An AI support agent is a software layer that sits in front of a help desk, reads the inbound message, decides whether it can answer it confidently from a curated knowledge base, and either replies directly or routes the conversation to the right human with full context attached. Built well, it deflects 30–50% of tickets on a typical US SMB support inbox — billing questions, hours, scheduling, refund policy, "where is my order," account resets, basic how-to. Built badly, it lies to customers and burns the trust your support team spent years earning. This post is the build pattern Horsiq uses, the stack behind it, the steps, the costs, and the four ways it goes wrong.
Support volume is the quiet tax most US small businesses pay without measuring. A 12-person clinic fields 400 questions a week. A regional ecommerce brand routes 1,200. A SaaS doing $2M ARR sees its founders drift back into the inbox every Friday. The questions are 80% repeats. The team knows it. Nothing changes, because the fix has always been "hire someone" — and nobody wants another seat.
The 40% number — what it actually means
40% is not a marketing number. It's the median deflection rate we see on tier-1 inboxes after the agent has been live for 6–8 weeks and has had two rounds of correction.
Three things to be clear about:
- Tier-1 tickets only. Hours, locations, order status, password resets, refund policy, basic product questions, appointment changes. Not medical advice. Not legal. Not anything the agent could plausibly get wrong in a way that costs the business money.
- Resolved means the customer didn't reply again. Not "the agent said something." If the customer comes back with a follow-up question, that ticket counts as not resolved.
- The baseline matters. 40% of a clean inbox with a good FAQ is different from 40% of an inbox where the team has been answering "what are your hours" 30 times a day. The latter is easier. The deflection rate is highest where the operational chaos was worst.
If a vendor pitches you 80% deflection on a generic chatbot, they're either measuring "agent said words" or pitching a product that will hallucinate inside a week.
The 3-layer AI support pattern
The agents that hold up in production are not single prompts. They are three small agents wired in sequence, each doing one job.
Layer 1 — Triage. Reads the inbound message. Classifies it: routine question, account-specific question, sensitive (billing dispute, medical, legal, emotional), or escalation-only (cancellation, complaint, churn signal). Outputs a category and a confidence score. Does not answer anything. Its only job is to decide who answers.
Layer 2 — Knowledge-base retrieval. If triage says "routine," this layer runs. It searches a vector database of the company's actual answers — FAQ, policy docs, product specs, past support replies that were marked correct — and pulls the 3–5 most relevant passages. Then it asks Claude to draft a reply only from those passages. If the passages don't contain a clean answer, the agent must say so and hand off. No improvisation.
Layer 3 — Escalation router. When triage or retrieval flags handoff, this layer decides which human, attaches the full conversation, the customer record, the agent's draft answer if any, and the reason for escalation. The human opens the ticket already knowing the context. Average handle time on escalated tickets drops 40–60% on its own, because the human isn't reconstructing.
Three layers, not one. The temptation is to write one fat prompt that does everything. Don't. Each layer is testable in isolation, and when something goes wrong you know exactly which layer failed.
What you need to build it
The stack is boring on purpose. Nothing here is exotic.
- Claude API (Sonnet or Opus). The reasoning layer. Sonnet handles triage and retrieval-grounded replies cheaply. Opus for the harder cases or for offline evaluation.
- n8n. Self-hosted on a $20/month VPS. Glues the help desk, the vector store, and the Claude API together. Webhooks in, webhooks out. Visual flow, version-controlled.
- A help desk with a webhook API. Intercom, HelpScout, Zendesk, Front, Freshdesk. All work. If the inbox is still raw Gmail, fix that first — you can't measure deflection in Gmail.
- A vector database for the knowledge base. Supabase pgvector or Pinecone. Hold maybe 200–2,000 chunks for a typical SMB. Embeddings via Voyage or OpenAI.
- A real knowledge base. The hard part. Not a 14-month-old FAQ. The actual answers your team gives, written in your voice, marked correct. If you don't have this, building the agent is the second job — building the KB is the first.
