The Support Team was the AI

How companies are eliminating the people who built the knowledge they're now automating.

When you look at my career, one pattern holds across every role I have ever taken, I always started in support first. That was not an accident. I always felt that starting in a support role provided more information about a company than any training ever could. You learn how customers actually refer to features, not the product names that appear in the documentation, but the words real people use when something breaks or confuses them. You learn how they interact with the software, what they feel, and which enhancements nobody thought to ask for because nobody was listening closely enough.

Support also gave me something harder to quantify: access. I was talking to a CSM one day, an account manager the next, an executive the day after that. Not because I had earned a seat at the table, but because support was where everyone's problems arrived. I got to have meaningful conversations with people in every part of the company, about their roles, about what they did, and about whether I could see myself doing it. That is how I learned what I wanted to do next. And it is how I learned what the company actually was, beneath its own marketing.

So when I started reading about what was happening to support teams, not just at Klarna, but across the industry, I could not let it go.

The Promise

Through June 2026, AI has been cited in 101,743 U.S. job cuts, nearly double the 54,836 attributed to it in all of 2025. About 59% of that all-time total came in the first half of 2026. Customer service and support is one of the primary categories, alongside logistics and tech.

The logic behind every one of those decisions sounds reasonable on the surface. AI can handle routine queries faster and cheaper than humans. Customers say they prefer not waiting on hold. The technology is improving. The savings are real.

Klarna made this case more publicly than anyone else. In February 2024, the company announced that its AI assistant, built with OpenAI, was handling two-thirds of all customer service chats. The official press release claimed the bot was "on par with human agents in regard to customer satisfaction score" and had driven "a 25% drop in repeat inquiries." Resolution time fell from eleven minutes to under two. The company projected $40 million in profit improvement for 2024. Wall Street responded warmly.

What the press release did not include: the underlying survey methodology. The sample size. The breakdown by issue type.

What Actually Happened

By early 2025, the picture had changed. Customer satisfaction scores fell 22%. Klarna posted a $99 million net loss in Q1 2025. CEO Sebastian Siemiatkowski publicly admitted that the pursuit of efficiency had gutted service quality, producing outputs he described as "generic."

The reversal was not announced the way the original deployment had been. The company quietly began rebuilding its human customer service capacity, rehiring agents, testing an "Uber-style" model with flexible remote workers.

"From a brand perspective, a company perspective, I just think it's so critical that you are clear to your customer that there will always be a human if you want."

— Sebastian Siemiatkowski, CEO of Klarna, Bloomberg 2025

That is a striking statement from the CEO who eighteen months earlier had used his own company as the headline argument for replacing 700 people with AI.

What happened in between is a data story most executives were not watching. The headline metrics looked good throughout 2024: overall chat volume handled, time-to-first-response, resolution rate, all of these stayed strong or improved. What deteriorated silently were two other signals: direct satisfaction scores on complex interaction types, and repeat contact rate the number of customers who had to contact support multiple times for the same issue.

A customer who contacts support twice for a refund dispute that should have been resolved on the first call is not a resolved case. They are a brand problem. And that brand problem does not show up in the metrics most executives were reviewing.

Curiosity

To be a good support agent, you need curiosity. Even when presented with an initial answer, a good agent keeps digging. They stay on the call a little longer, ask the question the customer did not think to ask, trace the problem back one more step. That is where the true answer lives, not in the first response, but in the second question.

AI does not do this. It produces the statistically likely answer and stops. For issues involving professional knowledge, requiring specific analysis, or needing personalized judgment, intelligent customer service cannot provide sufficient solutions. When users cannot express their demands clearly, when they are upset, or speaking a dialect, or circling around a problem they do not have words for, human customer service can understand and pinpoint what is actually wrong. AI processes what was said. A skilled agent hears what was meant.

CSAT dropped on complex and emotional tickets even when AI gave correct answers. The AI can be technically right and still fail the interaction. That failure does not show up in aggregate CSAT. It shows up in churn.

When AI Makes Things Up

The consequences of removing human judgment are not just statistical. They are legal.

