DEV Community

T.M. Gunderson
T.M. Gunderson

Posted on

I Built an AI Agent That Handles 70% of My Customer Support (Here's the Exact Setup)

You know what kills most AI automation projects?

Perfectionism.

Small business owners (and honestly, developers too) think an AI agent needs to handle 100% of a task before it's worth deploying. So they spend weeks tweaking prompts, testing edge cases, and arguing about which platform to use — and never ship.

Here's the pattern I've noticed across automation projects: the businesses that win with AI are the ones that deploy at 70% accuracy and improve from there.

Not 100%. Not 95%. Seventy percent.

Why 70% Is the Magic Number

A customer support agent that handles 70% of tickets automatically still saves you 4+ hours per day. A scheduling agent that books 70% of appointments still eliminates 2+ hours of phone tag. A reporting agent that generates 70% of your weekly numbers still saves 3+ hours of spreadsheet work.

The remaining 30% is where human judgment matters. The furious client who needs a real person. The appointment with weird constraints. The report with an anomaly.

Ship at 70%. Review the 30%. Improve weekly. The alternative — waiting for 100% — means you ship nothing.

The 5-Day Sprint: From Zero to Deployed Agent

Here's a realistic timeline for getting your first AI agent live:

Day 1: Pick Your Agent Type

Start with the task that eats the most time and has the most repetitive patterns. For most small businesses, that's one of these:

Agent Type Time Saved Monthly Cost Setup Time
Customer support auto-responder 4+ hrs/day $5-20 2-3 hours
Invoice follow-up bot 3-5 hrs/week $0-20 1-2 hours
Meeting notes & action items 1-2 hrs/week $0-10 30 minutes
Email triage & drafting 2-3 hrs/day $0-5 1 hour
Review request automation 1-2 hrs/week $0 15 minutes

Pick one. Not three. Not five. One.

Day 2: Write Your Prompt

Here's the exact prompt structure that works for customer support (adapt the brackets for your business):

You are a customer support agent for [BUSINESS NAME], a [INDUSTRY] company.

Read the customer message and:
1. Categorize: billing, technical, general inquiry, complaint, or urgent
2. Draft a response using this tone: helpful, concise, professional
3. If the issue involves refunds over $[AMOUNT], account changes, or legal concerns → flag for human review
4. Otherwise → prepare to send

Our policies:
- Refunds under $[AMOUNT]: approve automatically
- Refunds over $[AMOUNT]: escalate to manager
- Technical issues: provide troubleshooting steps from our FAQ
- Response time SLA: under 2 hours during business hours

Customer message: [PASTE HERE]
Enter fullscreen mode Exit fullscreen mode

The key elements every good agent prompt needs:

  • Role definition — "You are a customer support agent for..."
  • Decision framework — categorize, then route
  • Escalation triggers — know when to hand off to a human
  • Business-specific policies — not generic advice, YOUR rules
  • Confidence scoring — if unsure, flag for review

Day 3: Set Up the Workflow

This is where most people overcomplicate things. You need exactly three things:

  1. Trigger — New email/chat message arrives
  2. AI step — The prompt above processes the message
  3. Action — Send response OR queue for human review

On Make.com (free tier available):

  • Create a new scenario
  • Gmail module → "Watch emails" (trigger)
  • OpenAI module → "Create chat completion" (AI step)
  • Gmail module → "Send email" OR Google Sheets → "Add row" (action)

On Zapier ($20/mo):

  • Trigger: New email in Gmail
  • Step: OpenAI — send prompt
  • Step: Gmail — send draft OR add to review spreadsheet

On n8n (self-hosted, free):

  • Same workflow, but you control the hosting
  • Best for businesses with data privacy requirements
  • Self-hosted means no monthly platform fee

Don't overthink the platform. Pick the one you're most comfortable with. You can always migrate later.

Day 4: Test With Real Data

Run 20-30 real messages through your agent. Don't use synthetic test data — use actual customer messages from the past month.

Track these metrics:

  • Response accuracy — Does the answer actually address the customer's question?
  • Tone match — Does it sound like your business?
  • Escalation rate — How often does it flag for human review?
  • Time saved — How long would a human have taken on each?

You're looking for 70% accuracy or better on the first three metrics. If you're below that, refine your prompt — but don't aim for 95%. Good enough to ship is good enough.

Day 5: Deploy and Set Your Review Cadence

Turn it on. Set up a daily review of the 30% your agent flagged for human attention. Here's the cadence that works:

Time After Launch Review Frequency What to Review
Week 1 Every 4 hours All agent responses
Week 2 Twice daily Flagged items only
Week 3-4 Once daily Flagged items only
Month 2+ Once weekly Edge cases and failures

After week 2, you should be spending 15-20 minutes per day reviewing flagged items. That's your 70% → 85% improvement path.

The Cost Breakdown (Real Numbers)

People ask "how much does an AI agent actually cost to run?" Here are real numbers from businesses running these workflows:

Component Free Option Paid Option
AI API (OpenAI) Free tier (limited) $5-50/month
Automation platform n8n (self-hosted) Make ($9/mo) or Zapier ($20/mo)
Email/chat integration Gmail (free) Same
Review spreadsheet Google Sheets (free) Same
Total monthly $0-5 $15-70

Even at the high end ($70/month), if your agent saves 4 hours per day at a typical small business labor rate ($30-50/hr), you're looking at:

$30/hr × 4 hrs/day × 22 days = $2,640/month in time value

That's a 37x return on the paid option. Even conservatively (2 hours saved, $25/hr), it's still a 15x return.

Where This Breaks Down (And What to Do About It)

The 70% approach isn't perfect. Here's where it fails and how to handle it:

1. Your agent confidently gives wrong answers

  • Fix: Add explicit "if unsure, say you'll check and follow up" to your prompt
  • Fix: Add a confidence threshold — below 80%, always escalate

2. Edge cases eat all your time

  • Fix: Track edge cases weekly. After 3 occurrences of the same type, add it to your prompt's policies
  • Fix: Your prompt should grow over time. Add 1-2 rules per week based on real failures

3. Customers notice they're talking to AI

  • Fix: Add "I'll have a team member follow up on this" to low-confidence responses
  • Fix: Match your brand voice in the prompt. Most customers don't care if they get a fast, accurate answer

4. Costs spiral as volume grows

  • Fix: Monitor your API usage. Most agents cost $5-20/month even at high volume
  • Fix: Switch to GPT-4o-mini for simple tasks (10x cheaper than GPT-4)

What to Automate First (Decision Matrix)

Not every task is worth automating. Use this framework:

High Volume Low Volume
Repetitive ✅ Automate first 🟡 Automate if painful
Variable 🟡 Use 70% rule ❌ Don't automate yet

High-volume repetitive tasks are your bread and butter. Start there. That's where 70% accuracy saves the most time.

Want the Full Playbook?

This post covers the framework. If you want the complete implementation guide — all 12 workflows, every prompt, the 5-day sprint plan, cost calculator, security checklist, and scaling playbook — it's in the Small Business AI Agent Starter Kit ($59).

Or start free with the AI Automation Cheat Sheet — 10 prompts you can use today.


The best AI agent is the one that's running. Not the one you're still designing. Ship at 70%, review the 30%, improve weekly. That's how small businesses actually get value from AI — not by waiting for perfection.

What's the first task you'd automate? Drop it in the comments — I'll suggest a prompt for it.

Top comments (0)