By Jonathan West · Updated Nov 11, 2026

How Businesses Use AI in Customer Service Today (2026)

What AI does in customer service and how strong the evidence is for each use.

By Jonathan West · Updated Nov 11, 2026

In the sources we reviewed, businesses use Artificial Intelligence (AI) in customer service for three jobs: answering routine questions, helping human reps write replies, and scoring conversations. At Layer3Labs, we build custom AI agents and integrations for small and mid-sized businesses, including customer-service automation.

We compared the study, regulator report and vendor documentation below to see which AI uses have academic evidence.


How Businesses Use AI in Customer Service

Agent assist has one academic study. The other two jobs rest on company reports or vendor descriptions.

The only result measured by academic researchers among the sources we reviewed applies to agent assist rather than autonomous bots. In National Bureau of Economic Research (NBER) Working Paper 31161 by Brynjolfsson, Li and Raymond, researchers studied a generative AI assistant used by 5,179 customer support agents, later published in 2025. Access to the assistant increased productivity by 14% on average as measured by issues resolved per hour. That gain reached 34% for novice and low-skilled workers, but showed minimal impact on experienced and highly skilled workers.

By contrast, results for fully automated self-service support in the sources we reviewed are company-reported and depend on escalation and accuracy controls. Those sources describe quality problems when complexity or cost pressure is not managed.

If your business handles mostly complex, regulated, or low-volume contacts, adding AI first is not the right starting point for you. Keep human support reps, or use a hybrid model with a customer service escalation workflow template to route hard cases to people.

  • Self-service AI agents: customer-facing AI that answers routine questions.
  • Agent assist: copilot tools that suggest next steps, draft responses, and execute approved actions for human reps based on internal knowledge and procedures, as described on the Zendesk Copilot page.
  • Quality scoring: automatic scoring of conversations for both human reps and AI agents, as outlined on the Zendesk AI agents page.
Layer3Labs Implementation & Advisory

Adding AI to customer service is one workflow decision among many. Layer3Labs can review where AI fits the support workflows your team already runs.

2-Wk AI Architecture Sprint: $2,500 fixedTurnkey Custom Agent Build: From $5,000Fractional Chief AI Officer: From $3,500/mo
Book a Consultation

Examples of AI in Customer Service by Job

The table lists each job, who reports its results and how strong that evidence is.

AI jobWhat it doesSource and who reports itEvidence strengthMain limit
Self-service AI agentAnswers routine customer questionsKlarna press release and Zendesk AI agents pageCompany-reported figures and vendor marketing claim, with no measured study among the sources we reviewedConsumer Financial Protection Bureau (CFPB) reports that in consumer finance effectiveness wanes as problems become more complex
Agent assist (copilot)Drafts responses and suggests next steps for human repsNBER Working Paper 31161 and Zendesk Copilot pageOne academic study of 5,179 agents at one company with one assistant, plus vendor feature descriptionMinimal impact for experienced and highly skilled workers
Quality scoringScores conversations automatically, for both human reps and AI agentsZendesk AI agents pageVendor feature description only, with no measured study among the sources we reviewedNo measured study among the sources we reviewed

Only one row rests on a measured study. The other two rest on company reports or vendor descriptions.


Benefits the Evidence Supports

Published research and corporate announcements report different gains, and the evidence behind them varies by use. The only measured result among the sources we reviewed is for helping human reps, not replacing them.

In the 2025 publication of NBER Working Paper 31161, researchers observed a 14% average lift in issues resolved per hour across 5,179 customer support agents using one generative AI assistant. The productivity increase was concentrated among novice and low-skilled workers, who saw a 34% improvement in issues resolved per hour. For experienced and highly skilled workers, the assistant produced minimal impact. The study abstract also reported that the AI assistant improved customer sentiment and increased employee retention. It noted that AI assistance may lead to worker learning within that single company setting.

For customer-facing self-service, the available data in our review relies on corporate announcements. A February 2024 press release from Klarna reported that its AI assistant handled 2.3 million conversations in its first month, representing two-thirds of the firm's customer service chats. Klarna also reported a 25% drop in repeat inquiries. It stated average resolution time fell to less than 2 minutes compared to 11 minutes previously. Those numbers reflect company-reported operational figures rather than an outside audit.

None of the sources we reviewed measured whether self-service AI handles complex questions well.


Limits and Failure Cases to Plan For

The sources we reviewed describe four limits of AI in customer service. These come from a CFPB report, Klarna's chief executive as relayed by Customer Experience Dive (CX Dive), a British Columbia tribunal decision, and a Zendesk marketing claim.

In a June 6, 2023 research report on chatbots in consumer finance, the CFPB concluded that automated tools may resolve basic inquiries. Their effectiveness wanes as problems become more complex. The CFPB estimated that approximately 37% of the United States (US) population interacted with a bank chatbot in 2022. For financial institutions and related organizations, our guide on AI customer service in banking covers bank-specific capabilities, CFPB risks and regulation.

CX Dive reported on May 9, 2025 that Klarna's chief executive said the following. "As cost unfortunately seems to have been a too predominant evaluation factor when organizing this, what you end up having is lower quality."

