ai customer service automation in 2026: what the record actually shows
2026-08-04
Vendors promise deflection and speed. Three deployments TIN checked against the public record show where the numbers hold, where they moved, and where a company walked the whole thing back.
Built on verified case files. The argument below leans on evidence The Internet Ninja validated against the public record and published in full, method included.
Every AI vendor selling to a support team leads with the same two numbers: deflection up, resolution time down. The pitch is easy to make and hard to check, because the figures almost always originate with the company running the pilot.
So the useful question is not whether ai customer service automation works. It is which numbers survive a look at the public record, and what the honest ones have in common.
TIN checked three deployments against their sources. In one, an AI agent doubled deflection at a named client. In one, AI drafted a third of a utility’s support email with a person sending every reply and no job cuts. In the third, a fintech went AI-first, then reversed course: its CEO said the cost-led push produced “lower quality,” and the company began re-recruiting humans (source).
What ai customer service automation means
ai customer service automation is software that resolves or drafts responses to customer inquiries with little or no human input, deflecting repetitive tickets and routing the rest to a person. The load-bearing word is “routing”: the deployments that report durable numbers keep a human in the loop for the cases the machine should not close.
Can ai replace customer support?
Not on the current record. The clearest attempt to replace it, rather than assist it, is the one that got reversed.
Klarna’s OpenAI-built assistant handled two-thirds of its customer-service chats, 2.3 million conversations, in its first month in 2024, and was said to do the equivalent work of 700 full-time agents (source). That 700 figure is a modeled workload equivalence Klarna stated, not 700 documented layoffs, and TIN’s case file on Klarna carries it as exactly that.
By 2025 the story changed. CEO Sebastian Siemiatkowski said cost had been “a too predominant evaluation factor” and that “really investing in the quality of the human support is the way of the future for us” (source). The company that reported the most aggressive automation numbers is the one that publicly walked them back.
Does ai customer support automation reduce resolution time?
Yes, where there are figures, but they come from the deploying companies, not third-party audits.
Notion deployed Decagon’s AI customer-experience agent and reported a 34% improvement in ticket resolution time, a 2x increase in deflection, and a 3.4% ask-for-human rate (source). TIN’s case file on the Notion deployment treats those as vendor-published figures anchored to a named, on-record client executive who could refute them, not as independently measured results.
Klarna reported customers resolving issues in “less than 2 mins compared to 11 mins previously” and a 25% drop in repeat inquiries (source). Real speed gains, self-reported, and later overshadowed by the quality reversal.
What does keeping a human in the loop actually look like?
It looks like Octopus Energy. Its Kraken “Magic Ink” tool drafted customer-service email replies, with a human agent reviewing and sending each one.
By the end of April 2023, CEO Greg Jackson said it answered 34% of all customer queries, “the work of 250 people in the UK alone” (source). He framed that as workload equivalence, not layoffs, called cuts “unlikely,” and the company announced 4,000 new jobs in 2024. TIN’s case file on Octopus records the satisfaction figure moving across statements (80/65, around 70, 65/55) and shows each rather than merging them.
The satisfaction claim is unstable across sources. The structural choice, a person sends every reply, is not.
The proof
These three cases are the ones TIN validated against the public record, and the pattern across them is the argument. Every figure below originates with the deploying company; none is a third-party audit, and the case files say so on their face.
| Deployment | Headline number | Human in the loop | What happened next |
|---|---|---|---|
| Notion + Decagon | 2x deflection, 34% faster resolution | 3.4% ask-for-human rate | Reported stable; single vendor source |
| Octopus + Kraken | 34% of queries drafted by AI | Person reviews and sends every reply | 4,000 new hires announced in 2024 |
| Klarna | Two-thirds of chats, month one | Went AI-first, minimal human fallback | Reversed in 2025; re-recruiting humans |
The deployments that kept a person on the hard cases did not have to walk anything back. The one that treated automation as a headcount lever did.
The bottom line
ai customer service automation deflects repetitive volume and cuts resolution time on the figures companies publish about themselves. It has not replaced customer support anywhere the record can confirm, and the loudest replacement claim is the one its own CEO reversed. Buy it to handle the repetitive two-thirds and route the rest to a human. Treat any vendor number you cannot trace to a source, or that has no person behind the hard cases, as a pitch, not a result.
Sources
- Decagon, “Notion” (case study), checked 2026-07-08. https://decagon.ai/case-studies/notion
- Sifted, “Octopus Energy CEO on AI answering customer emails,” 2023-05-09. https://sifted.eu/articles/octopus-energy-ceo-chatgpt-news
- City AM, “AI doing the work of over 200 people at Octopus, chief executive says,” 2023-05-08. https://www.cityam.com/ai-doing-the-work-of-over-200-people-at-octopus-chief-executive-says/
- Klarna, “Klarna AI assistant handles two-thirds of customer service chats in its first month,” 2024-02-27. https://www.klarna.com/international/press/klarna-ai-assistant-handles-two-thirds-of-customer-service-chats-in-its-first-month/
- CX Dive, “Klarna changes its AI tune and again recruits humans for customer service,” 2025-05-14. https://www.customerexperiencedive.com/news/klarna-reinvests-human-talent-ai-customer-service-buy-now-pay-later/747586/
Questions
Can ai replace customer support?
Not cleanly. The deployments with durable numbers kept a human reviewing or handling the hard cases. Klarna went AI-first, said the result was lower quality, and started re-recruiting human agents in 2025.
Does ai customer service automation actually reduce resolution time?
Yes, in the cases with public figures. Notion reported a 34% improvement in ticket resolution time with Decagon's agent, and Klarna reported dropping average resolution from 11 minutes to under 2. Both figures originate with the companies deploying the tools.
What is a realistic deflection rate for ai customer support automation?
Notion reported a 2x increase in deflection with a 3.4% ask-for-human rate. That is a doubling, not full automation: roughly one in 29 conversations still routed to a person.
Sources
- Decagon (case study), Notion , 2026-07-08
- Sifted, Octopus Energy CEO on AI answering customer emails , 2023-05-09
- City AM, AI doing the work of over 200 people at Octopus, chief executive says , 2023-05-08
- Klarna, Klarna AI assistant handles two-thirds of customer service chats in its first month , 2024-02-27
- CX Dive, Klarna changes its AI tune and again recruits humans for customer service , 2025-05-14
This is analysis, not a verified outcome. It carries no verification badge and never will. The proof lives in the case files, where every figure is checked against the public record and the method is printed on the page.