AI and Support Teams: Replacement or Transformation?

Customer support is where the AI replacement question stopped being theoretical. Stanford's Digital Economy Lab lists customer service representatives among the occupations most exposed to AI automation, and the two most cited corporate AI stories of the past two years, Salesforce and Klarna, are both support stories. One is held up as proof that AI can shrink a support organisation by thousands of roles. The other became the cautionary tale of the decade when it reversed course and rehired humans. Between those two outcomes sits the model that actually works: a tiered support operation where AI absorbs the repetitive volume and people own the complex, emotional, and high-value conversations. Here is what both experiments teach, the metrics that separate genuine transformation from cost-cutting theatre, and a practical path for adopting the tiered model.

The Replacement Experiments: Salesforce and Klarna

The numbers that launched a thousand board discussions came from two companies. In September 2025, Salesforce CEO Marc Benioff said the company had cut its customer support staff from 9,000 to roughly 5,000 as its Agentforce AI agents took over about half of all customer interactions, telling an audience bluntly, "I need less heads." Support costs fell 17 percent, according to Fortune's reporting.

Klarna moved earlier and harder. In 2024 the Swedish fintech said its AI assistant was doing the work of around 700 customer service agents, a figure later revised to 853 full-time equivalents, and the company shrank from roughly 5,527 employees to about 2,907 through a hiring freeze. CEO Sebastian Siemiatkowski publicly credited AI with a 40 percent workforce reduction.

On paper these look like the same story: AI replaces support staff, costs fall, investors applaud. They are not the same story, and the difference between them is the most useful lesson available to any business leader weighing this decision in the USA, the UK, or anywhere else.

Where Full Replacement Went Wrong

Klarna's reversal is now well documented. Customer satisfaction deteriorated on complex interactions, the kind where someone is disputing a charge, dealing with suspected fraud, or simply upset, and the company began rehiring human customer service agents. Siemiatkowski admitted "We went too far," and Klarna moved to a hybrid human-AI service model.

The pattern extends well beyond one fintech. Forrester's 2026 Future of Work report estimated that 55 percent of employers regretted laying off workers for AI-related reasons, and Forrester predicts half of all AI layoffs will be reversed in some form by the end of 2026. Research by outplacement firm Careerminds found roughly two-thirds of companies that ran AI-led layoffs are already rehiring, with 35.6 percent bringing back more than half of the eliminated roles, and about one in three employers spent more on restaffing than the layoffs saved. We unpack that boomerang effect in detail in our piece on the hidden costs of AI layoffs.

The failure mode is consistent. AI handles repetitive contacts well, and Salesforce's own figure of roughly 50 percent of interactions is a realistic ceiling for a mature deployment. But the remaining cases are precisely the ones that decide whether a customer stays. A frustrated customer who fights through three chatbot loops before reaching a human has already learned that your company would rather deflect them than help them. The savings show up immediately on a spreadsheet; the damage shows up months later in churn, refunds, and public reviews.

The Tiered Model: How AI and Humans Divide the Work

The operations that work in 2026, from SaaS firms in Canada to retailers in Australia, share a three-layer structure rather than a binary choice between bots and people.

Tier 0: instant AI resolution. A well-built chatbot answers the repetitive majority of questions immediately, around the clock, in any language: order status, password resets, billing explanations, plan changes, returns within policy. This is where deflection genuinely helps customers, because the alternative was a queue. A purpose-built system from a specialist AI chatbot development team, grounded in your actual help content and order data, behaves very differently from a generic widget bolted onto your site.

Tier 1: AI agents with bounded authority. The next layer is agentic: AI agents that do not just answer but act, processing a standard refund, rebooking a delivery, applying a credit within preset limits, or updating account details after verification. Every action is logged, capped, and reversible.

Tier 2: human escalation. Complex, emotional, regulated, and high-value cases route to trained people, with the full conversation history and customer context attached so nobody has to repeat themselves. Escalation triggers should include detected frustration, vulnerable-customer signals, legal or compliance keywords, high account value, and any case the AI scores as uncertain.

The handoff design matters more than the model choice. An AI that escalates too eagerly saves no money; one that clings to conversations it cannot resolve burns trust. Getting that threshold right, and revisiting it monthly, is the core engineering work of the tiered model.

The Metrics That Matter (and the Vanity Metric That Lies)

Deflection rate is the number most vendors lead with, and on its own it is a vanity metric. A bot can "deflect" a ticket by frustrating the customer into giving up, and the dashboard will call that success. The metrics that actually describe a healthy tiered operation are different.

Resolution rate, measured by whether the customer's issue was actually solved with no reopen within a defined window. CSAT split by path: AI-resolved, human-resolved, and escalated-then-resolved conversations scored separately, because a blended average hides a failing tier. Escalation accuracy: how often the AI hands off the right cases at the right moment. Reopen and repeat-contact rates, the clearest signals of fake deflection. Cost per resolution rather than cost per ticket. And churn among customers who contacted support, the metric that ultimately justifies or condemns the whole programme.