Air Canada's AI chatbot invented a bereavement fare policy that did not exist, telling a passenger who had just lost his grandmother that he could receive a retroactive discount. When Air Canada was taken to the British Columbia Civil Resolution Tribunal, the airline argued, in an unprecedented move, that the chatbot was a separate legal entity responsible for its own actions. The tribunal rejected this entirely and held Air Canada liable.
Cursor's AI support bot told a developer that a new policy limited each subscription to one device. The policy did not exist. Before the correction reached users, the invented policy had circulated on Hacker News and Reddit.
WestJet's chatbot, Juliet, incorrectly interpreted a happy customer's glowing comment about a cabin crewmember's care over her succulent cutting and referred that customer to a suicide hotline.

Two court rulings, Air Canada in 2024 and a German appeals court in May 2026, have now settled the question on two continents: a business owns what its support chatbot says, even when the AI makes it up. The legal precedent has caught up to the operational reality faster than most executives anticipated.

The Data Behind the Decisions

According to a 2025 Assembled survey, 40% of support professionals say scaling support was their main focus, yet 73% flagged customer resistance to AI interactions as a major obstacle. That resistance is not irrational. Gartner research reveals that 64% of customers would prefer that companies did not use AI for customer service at all.

The paradox runs deeper. In 2024, 82% of consumers said they would use a chatbot instead of waiting for a customer representative. An outstanding 96% of surveyed shoppers believed more companies should opt for chatbots over traditional support. Customers say they want chatbots and then report lower satisfaction when companies deploy them. That gap between stated preference and experienced reality is the most important sentence in this entire conversation. People prefer not to wait. They do not prefer being misunderstood by something that cannot tell the difference between a routine question and a crisis.

AI-handled tickets average 4.10 out of 5 in customer satisfaction versus 4.30 for human agents, a 0.20-point gap.

With a hybrid escalation model, that gap narrows to 0.05 points.
The difference between full AI replacement and human-AI collaboration is a 0.15-point improvement in satisfaction.
That 0.15 points is the difference between a strategy that works and one that gets reversed.

Gartner finds AI deflects 45% or more of queries, while only 14% of issues reach full self-service resolution. Executives are measuring deflection. They should be measuring resolution. Those are not the same metric.

Table: Company Comparison — Action and Outcome

Company	Action	Outcome
Klarna	Replaced 700 agents with AI	CSAT fell 22%. Q1 2025 net loss $99M. Reversed — now rehiring.
Salesforce	Cut support from 9,000 to 5,000 citing Agentforce	Credited AI for case decline. Own CSAT data not published.
Paycom	Cut 500 customer service and implementation roles	Profitable when they cut. WARN Act investigation.
Uber	Cut 10% of Community Operations	Employees with 45+ combined years let go.
Dukaan	Replaced 90% of support team	Costs fell 85%. Massive public backlash.
Air Canada	AI chatbot invented a bereavement fare policy	Held legally liable. Two-continent court precedent.
Bank of America	AI assists 18,000 human agents	86% satisfaction rate. Record engagement.
IKEA	AI handles routine queries. Humans retrained as advisors.	Higher satisfaction AND higher retention.
JPMorgan	COIN automates contracts. Staff redeployed.	360,000 hours saved. No mass layoffs.
Thrasio	AI equips agents with brand-specific replies instantly.	CSAT rose from 87% to 97%. $1.8M saved annually.

Who Got It Right

The companies that succeeded did not ask whether to use AI. They asked which problems AI should solve and which ones it should not.

