Blog AI and automation

AI in customer support: what actually works

Triage, sourced drafts and knowledge base search: where AI genuinely helps the team, and where a person still has to decide.

AI in customer support: what actually works
In this article
  1. What "AI in support" means today
  2. How we got here
  3. What works: 3 applications that improve service
  4. 1. Ticket classification and routing
  5. 2. Suggestions to the agent while they write
  6. 3. Intelligent knowledge base search
  7. What does NOT work: 3 kinds of AI that frustrate customers
  8. ❌ 1. The chatbot with no way out
  9. ❌ 2. AI that "hallucinates" information
  10. ❌ 3. An intrusive proactive chatbot
  11. The hybrid model: AI and people
  12. Three levels
  13. Transparent handoff
  14. How to work out whether it pays off
  15. How costs work in Uppo
  16. How to start, over two weeks
  17. Week 1: foundations
  18. Week 2: expansion
  19. In conclusion: AI and people work together

What "AI in support" means today

"AI in customer support" has become a label covering very different things, from the chatbot that traps you in a menu to the assistant that saves an agent half an hour. It is worth telling them apart.

How we got here

First generation: rule-based chatbots (if-then-else)

  • ❌ Rigid
  • ❌ Frustrating: off-script, they understand nothing
  • ❌ The customer immediately looks for a way to reach a person

Second generation: generative chatbots with no context

  • ⚠️ Flexible but inaccurate
  • ⚠️ They make answers up (the famous hallucinations)
  • ⚠️ They do not know your company, your policies, your prices

Today: hybrid AI, with a person in the loop

  • ✅ AI can use company content as sources when drafting a reply
  • ✅ It assists the agent instead of replacing them
  • ✅ With Uppo Assist, an operator reviews and sends replies; AI Agent can reply autonomously within configured rules
  • ✅ Provide a path to a person when the customer needs further help

What works: 3 applications that improve service

1. Ticket classification and routing

How it works: the model reads the ticket as soon as it arrives and proposes

  • Category: technical / administrative / pre-sales / bug
  • Priority: urgent, high, medium, low
  • Assignee: the team or the person with the right skills

Why it works: as long as a ticket sits in a generic queue with no owner, nobody is looking at it. Classifying it on arrival sends it straight to whoever can solve it instead of letting it bounce around. The same assignments can be made with fixed rules and no AI at all: those are the five helpdesk automation workflows.

What to ask the vendor: is classification on by default, or do you decide? How much does it cost per ticket? And above all: can the agent correct it? In Uppo, automatic triage exists but is off by default, is switched on deliberately and has a configurable monthly request cap.

2. Suggestions to the agent while they write

How it works: while the agent reads the ticket, the assistant proposes

  • Similar replies already sent in the past
  • Relevant knowledge base articles
  • A draft reply to review, not to send with your eyes closed

Why it works: the agent starts from context rather than a blank page, and relevant knowledge base articles are easier to find in the context of the request.

The non-negotiable condition: with Uppo Assist, an operator reviews and sends drafts; AI Agent may reply autonomously within the configured rules.

How it works: search stops looking for the exact word and looks for meaning

  • Before: the customer searches "reset pwd" → zero results, because the article is titled "changing your password"
  • After: search understands the intent and finds the right article even with different words

Why it works: a search that ends in an article is a ticket that never arrives. It is the lever that carries growth without adding people, and how to build it is covered in the guide to the knowledge base and self-service portal; the product side in Uppo is the knowledge base.

What to ask the vendor: which sources does the AI use, and what happens when information is missing? Uppo uses company content to find relevant information and support the response. Keep sources up to date and check when the conversation needs your team’s involvement.

What does NOT work: 3 kinds of AI that frustrate customers

❌ 1. The chatbot with no way out

The problem: the customer is stuck going round in circles and cannot find a way to reach anyone.

Customer: "My order hasn't arrived"
Bot: "I can help you with order tracking! What is your order number?"
Customer: "I don't have a number, I want to speak to a person"
Bot: "I can help you with order tracking! What is your order number?"
Customer: gives up

What is needed: a way out that is always visible, along the lines of "Talk to a person", not buried three menus deep.