The augmentation evidence is strong when these metrics are managed honestly. A study of 5,172 support agents at a Fortune 500 software firm by Brynjolfsson, Li and Raymond, published through NBER and the Quarterly Journal of Economics, found agents using a generative AI assistant resolved roughly 14 to 15 percent more issues per hour, with the largest gains going to the least experienced agents. That is the quiet, unglamorous version of AI in support: not headcount theatre, but every agent performing closer to your best agent.

What Transformation Means for the People on Your Team

In a tiered model the human roles change rather than vanish. Agents become escalation specialists who handle only the hard conversations. New roles appear: AI supervisors who review transcripts and correct the system's mistakes, conversation designers who shape how the bot speaks, and quality analysts who audit both human and AI interactions. Harvard Business Review research from March 2026 found that while job postings fell 17 percent in the most automation-exposed roles, augmentation-friendly roles saw a 22 percent increase in demand. The work moves up the value chain for those given the chance to follow it.

The reskilling case is not theoretical. IKEA retrained 8,500 call-centre employees as interior design consultants with no layoffs, generating a reported $1.4 billion revenue uplift, because people who had spent years talking to customers turned out to be very good at advising them. The World Economic Forum reports 77 percent of surveyed companies plan to reskill or upskill existing workers between 2025 and 2030 to work alongside AI. Even IBM, which replaced around 200 HR roles with AI agents, tripled entry-level hiring for 2026, with its CHRO noting the work "still requires a human touch." For the broader evidence on why augmentation beats substitution, see our analysis of AI augmentation versus replacement.

A Practical Adoption Path, and the Regulatory Map

For most businesses the sensible sequence is: map your ticket taxonomy and identify the ten most repetitive intents; run the chatbot in shadow mode against historical tickets before it ever talks to a customer; launch on the safe intents with a visible "talk to a human" path; add agentic actions one at a time with hard limits; and only then revisit staffing, ideally through redeployment and attrition rather than cuts. The financial modelling, including the restaffing trap that catches one in three companies, is covered in our guide to the economics of replacing staff with AI.

Rules differ by market, so design for the strictest one you serve. In Europe, the EU AI Act introduces transparency obligations, meaning customers should be told when they are talking to a machine, and support transcripts fall squarely under GDPR. UK firms in regulated sectors face consumer-protection expectations around fair treatment that make silent bot-only support a genuine compliance risk. Canada's PIPEDA, Australia's privacy framework, and South Africa's POPIA all govern the customer data flowing through AI tools, and several US states already require businesses to disclose when customers are chatting with a bot. None of this blocks the tiered model; all of it punishes the lazy version of it.

The honest summary: replacement as a strategy has a documented failure record, and transformation has a documented success record. AI should be answering your repetitive tickets tonight. It should not be the only thing standing between your angriest customer and the door.

Frequently Asked Questions

Should AI replace my customer support team entirely?

No. The companies that attempted full replacement, most famously Klarna, reversed course after customer satisfaction dropped on complex interactions. The evidence favours a tiered model: AI resolves repetitive tickets instantly while trained human agents handle complex, emotional, and high-value cases. Even Salesforce, which cut support staff from 9,000 to about 5,000, still runs thousands of human agents alongside its AI.

What is the tiered support model?

A three-layer structure: an AI chatbot layer that instantly answers repetitive questions around the clock, an AI agent layer that takes bounded actions such as processing standard refunds or booking changes, and a human escalation layer for complex, emotional, regulated, or high-value cases. The handoff rules between layers matter as much as the AI itself.

How many support tickets can AI realistically handle?

Salesforce reported its Agentforce AI agents handling roughly 50 percent of customer interactions, which is a realistic benchmark for a mature deployment. The exact share depends on how repetitive your ticket mix is: order status, password resets, and billing questions automate well, while disputes, fraud, and complaints do not.

What went wrong when Klarna replaced support agents with AI?

Klarna's AI assistant was credited with the work of roughly 700 customer service agents, but customer satisfaction deteriorated on complex interactions. CEO Sebastian Siemiatkowski admitted the company went too far, and Klarna began rehiring human agents and moved to a hybrid human-AI service model.

Which metrics should I track for AI customer support?

Resolution rate rather than deflection rate, customer satisfaction split by AI-resolved versus human-resolved conversations, escalation accuracy, reopen rate, cost per resolution, and churn among customers who contacted support. Deflection rate alone is a vanity metric because a bot can deflect customers by frustrating them into giving up.

What happens to human support agents under the transformation model?

They move up the value chain into escalation specialist, AI supervisor, conversation designer, and quality analyst roles. IKEA reskilled 8,500 call-centre employees into interior design consultants with no layoffs, generating a reported $1.4 billion revenue uplift, and the World Economic Forum reports 77 percent of companies plan to reskill workers to work alongside AI between 2025 and 2030.