IKEA deployed an AI chatbot to handle routine customer inquiries, order tracking, return policies, store hours and retrained its call center workers as interior design advisors. The results showed both higher customer satisfaction and higher employee retention. The people who knew the customers did not disappear. They moved up.
JPMorgan's COIN platform reviews commercial loan agreements in seconds, work that previously consumed 360,000 hours of lawyers' time annually. Rather than mass layoffs, the bank redeployed legal staff toward higher-complexity negotiations and regulatory work. The lawyers did not become redundant. They became more valuable.
Bank of America's approach is the clearest model at scale. More than 18,000 employees use EricaAssist as a human-assisted AI agent. Client satisfaction is at an all-time high, with 86% of clients rating their digital experience at least a 9 out of 10. Twenty million users interacted with Erica nearly 700 million times last year.
Thrasio, an e-commerce company managing 190 unique brands and more than 70,000 customer interactions per month, implemented AI in a way that preserved the human layer. They pushed 420,000 customer tickets through machine learning to process six months of interactions, not to replace agents, but to equip them. AI now provides agents with brand-specific replies and relevant knowledge base articles instantly. The results: customer satisfaction rose from 87% to 97%, response times dropped to one minute, full resolution fell to twelve minutes, and the company saved $1.8 million annually. The agents stayed. They just got better tools.
Intercom, which builds its own AI customer service product called Fin, recognized that even companies using strong AI tools still need human agents for complex cases. Their approach: let AI manage simpler queries, let humans handle everything that requires judgment, and invest in workforce management infrastructure that makes the hybrid model actually work.

The distinction I want to make clear: these companies automated tasks. Every company that failed automated roles and eliminated the humans who held the institutional knowledge to handle what AI cannot. The ones who automated tasks kept their institutional knowledge and their customers. The ones who automated roles lost both.

The Pipeline Nobody Counted

Here is the part of this story that nobody is talking about yet. McKinsey explicitly identifies "identifying internal talent in other functions that are likely to produce product managers" as a primary sourcing channel and customer-facing roles are the specific function named. The support team is not just a cost center. It is a talent pipeline.

The pattern is documented in first-person accounts from people who made the transition. - One PM described it this way:

"The knowledge that you have acquired from working in customer support and listening to the ways in which actual people react to the product allows you to contribute meaningfully to the Product team in a way others without your experience cannot. As the customer support team is the first line of contact from the customer to the company, reports from their work are the kind of things that help make a product into something users can actually use and enjoy."

Tiwatayo Kunle, September 21, 2022 - Another, who went from customer support at Capgemini through sales at Freshworks to product manager at a SaaS startup, described the trajectory this way: "I started selling when I was in support, started speaking at conferences when I was in sales and started building products when I was in marketing. The transition was natural."

by Mukhil Maran, March 2021

I lived this pattern. The conversations I had in support, with CSMs, account managers, product teams, executives, are the reason I understand how organizations actually work. Not the org chart version. The real version.

A Wharton management study found that external hires receive significantly lower performance evaluations for their first two years on the job than internal workers promoted into similar roles. External hires also have higher exit rates and are paid 18% to 20% more. Internal promotions offer 18-20% cost savings over external hiring. Organizations that favor internal promotions are 32% more likely to be satisfied with those hires.

The calculation companies are running, cut support, save on headcount does not account for the hidden cost of losing the internal talent pipeline and replacing it with external hires who perform worse and cost more. The BLS projects customer service representative employment to decline 4% from 2024 to 2034. That is the government's forecast for the role that has historically been the entry point for the support → product specialist → implementation → product manager → leadership pipeline.

The Research Gap

When I researched this article, I found a gap I was not expecting. There are no published studies that directly measure what happens to companies that eliminated their support teams and subsequently struggled to fill product roles. The research shows each piece separately: That support is the pipeline. That internal hires outperform external ones. That institutional knowledge is lost when support is eliminated.

But no researcher has yet connected those three pieces into a longitudinal study of what happens to product quality in the two to three years after a company eliminates its support function.

The Klarna story shows the support quality failure, the declining CSAT, the repeated contacts, the reversal. But nobody has yet documented the downstream product failure that follows. The features built without customer insight. The roadmap decisions made without the edge case knowledge that only support teams carry. The products that become less useful not because the engineers got worse, but because the organization lost its connection to the people using what they built.

That is the story I am trying to tell.

What Most Companies Are Not Publishing

Salesforce, Uber, Paycom, Citigroup, none have published CSAT data tied specifically to their AI-cited support cuts. This is worth naming directly.

