All case studies

Case studies: customer support

Let AI resolve the routine 60-70% of tickets.

Your team takes the rest, with full context. We build an AI support agent that classifies each ticket, resolves the routine ones from your help docs and order data, and hands complex or upset customers to a person with the history attached. Klarna's AI cut average resolution time from 11 minutes to under 2. Then Klarna learned that full automation erodes CSAT, which is why we build hybrid.
Book a 30-min workflow reviewGet a free 90-day AI plan

Who handles what in a stable hybrid

  • AI resolves60–70%
  • Your team30–40%
Industry deployment data 2024-2025, and the Klarna case.

  • 11 → 2 min

    Average resolution time

    Klarna

  • 98%

    Chats resolved without a human

    Ruby Labs

  • −60%

    Human support need

    NIB

  • 6-12 wks

    Typical first deployment

    ATI Lab delivery

Case studies

Who got AI support right, and who had to walk it back

Klarna's full arc (savings, then rehiring), plus teams that built the hybrid from day one.

Third-party results from other companies, linked to source. They are not ATI Lab clients. How we would build it for you is below. See ATI Lab's own client builds.

Klarna - The most studied AI support deployment

Fintech - GlobalEnterprise referenceThird-party result

The bottleneck

150M users generating massive support volume. Customer service costs at $0.32 per transaction. Manual handling could not scale with growth.

What fixed it

OpenAI-powered AI assistant deployed globally. Handled 2.3 million conversations in its first month - equivalent to 700 full-time agents. Operates in 35+ languages, 24/7. Cost: $2-3M implementation cost (per CEO Sebastian Siemiatkowski).

Published result

  • $60M

    total savings as of Q3 2025

  • $39M

    saved in 2024 alone (IPO filing)

  • −25%(decrease)

    repeat issues

  • 2/3

    of all inquiries handled by AI

The Klarna arc

  1. Feb 24: Launch

    2.3M conversations in month one, the work of about 700 full-time agents.

  2. 2024: $39M saved

    Reported in the IPO filing.

  3. May 25: Rehiring for complex cases

    Full automation had cost customer satisfaction.

    CSAT pressure → hybrid

  4. Q3 25: $60M total savings

Average resolution time

  • Resolution time82% faster
    Before11 min
    After2 min
Third-party published result.

Customer-service cost per transaction (over 2 years)

  • Cost per transaction40% lower
    Before$0.32
    After$0.19
Third-party published result.

Sources: OpenAI case study (opens source in a new tab), CX Dive / Yahoo Finance (opens source in a new tab), CX Dive (opens source in a new tab), The Finanser (opens source in a new tab), Klarna press release (opens source in a new tab) - Q3 2025 earnings, Q1 2025 earnings, IPO filing analysis, and Klarna press release (Feb 2024).

The Klarna lesson - why hybrid wins

Critical contextThird-party result

The bottleneck

Klarna initially went all-in on AI, laying off staff and pausing hiring. By May 2025, they reversed course and began rehiring humans for complex cases after overpivoting on pure cost containment.

What fixed it

Hybrid models (60-70% AI, 30-40% human) typically deliver the strongest mix of cost savings and customer satisfaction. Full automation can produce short-term savings followed by CSAT erosion.

Published result

  • 60-70%

    AI handling in a stable hybrid model

  • 30-40%

    human handling for complex cases

  • Hybrid

    best long-term balance of cost and quality

Source: CX Dive (opens source in a new tab) - Klarna staffing and model shift in 2025 with analyst commentary on complexity-driven handoffs.

Ruby Labs - AI-First Support Resolution

SaaS - USAThird-party result

The bottleneck

Support team could not scale with user growth. Every new user added linear support cost. No way to maintain quality at scale.

What fixed it

AI agent on the front line: it classifies each ticket, writes and sends the answer for routine cases, and follows up. Only exceptions reach a person.

Published result

  • 98%

    of support chats resolved without human intervention

  • Flat headcount

    support without proportional headcount growth

Support chats at Ruby Labs

resolved without a human

Third-party published result.

