How to Build an AI Email Triage Agent for Your Service Business

The average professional spends 3.2 hours a day on email. For a UK service business — a consultancy, agency, or practice — that's not a productivity problem. It's a billable-hours problem. An AI email triage agent changes the equation: classify every inbound message, draft the replies that follow templates, escalate what needs a human, and archive what doesn't. Here's exactly how to build one that runs in the background while you do the work only you can do. Email volume hasn't declined. It's accelerated. The typical UK knowledge worker receives 126 emails a day in 2026 — a 34% increase since 2023, driven by AI-generated outreach, automated notifications, and clients who email instead of calling. Most of those emails don't need you specifically. They need a response that follows a pattern you've written a hundred times before. The cost of processing them yourself isn't just time. It's context — every time you stop work to read an email, it takes an average of 23 minutes to return to full concentration. An AI triage agent handles the classification and drafting. You handle the judgment calls. That split is where the time saving comes from. <h2>What an AI Email Triage Agent Actually Does</h2> <figure> <img src="https://images.unsplash.com/photo-1516321165247-4aa89a48be55?w=1200&q=80" alt="Diagram showing the four functions of an AI email triage agent: classify, draft, escalate, and archive" width="1200" height="800" loading="lazy" /> </figure> Before you build anything, be clear about what the agent does and doesn't do. It doesn't send emails autonomously — that's the line most service businesses won't want to cross, and for good reason. What it does is handle the four functions that currently consume most of your inbox time: <ul> <li>Classify. Every inbound email is categorised by type: new enquiry, existing client request, supplier/admin, newsletter, or noise. Classification happens within seconds of arrival, with a priority flag (urgent, normal, low) attached.</li> <li>Draft. For emails that need a response following a predictable pattern — meeting confirmation, document request, status update, standard question — the agent drafts a reply in your voice and drops it into your drafts folder ready to send with one click.</li> <li>Escalate. Anything genuinely urgent, emotionally sensitive, or outside the agent's competence is flagged clearly and delivered to you with a one-sentence summary of what it needs and why it can't be handled automatically.</li> <li>Archive. Newsletters, automated notifications, and marketing emails that don't need your attention are filed away automatically, keeping your inbox clear without you touching them.</li> </ul> The result isn't inbox zero. It's inbox irrelevance — your inbox stops being a place you need to monitor constantly and becomes a queue of decisions that genuinely require your judgment, curated and ready to process in a single focused session. <h2>What You Need Before You Start</h2> You don't need to be a developer to build this. You need four things: <ul> <li>An email account with API access. Gmail (via Google Workspace) or Microsoft 365 (Outlook via Graph API). Both work. Gmail is slightly easier to start with — the setup is more documented and the OAuth flow is more forgiving.</li> <li>A workflow automation tool. n8n (self-hosted or cloud, £20–40/month) or Make (from £9/month). Either works for this build. We use n8n because it handles more complex logic and is easier to debug at scale, but Make is quicker to get started with.</li> <li>An LLM API key. Claude 3.5 Sonnet is the most reliable for following tone guidelines and producing email copy that sounds human. Gemini 1.5 Flash is a cost-effective alternative if you're processing very high volumes.</li> <li>Roughly four hours for the initial build. Once you've done it once, subsequent iterations take 30–60 minutes.</li> </ul> If you've already built an <a href="/blog/build-ai-lead-qualification-agent">AI lead qualification agent</a> or an <a href="/blog/automate-sales-proposals-ai-agents">AI proposal writer</a> using our guides, you'll recognise the architecture. The email triage agent uses the same core components — trigger, classify, act — with a few email-specific additions. <h2>Step 1: Define Your Email Categories and Response Templates</h2> This is the step most people skip, and it determines whether the agent is useful or frustrating. Before you configure anything, spend 30 minutes classifying your last 100 emails by hand. You'll almost certainly find that 80% of your inbox falls into five or six recurring types. For a typical UK consultancy or agency, those categories look like this: <ul> <li>New enquiries. People asking about your services for the first time. High priority — these need a response within two hours.</li> <li>Client requests. Existing clients asking questions, requesting documents, or following up on work in progress.</li> <li>Meeting and scheduling. Confirmation requests, rescheduling, calendar invites.</li> <li>Invoicing and admin. Remittance advice, purchase orders, payment queries.