What the AI Chatbot Can't Answer Is the Most Valuable Number in Customer Service
When AI in customer service is discussed, the conversation almost always revolves around a single number: the share of inquiries a system handles autonomously. The higher, the better. The math behind it is cost accounting, and it ends right where it begins—with saved effort.
A case from furniture retail is currently often cited as a counterexample, usually reduced to "they didn't lay people off, they reskilled them." That is true, but it misses the bigger picture. The most interesting part lies in the number no one is talking about: the share the system was unable to answer.
What actually happened
In 2021, Ingka Group, the largest IKEA franchisee, introduced an AI-powered customer service chatbot named Billie. Between 2021 and 2023, according to company statements, it resolved around 47 percent of incoming inquiries—roughly 3.2 million interactions, yielding savings of nearly 13 million euros.
This is where most success stories end. The standard next step would have been to reduce headcount proportionally: if half of the inquiries are automated, half of the team is no longer needed.
Instead, the company looked at the other half.
The inquiries the chatbot couldn't resolve weren't harder variations of the same routine tasks. They were an entirely different kind of request.
Across the remaining 53 percent, a recurring pattern emerged: customers were asking for help with interior design and space planning. In other words, for advice—about taste, context, tradeoffs. None of this can be captured in a conversational system, which is precisely why it ended up marked as "unresolved."
The key point: This demand was previously invisible. As long as employees were consumed by delivery dates and return queries, no one had the bandwidth to even notice the pattern—let alone act on it. The chatbot absorbed the routine workload, revealing what customers actually wanted a human for.
The three decisions that turned the finding into a business
What happened next
Reskill
instead of cut
Build a dedicated
channel
Charge
for it
The remote sales channel hosting these consultations generated roughly 1.3 billion euros in fiscal year 2022—about 3.3 percent of the group's total revenue. The stated goal is 10 percent by 2028.
Two caveats to keep in mind
The 1.3 billion figure is currently circulating through many articles, usually without context. Two details must be added to prevent an instructive case study from turning into a myth.
The 1.3 billion represents the revenue of the entire remote sales channel, not the incremental gain from the reskilling program. A portion of this business existed previously. The figure illustrates the scale of the area this initiative contributes to—it is not a before-and-after comparison.
Anyone using it as proof of the reskilling program's impact is making the same mistake as presenting a record metric without a baseline. The direction is right, but causation is not proven.
This case is often cited as proof that automation and employment are not a zero-sum game. This narrative took a hit in 2026: in March, Ingka announced the elimination of around 800 office jobs, predominantly in Sweden and the Netherlands; in May, franchisor Inter IKEA followed with the announcement of around 850 additional job cuts worldwide.
This does not invalidate the retraining—it affected different areas and different roles. But it shows that the case is not a silver-bullet success story. What it proves is narrower, yet still valuable: that a company can redirect freed-up capacity into a revenue channel instead of cutting it. Not that retraining permanently prevents job cuts.
What is transferable—and what isn't
Furniture planning is an ideal case study: the need for advice is obvious, the outcome is visible, and the average order value is high. Things are different in banking and insurance, yet the mechanics remain the same. Four points can be transferred:
First: The non-resolution rate is a source of insight, not an error metric. In almost every service department, the rate of automation is measured and the rest is treated as leftover volume. Yet that remainder is precisely where customers want a human touch. Anyone who evaluates unresolved inquiries by topic rather than frequency will uncover the advisory services they should be offering.
Second: The bridge from inquiry to appointment is missing almost everywhere. This is the most expensive drop-off in the process. A customer asks a question the bot can't answer and gets a phone number, a contact form, or a transfer to a queue. The moment it becomes clear that advice is needed, a bookable appointment should be offered—with the right appointment type and with the right expert. This is not a technical question, but one of process design.
Third: Advisory services need their own revenue calculation. As long as customer service is managed as a cost center, every measure there is judged on cost reduction—and any shift toward advisory looks like extra effort. Only when the resulting deals are attributed to the channel does the evaluation flip.
Fourth: Retraining is a decision with lead time. Anyone expecting service reps to become advisors on their own as soon as the bot is live will be disappointed. It requires training, a defined service offering, and managers who actually want to fill this new role. This is the part where practice fails most often—and the part no software can provide.
What you should measure
Standard metrics in customer service are one-sided: automation rate, handling time, cost per contact. All three measure the same side of the coin. Four metrics complete the picture:
| Metric | What it answers |
|---|---|
| Unresolved inquiries by topic | What do customers actually want a human for? The topic list is your service menu |
| Share converted into an appointment | The bridge from inquiry to consultation—usually the largest gap in the entire process |
| Conversion rate per advisory call from the service channel | Whether the channel generates revenue or merely shifts workload |
| Contribution margin of the advisory channel | The metric that removes customer service from cost-center logic |
The first row is the entry point and costs nothing except an analysis you could already be doing. It answers the question this article started with: What can't your system do—and is that a problem or an offering?
Takeaway
The degree of automation measures the cost side. The key insight lies in the remainder: What a system cannot answer shows what customers want a human for — and thus, which advisory service could become a revenue channel. Three steps turn this into a business: train people for it, make the service bookable, and give it its own revenue calculation. And the biggest gap lies between steps one and two — right where an inquiry should turn into an appointment and a phone number appears instead.
Sources
Metrics on the chatbot resolution rate (around 47 percent, 3.2 million interactions, nearly 13 million euros in savings, 2021–2023) and on the reskilling of 8,500 employees: Ingka Group newsroom release "AI and Remote Selling bring IKEA design expertise to the many" (ingka.com).
Remote sales channel revenue of around 1.3 billion euros in fiscal year 2022, representing 3.3 percent of total revenue, as well as the target of 10 percent by 2028: reported in multiple sources, including robertjuliansmith.com and fluentsupport.com. For context on this figure as channel revenue rather than program-driven growth: scottmrobertson.substack.com.
Job cuts in 2026 (around 800 positions at Ingka in March, around 850 at Inter IKEA in May): employerbranding.news citing Eurofound. 2025 business figures and data on the number of customers advised: pymnts.com.
The inspiration for this article came from a LinkedIn post by Markus Brunold on the same case. The contextualization and application to banking and insurance are my own additions.
Share with your team
Three things a leader can implement this month after reading this article:
- Analyze unresolved requests by topic. Not by frequency, but by content. This list exists in every service system and is hardly ever read—it is the cheapest market research you have.
- Build the bridge. Where a phone number or a form appears today, there belongs a bookable appointment—with the right appointment type and with the responsible person. This is the point where a service contact turns into advisory.
- Give the advisory channel its own line item in the calculation. As long as closed deals from the service channel don't show up anywhere, every shift toward it remains a cost increase—and gets cut accordingly.