❌ 2. AI that "hallucinates" information

The problem: a model left free to answer from memory, without drawing on your content, fills the gaps by inventing.

Customer: "Do you offer refunds within 30 days?"
AI: "Yes, we offer a full refund within 60 days!"
[FALSE: the policy is 14 days]

The consequences: you have stated in writing a condition you do not apply, the customer discovers it when they try to use it, and the conversation starts from there.

Solution: RAG (Retrieval-Augmented Generation). The AI searches your content first, then answers by citing it. If it finds nothing, it has to say so rather than invent.

❌ 3. An intrusive proactive chatbot

The problem: the "Can I help you?" window that reopens on every page.

Consequences: it interrupts browsing, irritates the visitor and gets closed unread. At worst it covers the very content they were looking at.

What is needed: show it where it makes sense — on the pricing page, or when the visitor is about to leave — and never automatically everywhere. When to show it and when not to is the subject of the guide to website chat.

The hybrid model: AI and people

Three levels

Tier 1 (the bulk of the volume): repetitive questions that the knowledge base and self-service can resolve without opening a ticket

  • Password resets
  • Order tracking
  • Standard FAQs
  • Basic product information

Tier 2: AI + person

  • Refunds
  • Simple technical problems
  • Plan changes
  • The AI prepares the reply, the person approves or edits it

Tier 3: people only

  • VIP customers
  • Legal/compliance issues
  • Complex bugs
  • Custom contract negotiations

Transparent handoff

❌ Bad:

"I'm sorry, I can't help you. I'll transfer you."
[The customer gets annoyed because they are wasting time]

✅ Good:

"I've gathered the details to speed things up:
- Order #12345
- Problem: parcel not delivered
- Address: Via Roma 10, Milan
I'm connecting you with Sara from the shipping team. One moment, 30 sec."

How to work out whether it pays off

The calculation does not need a complicated spreadsheet. You need four numbers, and they are all yours:

  1. How many tickets you receive per month and how many of those are repetitive questions
  2. How many minutes it takes on average to answer one of those tickets
  3. How much an hour of your team costs
  4. How much the AI costs in the plan you are considering: pay-as-you-go, per agent, or included up to a cap

To estimate the monthly cost of repetitive requests, multiply their number by the average response time in minutes, divide by 60, then multiply by the team’s hourly cost. Measure the time actually saved during the trial, convert it to value using the same hourly cost, and compare it with the cost of AI. Do not use the savings percentages you find in marketing material, ours included: use your own minutes and your own hourly cost.

The right way to start: switch the AI on for a limited share of tickets, measure for two weeks against the previous period, and expand only if the numbers hold up. If they do not, you have lost two weeks instead of a year's subscription.

How costs work in Uppo

Uppo Assist is licensed per enabled operator and its interactive activities do not consume credits. AI Agent uses workspace credits for verified autonomous resolutions. The included allowance and add-ons are on the pricing page.

How to start, over two weeks

Week 1: foundations

  • Days 1-2: measure the baseline (volume, response time, repetitive tickets)
  • Days 3-4: put the knowledge base in order — without good content, no AI works
  • Day 5: switch on automatic classification for a single ticket category, not all of them
  • Days 6-7: check every proposal against what the agents would have done, and correct the categories where it gets things wrong

Week 2: expansion

  • Days 8-10: switch on reply suggestions for agents
  • Days 11-12: train the team on how to use them, and when to ignore them
  • Day 13: extend to the categories where the accuracy has convinced you, not to all of them
  • Day 14: define how you will monitor quality over time

In conclusion: AI and people work together

Myth: "AI will replace agents" Reality: AI takes the repetitive work away from agents, not the job itself. On the requests that really matter — a complaint, a refund, an angry customer — the person remains irreplaceable, and the customer knows it.

In practice:

  1. first level to self-service: the best ticket is the one that never arrives
  2. second level to AI: support for the team or autonomous replies within defined rules
  3. third level to people only: key customers, complex cases, complaints

Before any of that, measure your baseline: without it you will not know whether the AI is helping. The numbers to watch are in the customer service KPIs worth measuring in a small company.

Uppo

Turn the method into customer service that works.

Centralise conversations, assign clear ownership and keep response times and quality under control from one helpdesk.

Start your free trial Talk to the team