AI Agent for Inbox Triage & Drafting
A mailbox is a queue sorted by arrival time, which is the least useful order available. This agent reads threads to their current state, separates what needs an action from what needs only awareness, pulls out the commitments buried in them, and drafts replies for you to send.
What is an AI email assistant?
An AI email assistant is a governed software worker that processes a mailbox. It classifies each thread by what it requires rather than by arrival time, reduces long conversations to their current state, extracts commitments made in either direction from inside message text, and drafts replies that are only sent after explicit human approval.
What it does
What it is not
Forty messages, four that matter, sorted by arrival
Mail is the only enterprise system where the interface deliberately orders work by when it showed up. The consequence is predictable: the request that needed ten minutes sits behind thirty notifications, and the commitment someone made in the sixth paragraph of a reply is discovered when they do not honour it.
Arrival order is not priority
The message that blocks three people is below a newsletter because the newsletter arrived at nine.
Threads have to be reread
A twenty-message thread has one current state and you have to reconstruct it from the bottom up to find out what it is.
Commitments hide in prose
A colleague agrees to send something by Thursday in the middle of a paragraph about something else, and nothing records it.
Rules cannot read meaning
Filters key on sender and subject, which is why every mailbox eventually has forty rules and still needs triage.
A mailbox ordered by what it actually asks of you
Triage
Sorted By What It Asks Of You
Not by when it arrived.
Each thread is classified by what it requires — a decision, an action, a reply, or nothing at all — and the mailbox is presented in that order, with the reasoning shown so a misclassification is visible rather than silently burying something.
- Classified by what the thread requires
- Reasoning shown on each classification
- Awareness-only items grouped and collapsed
- Deadlines inside the text detected
Not by arrival
Threads
Where The Conversation Now Stands
Not a digest of every message.
A long thread is reduced to its current position — what was asked, what has been settled, what is still open and who is waiting on whom — rather than a chronological summary that leaves the reader to work out the state themselves.
Settled and open
Commitments
The Promise In Paragraph Six
Extracted and dated.
Undertakings made in either direction are pulled out of the prose with the person and the date, so something agreed halfway through a reply about another subject becomes a tracked item instead of a thing one party remembers.
From inside the text
How the AI Email Assistant runs a task
- STEP 01
Read the thread as one object
A conversation is treated as a single item with a state rather than as a sequence of messages, because what a reader needs is the current position and the list of messages is the raw material for working it out.
Thread assemblyState extraction - STEP 02
Classify by what it requires
Each thread is sorted into decision, action, reply or awareness based on what it actually asks of the recipient, and the reasoning is attached so that a thread pushed down the queue can be checked rather than trusted.
Requirement classificationReasoning capture - STEP 03
Pull the commitments out
Undertakings in both directions are extracted with the person and any date, including the ones stated in passing inside a paragraph about something else, which is where the majority of missed commitments actually live.
Commitment extractionDate resolution - STEP 04
Ground the draft
A reply is drafted from the thread, its attachments and any relevant internal knowledge, in the register the sender actually writes in, with anything it could not establish marked rather than filled with a plausible sentence.
Reply draftingKnowledge grounding - STEP 05
Stop before sending
Drafts, task proposals and record updates are all held for explicit approval, because a mail agent that can send is a mail agent that will eventually send something that should have been read first.
Approval gateSend log
Systems the AI Email Assistant connects to
Mail access
Content handling
Inputs, outputs and runtime
- Ingests
- Mailbox threadsAttachmentsContact directoryInternal knowledgePrior commitments
- Produces
- Action-ordered mailboxThread state summariesTwo-way commitment listDrafted repliesProposed tasks and records
- Triggered by
- New mail arrivingTriage run requestedCommitment date approaching
- Human oversight
- Every outgoing message is approved first
- Models
- Open-weight LLMs you host — Llama, Qwen or Mistral class
- Typical latency
- Seconds per thread
- Deployment
- On-premise or sovereign cloud with egress control
- Data residency
- Mail content never leaves your network
Where the Email Assistant pays back
Morning Triage
Present the mailbox ordered by what each thread requires, with awareness-only items grouped out of the way.
Thread State Summaries
Reduce a long conversation to what is settled, what is open and who is currently waiting.
Reply Drafting
Prepare a response in the sender’s own register, grounded in the thread and any attached documents.
Commitment Capture
Extract undertakings from inside message text with the person and date, in both directions.
Backlog Recovery
Work through an inbox after leave and identify what still needs action versus what resolved itself.
