AR · July 2026 · 9 min read

Setting one up takes an afternoon. Setting one up well takes a fortnight.

The order of operations that decides whether your AI agent helps or just annoys people into leaving.

Most teams get an AI agent answering questions on day one and regret it by week three. Not because the tool was bad. Because they pointed it at the wrong tickets, hid the escape hatch, and measured the wrong number.

Before you touch a tool

Two decisions matter more than which platform you buy, and both are free to make. First, which tickets you are automating. Second, what happens when the agent cannot help. Teams that skip these buy software to solve a problem they have not defined.

1. Count your tickets by type

Export a month of tickets and sort them into buckets. Most teams find that four or five categories cover 70% of volume, and that the largest is some version of where is my order or how do I do X.

This tells you two things: whether an AI agent can help you at all, and which capability you actually need. If your top category needs an order lookup, a tool that only reads help articles is useless regardless of how well it writes.

2. Fix your documentation first

An AI agent answers from your help centre. If the answer is not written down, the agent invents one or escalates. Teams that spend a week improving their top thirty articles get more from a cheap tool than teams that skip it get from an expensive one.

3. Pick the pricing model before the product

ModelExampleBest when
Per resolutionFin $0.99Under ~2,000 monthly resolutions
Per seatFreshdesk $18-95High volume, stable team size
Per ticketGorgias $10-900Predictable volume, seasonal staff
By contactHelp Scout $55-83Many agents, few customers

At 10,000 monthly resolutions, per-resolution pricing costs $9,900 and exceeds the staff it replaces in most markets. That is a decision you make now, not after the first invoice.

4. Start with one ticket type

Order status, or your single highest-volume documented question. One category, live, measured for two weeks. Resist the urge to switch everything on because the demo looked good.

5. Make escalation obvious

The most common complaint in consumer AI support is that customers cannot find a human. Put the escape hatch in the first message, not after three failed attempts. Your deflection metric will look worse and your customers will be happier, which is the correct trade.

6. Test the handoff, not the conversation

Escalate deliberately during evaluation and look at what the human agent receives. If the customer has to repeat themselves, you have added a step rather than removed one. Platforms differ sharply here and demos never show it.

7. Measure resolution, not deflection

Deflection counts a customer who gave up as a success. Ask your platform for resolution confirmed by the customer, and treat the gap between the two numbers as your error rate.

Sequencing here comes from how each support job behaves and what failure costs, not from benchmarking platforms. We do not run tests.

Frequently asked

How long does it take to set up AI customer service?

A basic agent answering from your help centre takes hours on tools like Chatbase or Tidio. Doing it properly, with documentation cleaned up, one ticket type live and escalation tested, takes about two weeks.

What do I need before starting?

A month of exported tickets sorted by type, your top thirty help articles in reasonable shape, and a decision on pricing model. All three are free and all three matter more than the tool choice.

Which tickets should AI handle first?

Order status, policy questions and other documented, system-answerable queries. Automating angry customers or technical diagnosis first is the most common and most damaging mistake.

Do I need a developer?

For deflection from help articles, no. For anything that queries your order system or takes an action, yes: custom actions are engineering work that teams routinely underestimate.

How do I know if it is working?

Confirmed resolution rate rather than deflection, hallucination count, and whether escalated conversations arrive with context. If deflection rises while satisfaction falls, customers are giving up rather than being helped.

What is the most common setup mistake?

Hiding the escalation path. It flatters the deflection number and makes customers considerably angrier than never offering AI at all.

Should I automate everything at once?

No. One ticket type, measured for two weeks, then expand. Switching everything on at once means you cannot tell what is working.

How much should I budget?

Between $0 and $9,900 a month for the same volume depending on pricing model. Model your actual ticket count against per-resolution, per-seat and flat-rate options before signing.

Can I use AI customer service with my existing help desk?

Yes. eesel and My AskAI layer on top of Zendesk, Freshdesk and others without migration, which avoids the largest cost of adopting AI support.

What if my documentation is bad?

Fix it first. The agent answers from your content, so poor documentation caps the quality of every answer regardless of which platform you buy.

Tools mentioned

Full reviews, pricing tiers and where each one breaks.

You can also look into

Why this exists. PAA returns 'How can we use AI in customer service?' and 'How to use AI to improve customer service?' across multiple queries. Every ranking guide is vendor-authored.

WRITTEN BY AR · UPDATED 2026-07-29

I read the fine print. Vendor pricing pages, billing definitions, terms, funding filings and acquisition notices — then I do the arithmetic nobody publishes: what a platform actually costs at your volume, what its headline metric is really counting, and who owns it now. I do not run benchmarks, and no page here pretends otherwise.

How we test · About the author