KB Kookmin's AI Resolved 41.3% of Calls, Staff Workload Rose

Even as AI handles more inquiries, the remaining calls can get harder

KB Kookmin Bank’s AI handled 41.3% of customer inquiries in 2025. That figure comes from data Kyunghyang Shinmun, a Korean daily newspaper, obtained through the office of Rep. Lee Jung-heon and reported this past April. The bank’s customer-service staff, the report noted, shrank from 1,133 to 869.

Choi Na-young, a customer-service representative quoted in the article, said she used to consider 70 calls a day a reasonable workload; now she handles 100-120. Her explanation: as AI took over simple inquiries, what’s left for humans are the harder ones. This figure is her own testimony, not a bank-wide average across all representatives.

What caught my attention in Choi’s account is this: even as the resolution rate rose, the number of calls she personally handles actually increased. Just because AI resolved 4 out of 10 inquiries doesn’t mean the time humans spend dropped by 40%. What changes together are which inquiries get resolved, how much time the remaining ones take, and how many people are splitting the load.

For instance, automated systems might handle brief informational inquiries, while humans still take on complex disputes or compensation negotiations. If staffing shrinks in the process, the workload concentrates on the representatives who remain. This is easy to miss if a rollout plan calculates headcount by looking only at resolution counts.

Bank-by-bank results didn’t all move in the same direction

It’s worth looking at the other banks’ figures that appeared in the same report.

Bank	Reported results
Shinhan Bank	After adopting AICC in 2024, counseling staff grew from 693 to 697 in 2025
NongHyup Bank	AI handling rates were 25.6% in 2023, 20.7% in 2024, and 21.7% in 2025
Toss Bank	AI handled seven out of ten inquiries, but in February 2026 chatbot satisfaction stood at 36% versus 72% for the traditional call center
Hana Bank	In an internal 2024 survey, 53% of customers who stuck with the call bot through to the end reported dissatisfaction

This table isn’t the result of testing the same technology under identical conditions. Adoption timing, systems, and the composition of customers and inquiries all differ. Nor can we assume the satisfaction surveys drew respondents from comparable user bases across channels. The table alone can’t tell us which model or which bank’s operations are actually better.

What it does show, though, is that adopting AI doesn’t automatically translate into fewer staff or higher satisfaction. Even when the handling rate dropped, we need to look harder at why. It might be that answer quality got worse — but it’s equally possible that the mix of inquiries being processed, or the counting criteria themselves, changed instead.

That’s why I think the first question, when comparing adoption outcomes, should be which inquiries the AI was actually applied to. Under the single label “AI counseling adoption,” systems that do entirely different jobs are being lumped together.

AI that assists agents and AI that faces customers directly play different roles

A 2025 academic paper by Brynjolfsson, Li, and Raymond analyzed 5,172 customer support agents at one company. When a tool that suggests responses to agents was rolled out in stages, resolutions per hour rose by an average of 15%. The tool only showed suggestions to the agent, who was free to revise or ignore them. Generative AI at Work

The effect wasn’t evenly distributed. Less experienced, less skilled agents benefited more, while the most skilled group saw only small speed gains — and in some cases, quality even dipped slightly. Gains were largest for moderately uncommon problems, not the very common or the very rare ones. Common and easy are not the same thing.

This study didn’t measure the effect of AI handling customers on its own. It measured the outcome of human judgment and AI suggestions working together. If you translate that 15% into “the share of support staff a chatbot lets you cut,” the conditions change entirely.

The two deployment approaches can be broken down as follows.

Where AI sits	Main expected help	Question to ask about what’s left for humans
Suggests answers to the agent	Supports information lookup and drafting responses	How long does reviewing the suggestion take, and for which inquiries does it help?
Customer talks to AI first	Handles some inquiries directly	Which inquiries get escalated to a human, and is the prior conversation passed along?

Real services can combine both approaches. There’s no need to generalize that AI takes over every easy inquiry, or that assistive tools never change the mix of inquiries at all. It’s more accurate to directly compare the types of inquiries and handling times before and after adoption.

In particular, when a customer gets escalated to a human after talking with AI first, the agent may have to correct guidance that already failed. If the customer has to repeat the same explanation, frustration can build further. Simply counting this as “one human handoff” doesn’t reveal this extra burden.