Task And Record Preparation
Draft tasks or CRM notes from a thread for confirmation, rather than writing them into systems directly.
AI Email Assistant vs chatbots and SaaS copilots
Every mail client eventually offers rules, and every heavy mailbox eventually has forty of them and still needs triage, because a rule matches a sender and the thing you need to know is what the message is asking.
| Generic chatbot | SaaS copilot | VDF AI | |
|---|---|---|---|
| Ordering basis | Not applicable | Arrival and rules | What the thread requires |
| Unit of work | One message | One message | The thread and its state |
| Commitments | Not extracted | Explicit tasks only | Found inside message prose |
| Classification reasoning | None | Rule id | Shown and checkable |
| Sending | Not possible | Can auto-send | Approval required every time |
| Draft grounding | Generic | Thread only | Thread plus internal knowledge |
| Where mail is read | Pasted to vendor | Vendor cloud | Inside your own network |
Governance and controls
Mail is where most organisations discover they hold data they never inventoried, and an agent reading a mailbox inherits every one of those obligations at once rather than gradually.
Approval before any send
Nothing leaves without a person
Single-mailbox scope
No access to other people’s mail
Classification reasoning kept
Why a thread was ranked is logged
No rule or filter changes
Mailbox configuration is untouched
Attachments stay internal
Files are parsed in your network
Retention follows your policy
Derived summaries expire with mail
Evidence it leaves behind
What changes after rollout
Who runs the AI Email Assistant
Account director
Opens a mailbox where the three client threads needing a decision today are at the top with their current state written out, and the fifty notifications are collapsed into a single line.
Operations lead returning from leave
Works through two weeks of backlog by state rather than by message, seeing immediately which threads resolved themselves and which are still waiting on a reply that never came.
Compliance officer
Can show that no message was ever sent without human approval and that every classification the agent made is logged with its reasoning, which is what makes mail automation defensible at all.
Questions about the AI Email Assistant
What is an AI email assistant?
It is an agent that works a mailbox: classifying each thread by what it requires rather than when it arrived, reducing long conversations to their current state, extracting commitments from inside the text, and drafting replies for approval.
How is an AI email assistant different from a generic chatbot?
A chatbot summarises a message you paste. This agent reads whole threads in place, orders the mailbox by required action, and tracks the undertakings made in both directions.
Can an AI email assistant run on-premise on mailbox data?
Yes. A mailbox is the least filtered record an organisation holds, containing commercial terms, personnel matters and legal discussion, so nothing is processed outside your own infrastructure.
What does an AI email assistant produce, and in what format?
A mailbox ordered by required action with the reasoning shown, per-thread state summaries, a two-way commitment list with dates, and drafted replies held for review.
Where does an AI email assistant fit in a governed AI programme?
It prepares; you send. No message leaves without explicit approval, no record is updated without confirmation, and scheduling requests are handed to the scheduling assistant.
Will it ever send an email on its own?
No. Drafts are prepared and wait for explicit approval, and that boundary is not configurable down. Mail is the channel where an error is immediately external, irreversible and attributed to a named person, and the time saved by removing one confirmation click is nowhere near the cost of one wrongly sent message. The drafting is where the minutes actually go.
How is classification different from mail rules?
Rules match on sender, subject and keywords, which is why they cannot distinguish a colleague asking a question from the same colleague copying you for information. This agent reads what the message asks of you. It also shows why it classified a thread the way it did, so the failure mode is a visible misclassification rather than a message silently filed in a folder you check monthly.
Can it read my colleagues’ mailboxes?
No. It is scoped to the mailbox of the person it works for, under their own credentials. Shared and functional mailboxes can be connected explicitly where the organisation intends that, but there is no path by which an individual assistant reaches another individual’s mail. This is a hard scope rather than a setting, for reasons that are obvious the first time anyone imagines the alternative.
What does it do with attachments?
It reads them where they bear on the thread — a contract sent for review, a spreadsheet the question is about — and uses them to ground the draft reply and the thread state. Parsing happens inside your network, which is the relevant point: attachments are typically the most sensitive part of any mailbox and are routinely the thing an external mail tool is least suitable for.
How does this relate to the customer support and SDR agents?
Those work a queue of external correspondence with a defined purpose: resolving a support case, running an outbound sequence. This agent works an individual’s general mailbox, where the traffic is mixed and the useful output is triage and drafting rather than case resolution. Where a support mailbox exists, the support agent is the better fit and this one stays with the person.
Order the inbox by what it asks of you
See the AI Email Assistant triage a mailbox and draft replies for approval.