AI Agent for Email Triage: Your Shared Inbox as a Process 2026 (automation from €3,500 net)
An AI agent for email treats the shared inbox (office@, sales@) as a process: it classifies, extracts data, drafts replies for approval, escalates exceptions, and leaves a trail. Automation from €3,500 net, an agent that runs the process from €6,000. You calculate the hours triage costs on your own numbers. You start with a free process scan.
An AI agent for email treats the company's shared inbox (office@, sales@, contact@) as a process, not a folder: it reads every message, classifies the case type and urgency, extracts key data, routes the case onward, and drafts a reply for approval. Automating such an inbox starts from €3,500 net, and an agent that runs a case across several systems from €6,000. How many hours triage costs, you calculate on your own numbers. The first step, a free process scan, costs €0.
Quick answer
The price depends on how much work the system actually does in the inbox, not on the label "agent":
- free process scan (€0): a 30-minute engineer call plus a written takeaway in two business days,
- inbox automation with an LLM (from €3,500 net): classification, data extraction, routing, and draft replies, sent after human approval,
- agent that runs a case (from €6,000 net): carries the case across several systems (inbox, CRM, ERP) within the boundaries you set, with escalation and a trail,
- maintenance (priced individually): hosting, monitoring, SLA, and changes after launch, plus a variable model cost measured as a few cents per email times volume.
If you want a portable document with architecture and a fixed quote before a bigger decision, the current price of the implementation specification is €1,200 net. Prices are Polish-market net figures (net of VAT). This piece is about text handling: email, the contact form, the shared inbox. For customer-service tickets more broadly, we cover the AI agent for customer service separately.
How an agent turns the office@ inbox into a process
The shared inbox is cheap to set up and expensive to run, because nobody owns it. Several people read it "in passing," some cases get lost, and the sorting rules live in people's heads. An agent does not "handle the mail" by magic: it breaks the inbox into five stages, each with a boundary and a person responsible for the decision.
| Stage | What the agent does | Boundary: who decides |
|---|---|---|
| Classification | Recognizes the case type (quote request, invoice, complaint, order status) and assigns urgency. | A confidence-threshold rule: an unclear case goes to a human, not forced into a bucket. |
| Extraction | Pulls key data: order number, VAT ID, amount, deadline, address, and attaches it to the case. | Uncertain data is flagged for review, not written silently into a system. |
| Routing | Sends the case to the right person or team with the thread pulled out, instead of forwarding a raw email. | An unknown case type lands in a "sort manually" queue. |
| Draft reply | For routine cases, prepares a first version of the reply from approved content. | A human approves (or a rule sends only for narrow, stable types). |
| Trail | Records, per case: what it recognized, the decision it made, the policy version, and the query cost. | Without a reproducible trail it is not an agent, just roulette with a nice interface. |
The order of the rollout matters more than the model here. We start production in observation mode: the agent classifies, extracts data, and prepares drafts, but nothing goes out to the sender without approval. We enable auto-send only where classification is stable and the content is approved. That order, not the quality of the model, is what turns a demo into a system.
How many hours triage costs: run it on your own numbers
The cost of the inbox is not our promise, it is your substitution. Start with what handling the mail manually costs today:
Annual manual inbox cost =
hours per week reading, sorting, and first-replying
x hourly rate of the people doing it
x 52Then calculate the payback. Subtract maintenance and model cost from the saving, because those are real lines after launch:
Monthly saving =
(hours recovered per month x hourly rate)
- monthly maintenance
- AI model cost
Months to payback =
build cost ÷ monthly savingThe scale is often larger than it feels. The McKinsey Global Institute estimates knowledge workers spend about 28% of the workweek on email, close to 11 hours. The Microsoft Work Trend Index (2025, telemetry from 31,000 workers) reports the average worker receives about 117 emails a day. On top of that sits the hidden cost of switching: research by Gloria Mark at UC Irvine finds it takes about 23 minutes to fully refocus after an interruption, and an overloaded inbox breaks the day into dozens of them. These are figures from other markets, so treat them as context for scale, not a result for your company. Your result comes from the formula above, with your own hours.
For comparison, again as a reference and not a promise: Microsoft's peer-reviewed Copilot study (6,000 workers across 50+ companies) found about 3 hours a week less on email and a 25% cut in email time. How much you recover depends on how repeatable your cases are, and that is what we check on the scan before quoting anything.
