PostNord Swapped Its Voice Bot in 90 Days: 60% of Calls Resolved Instead of 45%
PostNord's voice AI bot in Sweden hit a ceiling: it closed about 45% of calls without handing them to a human. The company replaced it with ElevenAgents, where an orchestrator identifies the request and routes it to one of four narrow subagents, including a dedicated parcel-tracking agent. The old system was switched off in roughly 90 days, and the share of resolved calls rose to 60%. If your bot has stalled on a plateau, try narrow agents and testing on live traffic.
AI-processed from ElevenLabs Blog; edited by Hamidun News
Most companies that deployed a voice AI bot a couple of years ago eventually run into the same wall: the share of calls resolved without a human stops growing. PostNord, the leading postal and logistics operator in the Nordics, hit this wall in Sweden. Instead of tweaking the old system, it decided to replace it entirely. The result: the bot was swapped in roughly 90 days, and the share of requests closed without escalation to a human rose from 45% to 60%.
An important caveat: this case study was published by ElevenLabs itself, so the numbers should be read as a vendor claim, not an independent audit. But the engineering details in it are more interesting than the percentage.
Why the old bot hit a ceiling
PostNord's previous system was also AI-based. It had evolved from in-house intent recognition models to LLMs from several providers. Even so, it stalled at roughly 45% of calls resolved. The new system had three requirements: more natural Swedish, a higher share of routine requests closed, and a platform that could be tuned directly in production.
Note that the 33% growth figure in the release is a relative number. In absolute terms, the gain is 15 percentage points.
How the multi-agent router works
At the core of ElevenAgents is an orchestrator. It listens to the caller, determines what they need, and hands the conversation to one of four specialized subagents. One of them handles only parcel tracking, the most common request. The logic is simple and familiar to anyone who has built a call center: high-volume questions go to a simple, cheap configuration, so a complex, expensive model is not spent on "where is my parcel?"
The agents are surrounded by guardrails, an intent classifier, and fallback rules for what to do when a request fits none of the skills. There is also separate A/B testing infrastructure: live traffic is split between different model configurations, and every change is validated on real calls. That is why 60% is the current value here, not a ceiling.
ElevenLabs also helped PostNord calculate LLM costs not per minute, but across the whole call path, from the start of the conversation to the resolution of the problem. A cheap minute does not help if the call ends in a transfer to a human. PostNord adopted this framing in its internal business case. The material gives no specific savings figures, only talk of "tangible savings."
Swedish and the phone network: where things usually break
The most practical part of the case is the list of problems that do not show up in a demo.
- Recognition. Transcription and end-of-speech detection were tuned for compressed telephone audio and regional dialects. The system mistook a short "ja" for background noise, so the bot kept talking over a person who had just answered.
- Voice. The team went through ElevenLabs voice models, balancing naturalness, speed, and reliability of sound.
- Carrier-grade telephony. ElevenLabs independently reproduced the SIP issues and worked directly with Telia's architects, while PostNord supplied packet-level analysis of SIP traces. Fixes came from both sides.
A launch without a hard cutover
Instead of switching over in one move, the rollout ran in four stages: an isolated weekend test, a hybrid mode with handoff to a human, a partial weekday launch, and full production. Before each go/no-go decision, both teams reviewed the calls from the previous phase. ElevenLabs engineers also held architecture walkthroughs and prompt-optimization sessions with PostNord's technical lead, so the team could adjust the system on its own after launch.
What to take from this
First, a plateau in the share of resolved calls is more often cured by architecture than by "one more prompt": splitting into an orchestrator and narrow subagents gives you both quality and cost control. Second, build A/B infrastructure from day one, or you will have nothing to optimize against. Third, most voice AI problems live not in the LLM but at the seam with telephony and language: dialects, short confirmations, SIP. Budget time for them.
Next, PostNord is carrying the same architecture over to Denmark, Norway, and Finland. That will test the key question: how easily the finished setup can be reconfigured for new languages.
- Nov 18, 2026AWS releases SMART: spin up a WhatsApp agent on AgentCore with one command
- Nov 18, 2026ElevenLabs Voice AI Handles 90% of Rohlik's Calls and Lifts Lead Qualification 3.3x
- Nov 18, 2026Anthropic: Claude with 413 note files earned 1 badge in Pokémon, but with a handful of files it earned 4
- Nov 18, 2026Configuring Qwen Code: QWEN.md, Subagents, and a Sandbox Cut Wasted Tokens
Key Questions Answered
Why the old bot hit a ceiling?
PostNord's previous system was also AI-based. It had evolved from in-house intent recognition models to LLMs from several providers. Even so, it stalled at roughly 45% of calls resolved. The new system had three requirements: more natural Swedish, a higher share of routine requests closed, and a platform that could be tuned directly in production. Note that the 33% growth figure in the release is a relative number. In absolute terms, the gain is 15 percentage points.
How the multi-agent router works?
At the core of ElevenAgents is an orchestrator. It listens to the caller, determines what they need, and hands the conversation to one of four specialized subagents. One of them handles only parcel tracking, the most common request.
What are the key details on Swedish and the phone network: where things usually break?
The most practical part of the case is the list of problems that do not show up in a demo. Recognition. Transcription and end-of-speech detection were tuned for compressed telephone audio and regional dialects. The system mistook a short "ja" for background noise, so the bot kept talking over a person who had just answered. Voice. The team went through ElevenLabs voice models, balancing naturalness, speed, and reliability of sound. Carrier-grade telephony.
Ship AI Faster with Founders, Engineers & Weekly Architecture Reviews
Inside Vibe Club (hamidun.com/club) — weekly live masterminds with Zhemal Khamidun, Tech Radar, full LMS access, and 370+ production Claude Code skills.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.