Moffatt v. Air Canada is a single decision of the British Columbia Civil Resolution Tribunal, a small-claims tribunal, dated February 14, 2024 (2024 BCCRT 149). We know it only through two law-firm summaries, one from Torkin Manes and one from Deeth Williams Wall (DWW), not through the decision text. According to the DWW summary, the tribunal found Air Canada did not take reasonable care to ensure its chatbot was accurate. It awarded a total of $812.02 in damages, pre-judgment interest, and fees. The decision is not settled law for other businesses.

The Zendesk AI agents page says its product can "Resolve even the most complex issues on any channel autonomously," which is a vendor marketing claim and not a measured result. Three practices follow from our reading of these sources:

  • A human handoff path: establish routing that transfers complex or sensitive issues to human reps.
  • Accuracy review: check what the AI says before and after launch.
  • Balanced evaluation metrics: measure resolution quality alongside cost.

How Vendors Price AI Support

Intercom prices Fin, its AI agent, per outcome, so what counts as an outcome directly affects your bill.

On its public pricing page as of 11 November 2026, Intercom lists Fin at $0.99 per outcome. Intercom also notes that a minimum monthly commitment applies, citing for example 50 outcomes.

A per-outcome price and a per-seat price count different things, so you need your own monthly count of resolved questions before comparing them.

Intercom counts an outcome when any one of these three things happens, and because handoffs are in the third condition, a handoff can be billed:

  • A customer confirms their issue is resolved.
  • They do not ask for more help after Fin responds.
  • Fin completes a workflow (Procedure), including handoffs.

Where to Start with AI in Customer Service

Start with agent assist where many reps are new or low-skilled.

Your situationFirst AI job to tryWhat would change that
Many new or low-skilled support repsAgent assist (copilot)A second study contradicting the novice productivity finding
High volume of narrow repetitive questionsSelf-service AI agent with human escalationIf most contacts are complex, regulated or low-volume, where human service stays better
Mostly complex, regulated, or low-volume contactsKeep human service or a hybrid escalation modelVolume becoming repetitive and narrow enough to justify automation setup

Add a self-service AI agent only for narrow, repetitive questions at high volume. Pair that agent with an immediate escalation path to human staff, and judge success on resolution quality rather than cost per contact alone. The 34% novice gain in the NBER study and the lower quality Klarna's chief executive described under cost-led design are the two findings behind this order.

This applies to teams with enough repetitive volume to justify setup. Our verdict changes if a second academic study contradicts the NBER novice finding, or if most contacts are complex, regulated, or low-volume. Our AI customer service automation page covers how to automate support, and building a customer service chatbot covers the build. Pick one job from the table and list the contacts it would take before you add AI in customer service.

Frequently Asked Questions

  • Yes. Our guide on using ChatGPT for customer service covers reply drafts, ticket summaries, knowledge base updates and escalation rules, and our ChatGPT for customer support guide covers where it fails. The NBER figures come from one company's assistant and should not be read as a ChatGPT result.
  • No source we reviewed ranks AI tools for customer service, because the one academic study covers a single assistant at one company. For tool selection, see our Intercom vs Zendesk comparison and our best AI chatbot for business comparison.
  • The main disadvantages are falling effectiveness on complex problems and lower quality when cost drives the design. The CFPB found in consumer finance that chatbot effectiveness wanes as problems become more complex. Klarna's chief executive said, as relayed by CX Dive, that treating cost as too predominant a factor produced lower quality. A single tribunal decision, known through two law-firm summaries, found Air Canada did not take reasonable care to ensure its chatbot was accurate. Human handoffs and accuracy reviews are the practices our reading of these sources points to.
  • As of 11 November 2026, Intercom prices Fin at $0.99 per outcome, and a minimum monthly commitment applies, for example 50 outcomes. Intercom counts an outcome when a customer confirms their issue is resolved, when they do not ask for more help after Fin responds, or when Fin completes a workflow (Procedure), including handoffs.
  • One decision cannot tell you how liability applies to your business, so ask a qualified lawyer. In Moffatt v. Air Canada, a law-firm summary reports that the British Columbia Civil Resolution Tribunal found Air Canada did not take reasonable care to ensure its chatbot was accurate. It awarded a total of $812.02 in damages, pre-judgment interest and fees. That is a single small-claims decision. It is known through two law-firm summaries and is not settled law.
  • The sources we reviewed do not support replacing human agents outright. The NBER study found its gains concentrated in novice reps, with minimal impact on experienced ones. Klarna's chief executive said, as relayed by CX Dive, that treating cost as too predominant a factor produced lower quality. Our reading of these sources is to assist first. Keep human handoffs in place.

Want AI to Take Over Repetitive Support Steps?

Layer3Labs is an AI implementation partner for small and mid-sized business workflows. Our AI workflow audit reviews where AI can take over repetitive work.

Book a Consultation

This page is general information about reported studies and one tribunal decision, not legal advice, so talk to a qualified lawyer about liability for your own chatbot.