Support queues have a shape. A large share of tickets are the same ten questions with different names attached. A smaller share need judgement. A tiny share are emergencies. Human agents spend most of their day on the first group, which is precisely the group AI agents are good at.
This guide explains how an AI support agent actually works (it is not a chatbot), which tickets it should and should not handle, how to roll it out without a CSAT dip, what integration looks like with the tools you already run, and what it costs for teams in the US, UK and Australia.
Chatbot vs. AI agent
The chatbot era trained everyone to expect a widget that says "I didn't understand that". An AI agent is a different thing:
| Chatbot | AI support agent | |
|---|---|---|
| Understands free text | Limited | Yes (LLM) |
| Answers from your docs | Sometimes | Yes, with citations |
| Looks up an order or account | Rarely | Yes, via API |
| Takes actions (refund, reschedule, update) | No | Yes, within limits |
| Writes a personalised reply | Template | Yes |
| Routes and tags tickets | Basic rules | Yes, with reasoning |
| Knows when to escalate | Keyword-based | Confidence-based |
The agent is a language model with tools: read the ticket, search the knowledge base, call the order API, decide, act, write, log. That loop is what makes it capable of resolving a ticket rather than just responding to it.
What an AI agent should resolve end-to-end
Ticket types that are safe to fully automate once accuracy is proven:
- Order and delivery status
- Password resets and login help
- "How do I..." questions covered by documentation
- Billing questions (invoice copies, payment dates, plan details)
- Appointment rescheduling within policy
- Returns and refunds within a defined value limit
- Account updates (address, contact details) with verification
Ticket types that should get a drafted reply and human approval:
- Complaints and anything with negative sentiment
- Refunds above the limit or outside policy
- Technical bugs that need investigation
- Anything involving legal, medical or financial advice
- VIP or enterprise accounts
Ticket types that should escalate immediately:
- Safety issues
- Data breach or security concerns
- Threats of legal action or churn from high-value accounts
- Anything the agent's confidence score flags as uncertain
Across SMB support queues we have analysed, 40-70% of tickets fall into the first group. The exact number depends on how good your documentation is - better docs mean higher resolution rates. Improving the knowledge base is usually the highest-leverage week of the whole project.
A safe rollout in four stages
The mistake teams make is switching on full autonomy on day one. This is the sequence that protects CSAT:
Stage 1: Shadow mode (week 1-2)
The agent reads every ticket and writes a draft, but nothing is sent. Your team compares drafts to what they would have written. You get an accuracy baseline and a list of documentation gaps.
Stage 2: Draft-and-approve (week 3-6)
Agents see the AI draft in their help desk and send, edit or discard it with one click. Handling time drops immediately. Every edit is logged and used to tune the prompts and knowledge base weekly.
Stage 3: Autonomous on proven categories (week 7+)
Ticket types with a 95%+ approve-without-edit rate for two consecutive weeks go autonomous. Everything else stays in draft mode. The customer always sees a "talk to a person" option.
Stage 4: Expand and tune
Each week, review the categories just below the threshold, fix the underlying documentation or logic, and promote them. Review escalations to catch anything the agent is over- or under-escalating.
Integration: what connects to what
A typical setup for a US, UK or Australian SMB:
- Inbound channels: help desk (Zendesk, Intercom, Freshdesk, HubSpot, Help Scout, Gorgias), shared inbox (Gmail/Outlook), WhatsApp Business, website chat
- Knowledge: help centre articles, internal docs, past resolved tickets (with PII removed), policy documents
- Systems the agent can act on: order management or e-commerce platform, billing (Stripe, Chargebee), CRM (HubSpot, Salesforce), scheduling
- Guardrails: action limits (refund cap, no deletes), confidence thresholds, PII redaction, full audit log
- Reporting: weekly accuracy, resolution rate, response time, CSAT by category
Most of this runs on the help desk's own API plus an orchestration layer. No rip-and-replace.
What it costs
| Scope | USD | GBP | AUD | CAD | NZD |
|---|---|---|---|---|---|
| Managed AI support agent, per month | $2,000 - $5,000 | £1,550 - £3,900 | A$3,000 - A$7,500 | C$2,700 - C$6,800 | NZ$3,300 - NZ$8,250 |
| Model / API usage, per month | $50 - $500 | £50 - £400 | A$100 - A$750 | C$50 - C$700 | NZ$100 - NZ$850 |
| One-off build (no ongoing management) | $6,000 - $20,000 | £4,700 - £15,600 | A$9,000 - A$30,000 | C$8,150 - C$27,200 | NZ$9,900 - NZ$33,000 |
Non-USD figures are indicative conversions; we invoice in USD, GBP or AUD.
For comparison, a support agent in the US costs $45,000-$65,000 per year fully loaded; in the UK £28,000-£38,000; in Australia A$60,000-A$80,000. An AI agent that resolves half the queue typically replaces the growth in headcount rather than existing staff - the team stays the same size while volume doubles.
Metrics that matter
Track these weekly from day one:
- First-response time (target: under one minute for automated categories)
- Resolution without human touch (percentage of tickets)
- Approve-without-edit rate on drafts (your accuracy proxy)
- Escalation precision (were escalations actually needed?)
- CSAT by category - split AI-resolved and human-resolved
- Handling time for human-resolved tickets (should drop 30-50% thanks to drafts)
Working with GridCore
We build and run AI support agents for businesses in the US, UK and Australia as part of our AI Agents and Full-Service AI offerings. The engagement follows the four-stage rollout above, with weekly tuning and a report you can read in two minutes. Data stays in your region, and you own the whole system.
If you want a realistic estimate of what share of your queue could be automated, send us a sample of anonymised tickets and we will tell you. Related reading: AI automation for small business.
