Automatic
Clear and complete
Fields complete and confidence above the threshold. ARGUS looks up the rates, drafts the reply and leaves it ready to send. The person reads it and approves.
Inbound enquiry operations
The inbox decides the margin.
ARGUS reads every enquiry that arrives, extracts the data, scores the priority and drafts the reply. Your team approves. What was triage becomes judgement.
In arrival order
- Freight & Logistics Weekly
- Aurimetal, S.A.
- Internal
- Coelho Glassworks
- Nordeste Distribution
- Lima Stationery
By priorityScore
- 97Coelho GlassworksComplaint — order delivered damagedComplaint
- 94Aurimetal, S.A.Quote request — 12 pallets, Braga → BarcelonaQuote
- 71Nordeste DistributionCollection moved to tomorrowBooking
- 34Lima StationeryQuestion about transit times to IberiaInformation
- 2Freight & Logistics WeeklyInvitation: operations conference 2026Noise
- 1InternalRe: Re: team lunchNoise
< 30 s
Enquiry received → reply ready
classified, prioritised and drafted
< 15 min
Human review on ambiguous cases
with an alert to the team channel
6
Workflow states
from new to closed, with an audit trail
100%
Of replies approved by a person
nothing sends itself
The problem
The bottleneck is not the operation. It is the inbox.
On a pricing desk, the quote request, the complaint and the last-minute change all land in the same place. With no priority, they get worked in arrival order — not by value, and not by risk.
01
A slow quote loses the deal.
The customer asked three suppliers. Whoever answers in minutes wins, not whoever answers in hours.
02
The urgent one is buried in the volume.
A critical complaint sits behind forty routine emails and surfaces once it is already late.
03
Work is picked in arrival order.
The queue is chronological, not economic. The highest-value enquiry waits its turn.
04
The judgement leaves with the person.
Whoever knows how to price that lane is on holiday, and their reasoning is written down nowhere.
Every hour an enquiry waits is margin sliding towards whoever replied first.
How it works
Six steps, from email to an approved draft.
An agent that works the mailbox like an operator — and never sleeps. Every step leaves a record, and the last one is always a person.
- 01
Ingestion
Reads the mailboxes you point it at, continuously — body, headers and attachments, all stored in full.
- 02
Conversation reconstruction
Groups messages by Message-ID, In-Reply-To and References, with a heuristic fallback for when the customer breaks the thread.
- 03
Classification and extraction
Identifies the request type and pulls out the fields that matter: origin, destination, quantities, dates, references.
- 04
Priority
Scores each conversation from time without a reply, customer value, request type, urgency and SLA.
- 05
Routing
Chooses between answering, asking for what is missing, or calling in a person — according to how confident the classification is.
- 06
Assisted reply
Drafts in the language and register of the house, with the data already filled in. The person approves, edits or rejects.
What the machine sees
Prose goes in. Filled fields and a draft come out.
Classification and extraction run on local models. Email content never leaves your infrastructure.
QuoteAurimetal, S.A.conversation #2231
Confidence94%
Extracted fields
- Origin
- Braga
- Destination
- Barcelona
- Load
- 12 pallets
- Weight
- 3,240 kg
- Collection
- 18 Sep
- Reference
- AUR-4471
Draft · awaiting approval
- Rate
- €1,240
- Transit
- 48–72 h
- Valid for
- 48 hours
Dear Sirs, thank you for your enquiry. For 12 pallets (3,240 kg) from Braga to Barcelona, collected on 18 September, our rate is €1,240, delivered within 48 to 72 hours. This quote is valid for 48 hours.
How it decides
Three paths. One rule: nothing sends itself.
The machine handles what is clear; the person decides the rest. Every conversation is routed on the confidence of its classification, and no reply is sent without approval.
Ask for what is missing
One field short
Fields incomplete. ARGUS writes to the customer asking only for what is missing. When the answer arrives, the conversation moves itself to Path A.
Human review
Doubt or ambiguity
Low confidence or an ambiguous reading. The draft pauses, the team gets an alert, and the decision the person makes is recorded.
Path C is what makes the system improve.
Every human review is a high-quality example that becomes available to the enquiries that follow. The confidence threshold rises, and week by week fewer cases reach a person.
The engine
The intelligence runs in-house.
An internal LLM and a local RAG index built on your own history. Nothing leaves your network — unless you decide otherwise.
Internal LLM
The model runs on your hardware. Classification, extraction and drafting all happen without a single message leaving your network — not as a contractual promise, but as a matter of where the process runs.
Local RAG, tailored to your case
The index is built on your own history: your lanes, your customers, the replies your team has already sent. That is what separates a generic answer from the answer your firm would give.
It improves as it is used
Every new conversation enters the index and every correction stays as an example. Model weights are never retrained — the system gets better because it knows more about you, not because it was rewritten.
External providers, if you want them
For more capability on a specific step, connect Gemini, Anthropic or OpenAI. It ships switched off, you decide step by step, and it is reversible.
