running.

Where does my data go?

Message content goes to the model provider running the triage. OpenAI is the default; agents can be pinned to Anthropic models instead. Everything else — inbox, work items, knowledge base, event log — lives in the PostgreSQL database you point Provostal at, which can be your own. Provostal holds no compliance certifications — evaluate it against your own policy, not against a badge.

Can we run it ourselves?

Yes, with Docker on your own infrastructure, with database backup and restore in the admin interface. Note the answer above: self-hosting does not keep message content in-house, because the model runs elsewhere.

How new is Provostal?

Early, and specific about it: version 0.4.1, first commit June 2026. The inbox, the triage and decomposition agents, the knowledge base, work items, contacts and projects are built. Routing and escalation policy are not. If you want a finished product, this is not it yet. If you want to shape one, tell us what you receive all day.

Does it work in Dutch?

Yes. Agents answer in the language the installation is set to, and follow the language of the conversation in front of them when it differs. The interface is English.

How long before it is useful?

Connect one mailbox and triage starts on the next message; there is no training period. Getting it right takes longer, because the first week is you correcting decisions and filling in the knowledge the agents read. Start with one channel, not everything.

Point it at one mailbox and watch what it does

Connect a single mailbox and let AI email triage run on real messages for a week. You will see what it labelled, what it turned into work, and every step it took to get there. It cannot message anyone, so the worst case is a week of reading its homework.

Get started Talk to us

Every email and message arrives. None of it sorts itself.