</li> <li>Supplier and tools. Platform notifications, software updates, SaaS invoices.</li> <li>Noise. Marketing emails, newsletters, automated alerts you don't read.</li> </ul> For each category that needs a response, write a response template. Keep it conversational — not formal, not stiff. The agent will adapt the template to the specific email, but it needs a clear starting point. A template for "new enquiry" might be: "Thanks for getting in touch. I'd love to hear more about what you're working on — could you share a bit of context about your business and what you're hoping to achieve? Happy to jump on a call if that's easier." <blockquote>The quality of your templates determines the quality of your drafted replies. A template that sounds like you will produce drafts that sound like you. A template that sounds like a press release will produce drafts you delete.</blockquote> <h2>Step 2: Build the Classification Workflow</h2> <figure> <img src="https://images.unsplash.com/photo-1518770660439-4636190af475?w=1200&q=80" alt="n8n workflow diagram showing email trigger, AI classification node, and routing logic for an AI email triage agent" width="1200" height="800" loading="lazy" /> </figure> In n8n, the workflow starts with a Gmail or Outlook trigger that fires every 5–15 minutes. The trigger pulls in new emails and passes each one — subject line, sender, first 500 characters of body — to an LLM classification node. The classification prompt is straightforward. Here's the exact prompt structure we use: <pre><code>You are an email triage assistant for a UK {business_type} business. Classify the following email and respond with JSON only: - category: one of [new_enquiry, client_request, meeting_scheduling, invoicing_admin, supplier_tools, noise] - priority: one of [urgent, normal, low] - requires_response: true or false - summary: one sentence describing what the email is about - suggested_action: brief description of what should happen next Email details: From: {sender_email} Subject: {subject} Body preview: {body_preview}</code></pre> Route based on the category field. Urgent new enquiries route to your drafting workflow immediately. Noise routes directly to archive. Everything else gets classified and queued. One important detail: include the sender's email domain in your prompt context. An email from @microsoft.com or @hmrc.gov.uk has different handling rules than one from a Gmail address. Domain-aware classification dramatically reduces false positives in the noise category. <h2>Step 3: Build the Drafting Workflow</h2> For categories that need a drafted response, a second LLM call takes the classification output, the full email content, and the appropriate response template, and generates a first-draft reply. The drafting prompt includes three critical elements: your response template for that category, a brief description of your communication style ("direct, warm, plain English — never formal, no corporate jargon"), and any specific information from the email that should be referenced in the reply. The draft is saved to your Gmail drafts folder or Outlook drafts via the relevant API node. In n8n, this is a single node — the "Save Draft" action. The email appears in your drafts with the subject "RE: {original subject}" and is flagged with a label ("AI Draft") so you can review and send them in a single focused session rather than context-switching throughout the day. For most UK service businesses, 60–70% of inbound emails that need a response will have usable AI drafts. You review, adjust where needed, and send. The remaining 30–40% — emotionally nuanced messages, complex client situations, new business negotiations — are flagged for your full attention. <h2>Step 4: Configure Escalation and Safeguards</h2> This is the step that separates a useful agent from a liability. Escalation rules define what the agent cannot handle — and those rules need to be explicit. Hard escalation triggers that should always go directly to you, unfiltered: <ul> <li>Any email containing words like "complaint", "legal", "solicitor", "GDPR request", or "data subject"</li> <li>Any email from a known client address that contains language suggesting dissatisfaction ("disappointed", "not what I expected", "you said")</li> <li>Any email from a prospect previously marked as a priority opportunity</li> <li>Any email where the classification confidence is below 80%</li> </ul> Build these as filter conditions before the classification step — pattern matching on keywords is faster and more reliable than asking an LLM to spot legal risk in every email. LLMs are excellent at drafting; they're less reliable at consistently catching regulatory risk signals when processing at volume. This same principle applies across all agent builds. Our guide to <a href="/blog/ai-agent-security-prompt-injection">AI agent security</a> covers the broader defence stack you should have in place, particularly the prompt injection risks that apply when your agent is processing external email content. <h2>Step 5: Test on Two Weeks of Historical