Salesforce cut its support team from 9,000 to 5,000 people and then cut another approximately 1,000 more, crediting Agentforce efficiency. Salesforce is also the company that sells the tools other companies use to measure customer satisfaction. They literally define CSAT in their own marketing materials. They have not published their own post-cut CSAT data.

Executives want revenue metrics, not satisfaction scores. CFOs and CEOs are now asking customer success leaders how they are impacting net revenue retention and gross revenue retention and explicitly do not want to hear what they call "vanity metrics, like NPS scores."

The industry-wide picture: 83% of U.S. consumers say they are "happy" with service overall, yet 42% report more negative experiences than in prior years, up from 38% in 2024 and 40% in 2025.
Aggregate satisfaction masks touchpoint-level deterioration. Forrester reported that U.S. customer experience quality declined for a third consecutive year in 2024, with 39% of brands experiencing a significant decrease. The cross-industry average CSAT declined from 79 in 2024 to 78 in 2026.

The dashboard shows the average. It does not show the distribution. And the distribution is where the customers you are losing live.

The Questions Worth Asking

Before your next announcement about AI-driven support efficiency, ask these:

Who trained the AI, and what happens when those people are gone? The AI knows what it knows because support agents built the knowledge base, handled the edge cases, and generated the training data. When you eliminate the source, you eventually eliminate the knowledge.
What is your repeat contact rate, not your resolution rate? Resolution rate tells you whether the AI closed the ticket. Repeat contact rate tells you whether it actually solved the problem. These are not the same metric. Klarna looked healthy on one and was failing on the other.
When your AI escalates to a human, who is that human, and what do they know? If you have cut your support team, the cases the AI cannot handle now have nowhere to go. The hard cases are precisely the ones that require the most experienced people and those are often the first to leave when the team shrinks.
Who on your product team has spent time in customer support? If the answer is nobody, your product team is building without a direct line to the people using what they build. That gap compounds over time.
How is your AI gathering the data you need to not just change your product but innovate? AI can report what customers asked. It cannot tell you what customers meant, what they needed and did not know how to ask for, or what they gave up trying to do. That signal lives in the support interaction specifically in the moments a skilled agent noticed something the customer had not even articulated yet.
Who is training your team for a partnership with AI? Deploying AI without training your human agents to work alongside it is not a cost-saving measure. It is a setup for the outcome Klarna experienced. The hybrid model requires human skill, not just human presence.
How are you ensuring that your data is safe and compliant? Two courts have now ruled that companies are liable for what their AI systems say. Before your chatbot handles customer disputes, account closures, or any interaction involving personal data, the questions of SOC 2, HIPAA, and PCI-DSS compliance are not technical checkboxes. They are legal exposure. The Air Canada ruling did not turn on the sophistication of the AI. It turned on the fact that a customer relied on what the company's system said.
Are you documenting knowledge gaps in a meaningful way? A support team that documents what it does not know, the edge cases, the unusual configurations, the questions that arrive without good answers, is actively building organizational intelligence. An AI system that handles those same cases silently is burying them.

The companies that will get this right are not the ones that find the most efficient way to remove humans from the equation. They are the ones that understand what humans bring to the equation that cannot be automated: the curiosity to keep asking, the experience to recognize when a routine query is actually something else, and the institutional knowledge to know what "something else" usually means.

That knowledge was built by support teams. It lives in support teams. And when you eliminate support teams, you do not just cut a cost center. You cut the source.

Every role I built started with support. Every product I have understood, every team I have navigated, every insight I have brought to a room, it traces back to those first conversations with customers who were frustrated, confused, or just needed someone to actually listen. The companies that will get this right are the ones that understand what humans bring that cannot be automated: the curiosity to keep asking, the experience to recognize when a routine query is actually something else, and the institutional knowledge to know what that something else usually means.

That is not a soft skill. That is the work. And it cannot be automated. The data on what happens next to the products, the roadmaps, the companies that cut their way to efficiency, has not been written yet. The support teams are gone. The products are still being built. In two or three years, we will know what was lost.

I am already watching.

Triplebee|Human-in-Command