Where the model runs
Default
Internal LLMon your infrastructure
Optional — your choice
What changes
The inbox stops running the day.
The team stops triaging and starts doing what only a person does: deciding the hard cases and closing the deal.
| Before | With ARGUS | |
|---|---|---|
| Time to reply | Hours, whenever somebody picks it up | Minutes — drafted in seconds, approved and sent |
| Mailbox triage | Manual, in arrival order | Automatic, by priority and risk |
| Incomplete enquiry | Manual back-and-forth, if anyone remembers | Automatic follow-up, asking only for the gap |
| Critical complaint | Lost in the volume | Rises to the top of the queue |
| The pricer's judgement | In one person's head | Recorded and reused at every review |
| Sending the reply | Manual | Always with human approval |
Workflow and priority
Six states and a score the team tunes.
States
- New
- In progress
- Waiting on customer
- Ready to price
- Waiting on internal action
- Closed
Transition rules, an assigned owner, and a record of every action. Each conversation is a work object with a traceable history.
Score inputs
- Time without a reply
- Customer value
- Request type
- Urgency
- SLA
Weights are configurable without a redeploy. The people who know the operation decide what counts as urgent — not the software vendor.
Why it holds up in production
The difference between a chatbot and a decision layer is in the detail.
Which systems it sits on, who has the last word, and what gets written down.
It sits on what you already run
It reads the mailboxes that already exist and queries the systems you already have. We do not replace your email or your ERP — we build on top of what already works.
The person stays in command
No reply is sent without approval. The person approves, edits or rejects, and ARGUS does the clerical work. Accountability stays human.
Every decision leaves a trail
Classification, score, state change, draft and approval — all recorded, with an author and a timestamp. When somebody asks why this enquiry was handled that way, there is an answer.
What it sits on
Connections, not migrations.
Every integration is an isolated adapter with retry, a circuit breaker and a dead-letter queue. An integration that goes down never stops the system.
Email
- IMAP
- POP
- Microsoft Graph
Data and context
- PostgreSQL
- pgvector
- Attachments on disk or object storage
AI engine
- Internal LLM (default)
- Local embeddings
- Gemini · Anthropic · OpenAI (optional)
Business systems
- Pricing
- Booking
- TMS
- CRM
Sovereignty and GDPR
It runs in-house. It leaves a trail of everything.
Local processing
Processing happens on your infrastructure. None of the email content leaves for a third-party service.
Authentication and permissions
SSO, role-based access, and an audit record of every action — who did what, when, on which conversation.
Data minimisation
We can work from dedicated ingestion mailboxes, with the server-side copy deleted once read. The database becomes the source of truth.
Where it is today
In pilot at an Iberian haulier.
Working the operation's real mailboxes — quotes, complaints and collection requests — with the team approving every reply. The pilot's goal is a plain one: less time to reply, and fewer cases needing intervention, week by week.
4
Mailboxes monitored
one queue of work
< 30 s
Enquiry → reply ready
classified and drafted
< 15 min
Review on ambiguous cases
with an alert to the team
Always
Human approval before sending
no exceptions
Illustrative operational figures from a pilot in progress. The client is unnamed by our choice.
Questions
What the team asks first.
Does this replace our email?
No. ARGUS is not an email client and nobody stops using Outlook. It reads the mailboxes as an input of work and returns everything to the same place — the team keeps replying from where it always replied.
Does it send replies on its own?
No. Every draft waits for approval. The person approves, edits or rejects. That is an architectural decision, not a setting you can switch off without talking to us first.
Does it train on our email?
Model weights are never retrained. History is used to calibrate classification and as retrievable context, and your corrections become available as examples. None of that alters the model itself.
Does it run on our infrastructure or yours?
Yours, if that is what you need. The AI engine is local and depends on no external API, which makes fully in-house processing a real option rather than a compromise.
Will it work with our mail system?
If it speaks IMAP, POP or Microsoft Graph, yes. Ingestion is an isolated adapter, and adding a protocol is known work — not a rewrite.
And when it classifies something wrongly?
It goes to human review before it ever reaches the customer, which is precisely what the confidence threshold exists to guarantee. The correction is recorded and informs the enquiries that follow.
How long until it is working?
The first phase gets ingestion running on the real mailboxes, with classification calibrated against a sample of your own history. From there the pilot team works inside the system and tunes it with real use.
Does it work beyond logistics?
It works for any operation where the enquiry arrives as prose and the answer needs data. What changes is the taxonomy and the fields to extract; the classification, the priority and the human approval are the same.
ARGUS
Run the inbox on facts — not in arrival order.
ARGUS turns your inboxes into a single prioritised queue: every enquiry classified, routed and — where it is clear — already answered, waiting on one click. Operator-led. Implementation-focused. No transformation theatre.
Book a demoThe calendar is just below.
Talk to us
Thirty minutes, with the people who build it.
Pick a time. No sales deck — we show the system working and listen to how your inbox works today.
Open the calendar in a new window
The calendar is Calendly's and loads only once you reach this section. Using it means Calendly receives what you enter and sets its own cookies.