Email</h2> Before going live, run the agent against two weeks of your historical inbox — without it taking any action, just generating classifications and draft outputs for your review. This is the same parallel-run approach we use in every agent build we do for clients. What you're checking: classification accuracy (are new enquiries actually being classified as new enquiries?), draft quality (would you send these replies?), and escalation reliability (are sensitive emails hitting the escalation list?). In our experience, the first pass gets classification right about 85–90% of the time. You'll find two or three specific patterns that need refinement — usually edge cases in your specific industry that the initial prompt didn't anticipate. Fix those, re-run, and accuracy typically comes in above 94%. Don't skip this step. A week of parallel testing costs nothing and prevents the agent from misclassifying an important client email on day one. <h2>What to Expect After Two Weeks</h2> <figure> <img src="https://images.unsplash.com/photo-1611974789855-9c2a0a7236a3?w=1200&q=80" alt="Before and after stats showing 3.2 hours daily email time reduced to 45 minutes with an AI email triage agent for UK service businesses" width="1200" height="800" loading="lazy" /> </figure> The numbers vary by inbox volume and email type, but across the UK service businesses we've built this for, the pattern is consistent: <ul> <li>Email time drops from 3+ hours daily to 45–60 minutes. Most of that remaining time is review and send — reading AI drafts and approving them, not composing from scratch.</li> <li>New enquiry response time drops to under 15 minutes. The agent drafts a reply the moment an enquiry arrives. You review it at your next email session and send. For service businesses where speed of response materially affects conversion, this is the highest-ROI function of the whole system.</li> <li>Client satisfaction improves. Clients notice faster, more consistent responses. The draft quality, when tuned to your voice, is indistinguishable from responses you'd write yourself.</li> <li>Running cost: under £30 a month. At typical UK service business email volumes (100–200 emails a day), LLM API costs for classification and drafting run at £15–25 a month. Add n8n cloud hosting and the total is well under £50.</li> </ul> For context on how to keep those LLM costs low as volume scales, the <a href="/blog/ai-agent-cost-optimisation-uk">AI agent cost optimisation guide</a> covers the three approaches — caching, routing, and compression — that we apply to every build. <h2>Connecting It to the Rest of Your AI Operating System</h2> <figure> <img src="https://images.unsplash.com/photo-1553877522-43269d4ea984?w=1200&q=80" alt="Confident UK business owner reviewing a clean AI-triaged inbox with most emails handled — representing the time saving an AI email agent delivers to service businesses" width="1200" height="800" loading="lazy" /> </figure> The email triage agent integrates naturally with the other agents in an AI Operating System. If you've already built an <a href="/blog/build-ai-lead-qualification-agent">AI lead qualification agent</a>, the two can work in tandem: the email triage agent classifies an inbound enquiry and passes it to the lead qualification agent, which scores it and decides the appropriate follow-up sequence. The <a href="/blog/automate-sales-proposals-ai-agents">AI proposal writer</a> can receive a qualified lead and prepare a first-draft proposal — all without you touching the keyboard until it's time to review and send. The <a href="/blog/automate-client-reports-ai-agent">AI client reporting agent</a> handles the outbound side: regular status updates going to clients without manual effort. That's what an AI Operating System looks like in practice: not one agent, but agents that hand work to each other so that a new enquiry arriving on a Tuesday afternoon can be classified, qualified, and have a draft proposal ready for your review before you finish your next client call. <blockquote>The email agent is often where teams start because the ROI is immediate and visible. But its real value emerges when it's the first node in a connected system — the front door that routes work to every other agent running in your business.</blockquote> <h2>Build It or Let Us Build It for You</h2> If you're ready to work through the build yourself, everything you need is in this guide. Allow four hours for the initial setup, plan for a week of testing, and expect to spend a further couple of hours refining the classification and escalation rules once you've seen the agent running on real email. If you'd rather have the build done in a week with no trial and error — and have it integrated with any other agents you already have running — that's exactly what we do. <a href="/contact">Book a free 30-minute call</a> and we'll map your inbox patterns, identify your highest-value automation opportunities, and give you a clear picture of what an AI email agent would save you in time and money. No obligation — just a clear view of what's possible for your business.