Source: Botpress (opens source in a new tab) - "10 Real World Case Studies of AI Agents" (May 2025)

NIB Health Insurance (Australia) - "Nibby" AI Assistant

Health insurance - AustraliaThird-party result

The bottleneck

High volume of routine customer inquiries consuming licensed agent time. Agents tied up answering repetitive questions instead of handling complex cases.

What fixed it

AI digital assistant "Nibby" for routine inquiries, with a handoff to a human agent that carries the conversation over, so the customer doesn't repeat themselves. A hybrid model from the start.

Published result

  • $22M

    total savings

  • −60%(decrease)

    human support need

  • −15%(decrease)

    phone calls to agents

Indexed (before = 100)

  • Human support need60% less
    BeforeBefore: 100
    AfterAfter: 40
  • Phone calls to agents15% fewer
    BeforeBefore: 100
    AfterAfter: 85
Third-party published result.

Source: FactR / DataPeak (opens source in a new tab) - industry comparison in AI cost savings analysis

American Express - Travel Counselor Assist AI

Financial services - GlobalEnterprise referenceThird-party result

The bottleneck

5,000 travel counselors across 19 markets creating personalized travel plans. Enormous research burden per customer interaction.

What fixed it

AI assistant working alongside human counselors - pulling real-time travel data, matching suggestions to cardholder preferences using booking and spending behavior, handling research instantly.

Published result

  • 85%+

    of counselors say AI saves time and improves quality

  • Real-time

    location-specific travel data

  • Personalised

    suggestions from spending behavior analysis

Source: Botpress (opens source in a new tab) - "10 Real World Case Studies of AI Agents" (May 2025)

Waiver Consulting Group - Waiverlyn AI Lead Bot

Pre-sales chatThird-party result

The bottleneck

Website visitors dropping off without converting. No scalable way to greet, qualify, and book consultations without dedicated sales staff.

What fixed it

AI lead generation bot ("Waiverlyn") deployed to greet visitors, answer questions, collect contact info, book consultations, fill calendar invites, send reminders, update Google Sheets, and notify the sales team.

Published result

  • 3 weeks

    to pay for itself in booked consultations

  • Fully automated

    lead capture → qualification → booking → notification

Pre-sales questions rather than support tickets; shown because the handoff pattern is the same.

Source: Botpress (opens source in a new tab) - "10 Real World Case Studies of AI Agents" (May 2025)

How we'd build it for you

What we build for support teams

  • The pain: The same 20 questions, all day

    What we build: AI agent trained on your help center + order lookup

    Published proof: 98% of chats auto-resolved (Ruby Labs)

  • The pain: Bots trap customers in loops

    What we build: Scenario classifier + handoff rules + context passed to the agent

    Published proof: 60-70% AI / 30-40% human, stable hybrid

  • The pain: Slow answers to simple questions

    What we build: Instant AI first reply, 24/7, multilingual

    Published proof: 11 → 2 min resolution time (Klarna)

How each ticket is routed

  1. Ticket in

    Chat, email, form

  2. Classify scenario

    Guardrails: unsure → escalate

    Branch: Complex or upset → your team, with the full thread

  3. AI resolves

    Help docs + order data

  4. Follow-up sent

Industry benchmarks

Realistic targets for an AI + human support desk

Ranges reported for AI support

  • Efficiency gains in customer serviceHype Studio / industry benchmarks40–70%
  • AI share of volume in a stable hybridIndustry deployment data 2024-202560–70%
Third-party published result.

Typical call-center time savings

  • Time saved30%
Source: Agility at Scale, citing Deloitte and McKinsey. Third-party published result.

  • 40-70%

    efficiency gains in customer service from AI

    Hype Studio / industry benchmarks

  • $80B

    projected global contact center savings from AI by 2026

    Digital Applied analysis

  • 60-70%

    optimal AI handling rate for hybrid models

    Industry deployment data 2024-2025

  • 30% call center

    time savings is the typical industry benchmark

    Agility at Scale / Deloitte, McKinsey

Next step

We build the hybrid, not the 'replace everyone' bot.

We classify your ticket types, decide which ones AI can resolve safely, connect it to your helpdesk and order data, and route everything else to your team with the full thread. We agree a target share for AI (60-70% is typical in stable hybrids) and how it is measured against CSAT. Typical first deployment: 6-12 weeks.

Review my ticket mix