What the agent must never send without a human
An email agent is safe not because the model is clever, but because it works inside a narrow permission model. The next step, whether a case may be closed automatically, is decided by a deterministic process rule, not the model. Without human approval, these should never leave the inbox:
- quotes, prices, and estimates, because they are a commercial commitment,
- confirmations of dates, quantities, or terms, because a mistake has a real cost,
- replies to complaints and disputes, because they touch money and emotion,
- anything at low classification confidence, because an unclear email is better handed to a human than guessed.
In practice we enable auto-send only for narrow, repeatable case types with approved content, for example acknowledging receipt of a ticket or a standard status note. Everything else ends in a one-click draft, not an automatic send. Separately, we guard against prompt injection: an email from an outside sender may contain content that tries to hijack the agent's behavior, so such attempts go to quarantine rather than execution. More on that threat is in the piece on prompt injection in AI agents.
GDPR: an inbox holds personal data
A company inbox is a store of personal data: names, addresses, numbers, the content of correspondence. Letting a model into that inbox is data processing and needs full rigor. In practice that means four things we settle before launch, not after:
- a data processing agreement (DPA) with every provider that handles email content,
- scope and purpose: what data the system even sees and why, with minimization where the full content is not needed for classification,
- retention: how long the system keeps content, extracted data, logs, and the decision trail, and when it deletes them,
- a trail, not a copy: for accountability, a record of the decision and its reason is usually enough, not a second database with full email content.
The choice of model and processing location is part of this decision, not a technical detail. We unpack it further in the piece on GDPR and DPA when deploying AI. The rule is simple: if you cannot say where an email's content goes and how long it stays there, you are not ready for production yet.
When you do NOT need an email agent
Honestly: some companies should not buy this, however fashionable the topic is. We will not sell an agent to an inbox that gets a dozen emails a day.
- Low volume. At a dozen or so emails a day, a person handles the inbox for less than the deployment and its upkeep cost. The arithmetic above shows it at once.
- A plain filter is enough. If the problem is only sorting mail into a few buckets, filter and label rules in Gmail or Outlook do that for free. Do not build an LLM for it.
- A team agreement is enough. Sometimes the inbox "does not work" not because AI is missing, but because nobody is assigned to it. Agreeing who owns office@ and until what hour is sometimes the whole fix.
- The rules live in someone's head. If nobody can describe when a case is routine and when it is an exception, it has to be written down first. Writing the rules down is often most of the work before AI enters the picture.
An email agent makes sense only where three conditions meet at once: hundreds of emails a day, repeatable case types with rules, and a real cost of the hours triage takes from your people. If one of these is missing, we will say so plainly on the scan before you spend anything. This is usually also the first step in wider AI process automation, not a standalone gadget.
How to start
The cheapest sensible first step is to calculate the inbox, not to buy a tool.
- Book a free process scan and show one shared inbox.
- Prepare: how many emails a day, which case types repeat, who reads them today, what may be sent without a human, and where the exceptions appear.
- After the call you get a recommendation: a plain filter, inbox automation, an agent, or an honest "not worth it yet."
Book a free process scan | AI automations | See pricing
FAQ
How much does an AI agent for email cost?
Automating a shared inbox with an LLM (classification, data extraction, draft replies with human approval) starts from €3,500 net. An agent that runs a case across several systems starts from €6,000 net. Maintenance is priced individually, and the model cost is a few cents per email times volume. The first step, a free process scan, costs €0.
How does automatic email sorting work?
The system reads every message in the inbox (office@, sales@), recognizes the case type and urgency, extracts key data (order number, VAT ID, deadline), applies a label, and routes the case onward. For routine cases it drafts a reply for approval. A process rule, not the model, decides what may be sent automatically and what waits for a human.
What should an email agent never send without a human?
Without human approval, these should never go out: quotes and prices, confirmations of commitments, replies to complaints and disputes, anything touching money, contracts, or sensitive data, and any email at low classification confidence. Auto-send is switched on only for narrow, stable case types with approved content.
When is an email agent NOT worth buying?
At a dozen or so emails a day, manual handling is cheaper than building and maintaining an agent. If filter and label rules in Gmail or Outlook, or simply agreeing who owns the inbox, would do, you do not need an agent. An agent makes sense only at hundreds of emails a day, repeatable case types, and a real cost of triage hours.
Is an email agent GDPR compliant?
An inbox holds personal data, so a deployment needs a data processing agreement (DPA) with every provider that handles the content, a defined scope and purpose, data minimization, and a clear retention policy: how long the system keeps content, logs, and the decision trail. That is a condition for production, not an add-on.