An agent who joined your contact centre in 2023 spent the first week on address changes, invoice copies and “where is my parcel”. Forty calls a day, most of them easy. By Friday they knew the systems and had spoken to a couple of hundred customers before they met an angry one.
The agent you hire in 2028 gets none of that, because by then the AI answers those calls. Their first live call is a customer who has been through the AI twice, is disputing a charge and wants to know why nobody rang back.
When AI takes the simple contacts, your people are left with the hard ones. Handle time rises, first-contact resolution falls and the workload shrinks far less than the volume, without anyone doing a worse job. The calls new hires used to learn on disappear too. Norway’s customer service leaders are planning for AI to take the simple contacts. Their own answers suggest few have planned for what that leaves behind.
What 47 Norwegian service leaders are planning
KSIndeks and the technology company Heisenbug surveyed 47 Norwegian customer service leaders, mostly in mid-sized and large organisations. The report, AI i norsk kundeservice 2026, is dated September, and KSIndeks has written up the findings in two articles, on systems and barriers and on the people who remain.
- AI resolves 16% of customer contacts without a person today. By 2028, the leaders expect 40%.
- About six in ten expect fewer staff by 2028: 21% a significant reduction, 38% a moderate one.
- 55% name changed job content, with more complex tasks, among the most important changes towards 2030.
- 34% plan to use AI in recruitment and training. Of the nine uses the survey mapped, that is the lowest.
KSIndeks is careful about what this shows. The sample is small and self-selected, the figures are self-reported, “resolved by AI” has no shared definition, and the 2028 numbers are plans. The 34% counts AI used for training, and says nothing about training budgets. Even so, more than half of these leaders expect the job to get harder, and preparing people for it is where the fewest plan to use AI.
Some of the plans are already decisions. On 6 October, DNB told the stock exchange it would cut about 400 full-time equivalents in its Technology & Services unit before the end of the year, because the technology “replaces tasks that were previously handled manually”. Those cuts are not in the customer centre. But at the bank’s capital markets day on 10 November, its head of personal banking, Maria Ervik Løvold, told Nettavisen what happens when the digital assistant can’t help: “If you get stuck, the whole context follows you over to a person in the customer centre.”
That is how a handoff should work. It also means the person in the customer centre gets every customer who got stuck.
Same team, same skill, worse numbers
Illustration: 1,000 contacts a week. Half are simple, taking four minutes each, and 90% of them are resolved on first contact. Half are hard, taking ten minutes, and 70% are resolved on first contact. The AI resolves only simple contacts, and your people handle each kind exactly as well as they did before.
| AI resolves 16% | AI resolves 40% | |
|---|---|---|
| Contacts reaching your people | 840 | 600 |
| Of which hard | 500 (60%) | 500 (83%) |
| Average handle time | 7.6 min | 9.0 min |
| Resolved on first contact | 78% | 73% |
| Minutes of work | 6,360 | 5,400 |
Handle time is up 19% and first-contact resolution is down five points, and nobody got worse. The easy calls that pulled the averages up are gone. A team lead who reads those two rows without the mix behind them will coach people for a problem they don’t have.
The spreadsheet says 14 agents. The work says 17
Now compare the first row with the last. Contacts reaching your people fall 29%. Minutes of work fall 15%. KSIndeks makes the same point in its write-up: lower case volume doesn’t have to mean a correspondingly lower workload.
Staff a team of 20 by contact volume, and the spreadsheet gives you 14. Staff it by the work, and you need 17. The three missing people show up as queue time, on the calls where customers have the least patience left.
That gap is one route to the rehiring Gartner predicted in February: half of the companies that cut service staff because of AI will hire again for similar work by 2027. When Gartner surveyed 321 service leaders, 31% had made or were planning AI-driven frontline layoffs, and 85% were expanding what their agents are responsible for.
The numbers in the table are invented, and yours will differ. Your AI may take some hard contacts, and a good handoff saves a minute or two. Redo the table with your own contact reasons before anyone sets next year’s headcount.
New hires lose the calls they learned on
Easy calls were cheap volume, and they were also the training programme. A new agent learned the systems, the tone and the product on calls where a clumsy start cost nothing. Nobody complains about a slow address change.
Take those calls away, and a new hire starts on the hard ones. They take longer to get good, or they leave before they do. And the customer the AI has already failed once gets the least experienced person on the floor.
It doesn’t have to take as long as it used to. In a study of more than 5,000 support agents, Brynjolfsson, Li and Raymond found that an AI assistant raised issues resolved per hour by 15% on average, and the newest and least skilled agents gained the most. Agents two months into the job with the assistant performed like agents with more than six months without it.
An assistant helps during the call, though. It doesn’t give anyone practice. Before a new hire takes a disputed charge from a real customer, they should already have handled one several times, out loud, against a customer who pushes back. That is what role-play is for, and it works best when it’s built from your own hard calls. We’ve written about why most agents still learn the hardest conversations on real customers.
Four changes to make before the mix shifts
1. Report by contact reason. Blended handle time and first-contact resolution will move with the mix, whatever your people do. Billing disputes compared with last quarter’s billing disputes won’t. Check the targets as well: a handle-time target set in 2025 was set for a different job.
2. Staff by minutes. Forecast the contacts that will still reach people, multiply by the handle time for each contact reason, and build the headcount from that. Contact volume alone will tell you to cut too deep.
3. Let new hires rehearse the hard calls first. List the five contact reasons that will dominate the human queue. Build practice for each from real calls, and have new hires run them before their first live shift. Then check whether it worked on their real calls, against colleagues who didn’t practise.
4. Score the handoff. Every escalated call starts with what the AI did. Did the context arrive? Did the customer have to repeat anything? How many times did they ask for a person before they got one? Score the AI’s part and the agent’s part on the same scorecard, or the agent takes the blame for a call that was lost before they picked up.
The first disputed charge
The agent you hire in 2028 will take a disputed charge on day one. Nothing in the survey suggests otherwise. What you can still decide is whether that call is the first time they’ve had the conversation or the thirtieth.
Which calls will your next new hire learn on?
Bring one of the hard calls your AI will leave to your people. We’ll fully anonymise it and turn it into an AI Coaching voice role-play while you watch, so your next new hire has had the conversation before the customer rings.