Where the saved time actually went matters too

A research team led by Seo and Oh, backed by the Bank of Korea, surveyed 5,512 employed workers in Korea in May–June 2025. Among those who had used generative AI in the previous week, average time savings came to 3.8%; converted across all employed workers including non-users, that figure was 1.4%. Both the changes in working hours and in output are self-reported values. Generative AI and the Reallocation of Time

The correlation coefficient between time savings and output change was a mere 0.008. In other words, reporting a large drop in time spent didn’t linearly translate into reporting a large increase in output. The researchers point to fixed quotas, slack during work hours, and task reallocation as possible explanations.

Nor did every task see reduced time. Of the 156 tasks analyzed, 140 saw time decrease, but some tasks — like handling customer complaints or physical coordination — reportedly took longer. This alone can’t prove why a particular bank call-center agent’s workload increased. Still, it’s worth keeping in mind alongside the overall average, since individual tasks can move differently from it.

I don’t read this result as “Korea used AI and got nothing for it.” Handling the same volume of work with less strain is itself a gain. If that saved time went back to the worker as rest, that has value from the worker’s perspective too. If the goal is to increase output, you need to look at the constraints and incentives around other tasks as well.

Even for call centers, you first need to decide what you’re actually trying to get out of it. Whether you want to cut wait times with the same headcount, spend more time on difficult inquiries, or reduce work intensity — each goal calls for a different staffing plan and evaluation criteria. That choice has to come before converting all the saved time into headcount reductions.

There’s More to This Job Than Writing Replies

I find it useful to break the work of running a support AI into four separate tasks. That doesn’t mean you need four new job titles. It’s a way of making sure nobody’s time gets overlooked.

Task	What it involves	What it should produce
Handling responses and inquiries	Drafting replies, personally handling exception cases	Customer requests and how they were actually resolved
Knowledge management	Updating current policies, pricing, and exception rules	Reference material with clear effective dates and owners
Quality evaluation	Reviewing failure cases, maintaining evaluation criteria	A recurring set of cases to check and the reasoning behind each verdict
Exception recovery	Managing handoff conditions, compensation, and reprocessing procedures	A recovery process with a clear owner, authority, and deadline

Say a refund policy changes. Simply posting new wording may not be enough. You need to decide which policy applies to existing customers, and check whether the AI is correctly distinguishing exception cases. Someone also needs to be assigned to reach back out to customers who were given wrong information — and with what authority to resolve it.

Even if you can automatically score entire conversations, the evaluation work doesn’t disappear. You still need to check whether the scoring criteria match current policy and whether they correctly judge ambiguous cases. If some conversations the machine marked as “resolved” were actually cases where the customer just gave up, someone still has to find those and fix the criteria.

You can assign this work to your existing support staff and managers. But if you simply add it on top of their existing targets, total workload goes up. Training time, review time, and operational authority all need to be allocated alongside it. Workforce planning should account not just for the number of tickets automated, but for the time this work requires.

Every Product Defines “Resolved” Differently

To compare resolution rates, you first need to know what each vendor counts as “resolved.” Since a vendor’s own counting standard often doubles as its billing standard, this is something to verify before signing a contract.

Fin distinguishes between cases the customer explicitly marked as resolved and cases where the customer simply didn’t ask for further help after receiving an answer. The latter is an inferred resolution. It doesn’t automatically count a resolution just because the customer exchanged greetings or asked only clarifying questions without receiving an answer. There’s also a rule that if the customer later says they need more help in the same conversation, the earlier resolution charge gets reversed.

There’s also the case where Fin completes a defined procedure and then hands the conversation off to a human — this can still count as a billable outcome. That needs to be distinguished from a simple, unsuccessful handoff. Summarizing it as “handed off to a human, so it’s free” or “customer went silent, so it’s automatically billed” misses the actual conditions. Fin pricing outcomes guide

The tier system Zendesk introduced in May 2026 distinguishes between cases handed off to a human, cases where the conversation ended without further requests but didn’t pass verification, and verified resolutions. The first two are not deducted from the resolution quota; the last is confirmed by having a language model evaluate the conversation content. Existing accounts may still be on the previous system, so it’s worth checking which version applies. Zendesk resolution tiers

Even a vendor’s own verification can’t fully substitute for an operations team’s own checks. Whether the customer contacted support again about the same issue, switched to a different channel, or needed a refund or reprocessing — these are things you may need to track separately.