Step-by-step build (the Horsiq playbook)
- Audit two weeks of inbox. Export tickets. Tag the top 20 question types by volume. If the top 10 cover more than 60% of tickets, the project pays back. If the long tail dominates, the agent will deflect less and the build needs to be scoped smaller.
- Build the knowledge base. For each of the top 20 question types, write the canonical answer. Pull the best past replies from the inbox, edit for clarity, mark them as the source of truth. Chunk into 300–500 token passages. Embed. This is 60% of the project.
- Write the triage prompt. Three categories: routine, sensitive, escalation. Few-shot it with 30–40 real examples from the audit. Run it offline against a held-out set of 200 tickets. Iterate until classification accuracy is above 95%. Don't ship below that.
- Wire retrieval. Top-k of 5, similarity threshold tuned to your data. The retrieval prompt instructs Claude to answer only from the passages, to cite which passage it used, and to say "I don't have a confident answer for this" if the passages don't cover it. That last instruction is the difference between a useful agent and a liability.
- Build the escalation handoff. When the agent hands off, the human receives the original message, the triage category, the retrieved passages, the agent's draft (if any), and the reason for handoff. This single integration is where most "AI for support" projects underdeliver — they hand off without context and the human starts from zero.
- Shadow mode for 5–7 days. The agent runs on every inbound ticket, drafts a reply, and posts it as an internal note. Humans read both, send the human reply, and rate the agent's draft. You'll find the prompts that misfire, the KB gaps, the categories you missed. Fix before going live.
- Go live with a conservative confidence threshold. Auto-send only the highest-confidence categories. Everything else stays in shadow. Loosen the threshold weekly based on actual customer satisfaction signals, not your gut. Plan for 6–8 weeks before the deflection rate stabilizes.
What can go wrong
1. Generic chatbot disease. The team buys a SaaS chatbot with a pre-trained "ecommerce" or "healthcare" template. It confidently quotes pricing that doesn't match the website, hours that don't match the door, policies that don't match the contract. Trust dies in a week. The fix is to never ship an agent on top of generic context — only on top of your answers.
2. No "I don't know" path. The agent is prompted to be helpful, so it answers everything, including things it shouldn't. Every retrieval-grounded reply needs an explicit escape hatch: if the passages don't contain a clean answer, hand off. Test this path more than the happy path.
3. Aggressive auto-send on day one. The agent goes live full-volume on a Friday afternoon. By Monday, three customers have been told the wrong refund policy. Start conservative. Tighten the auto-send list. Loosen on evidence.
4. No feedback loop. The agent ships, then the team forgets about it. Six months later the KB is stale, the auto-send categories haven't been revisited, and deflection has quietly drifted from 40% to 18%. A weekly "agent didn't know" report and a monthly KB review is the difference between an asset and a decaying script.
Cost & timeline reality
Build cost. $5,000–$15,000 for a typical SMB build. The range is driven by the state of the knowledge base going in. If you have clean, current answers, it's the low end. If we're writing the KB from inbox archaeology, it's the high end. Voice agents (phone, not chat) add roughly $3,000–$8,000 on top.
Runtime cost. $100–$500/month for most SMBs. Breakdown: Claude API calls scale with volume (assume $0.01–$0.05 per resolved ticket on Sonnet), vector DB hosting is $20–$80, n8n VPS is $20, help desk seat is whatever you already pay. No per-conversation platform tax. No SaaS markup.
Timeline. 4–8 weeks from audit to live. 6–8 weeks more before the deflection rate stabilizes. Anyone promising "AI support agent in a week" is selling a template, not a system.
Payback. If your team spends 15 hours a week on tier-1 tickets at a $30/hour fully-loaded cost, a 40% deflection saves roughly $9,360 a year per support seat. The build pays back in 6–18 months and keeps paying.
Horsiq builds these systems for US small businesses — clinics, ecommerce, services, B2B SaaS. The agent runs on your accounts, your API keys, your help desk. You own it on day one. See our AI Automation approach, or scope a build.