Resolution rates are shaped jointly by the model’s capability, the knowledge and tools it’s connected to, the handoff rules, and the composition of inquiries. Changing the rules can change the numbers, but that doesn’t mean the model’s performance doesn’t matter. What matters is identifying exactly what changed to move the numbers.

Operational metrics are best read alongside the following:

• The number of cases the AI handled directly, and the counting standard used
• Repeat inquiries about the same issue and cases that moved to a different channel
• The type and handling time of inquiries escalated to humans
• Customer satisfaction and the composition of respondents
• Agents’ post-processing, review, and recovery time, and their overall workload

Satisfaction scores can miss customers who never respond, and repeat-contact rates can come out artificially low if you’re not connecting the same customer across channels. Don’t stop at bolting on one more metric — you also need to regularly read through actual complaints and failure cases.

Oswarld’s Lens

I think that when you look at an automation rate, you have to look at the work-assignment pattern behind that number too. If the number went up, you need to check which inquiries got automated, and who ended up handling the ones that remained — and whether that person can actually absorb the extra load on top of what they already do.

Who captures the gains when productivity rises also matters. The customer-support paper I discussed earlier didn’t have direct compensation data. The managers interviewed said incentive pay had probably increased, but warned that the gain might not last once targets were raised afterward. It wasn’t a study that concluded wages actually went up.

Domestic reporting has also surfaced analysis suggesting that the estimated average wage at call centers with heavy AI adoption is lower than at centers without it. That gap alone can’t confirm that AI is depressing wages, but it’s reason enough to examine how automation gains get distributed in a structure where call volume is tied to incentive pay.

The role of experienced staff also needs a second look. The paper noted that conversations from skilled agents carry value as training data in their own right. It also flagged, as a question for future work, the possibility that if skilled agents simply follow AI suggestions, fewer novel solutions might emerge. If you’re going to assign someone like that knowledge curation and quality evaluation on top of handling calls, I think you have to recognize the time and compensation that work deserves.

In outsourced/contracted structures, you need to make sure operational knowledge doesn’t get severed when contracts change. If the people who understand policy exceptions and the context behind customer interactions leave, documentation alone may not be enough. Someone has to be designated to maintain the materials, correct errors, and hand things off to the next person in charge.

In the end, what I want to confirm before any workforce plan is: who is responsible. Decide first who will update the knowledge base, who will review failure cases, and who will re-explain things to customers and fix problems. And I don’t think you can pile that work onto the remaining staff while calculating that “fewer inquiries means less work.”

Automation offers no guarantee of preserving net employment. The headcount needed for the same output can shrink. That’s exactly why throughput, working hours, and service quality need to be measured separately. Only by counting the reduced inquiries alongside the new operational work that’s been created can you tell whether the staff cuts are turning into overload for the people who remain.

I think the results of an AI rollout have to include both the customer problems it resolved and the workload of the people producing those results.

💬 Where do the inquiries AI can’t handle end up going? Is the person handling them given dedicated time for review and recovery?

References

• Kyunghyang Shinmun, “AI Adoption at Call Centers Cut Wages Too”, April 15, 2026: bank data and field interviews
• Brynjolfsson, Li & Raymond, Generative AI at Work, QJE 2025 published version: the effects and limits of AI-assisted support tools
• Suh & Oh, Generative AI and the Reallocation of Time, February 2026: shifts in time, output, and task composition found in a survey of Korean workers
• Fin, Outcome-Based Pricing Guide
• Zendesk, About Automated Resolution Tiers · Prior Methodology and Applicable Accounts

Read together: AI Customer Support: What Counts as One Resolution · Designing Verification Work and Reviewer Roles

The author is Oswarld (Kwangseob Ahn). Current roles: Adjunct Professor at Sejong University, Strategy Consultant at INLEVEL9. Career, research, books, and recent work are kept current on the About page. Latest · July 2026: HEMA-2: A Consolidation-Aware Tri-Memory Architecture with Multi-Channel Scheduling for Lifelong Conversational AI.