Norse Atlantic’s AI Agents: Cost Savings, Fraud, and Customer-Service Failures

Norse Atlantic’s AI Agents: How AI-Powered Customer Service Saved Costs—and Created New Risks

TL;DR

Norse Atlantic moved to an AI-first support model—chatbots branded Odin and Freya—to cut costs and staff. Vendors and the airline reported high automated-resolution rates.
WIRED obtained roughly 75 FTC complaints describing broken refund pages, missing phone numbers, and customers routed to scam phone lines. Of 41 complaints listing dollar amounts, 21 alleged losses over $1,000.
High “no-human-intervention” resolution rates can mask harm. Businesses must add fraud and safety KPIs, publish verified human escalation paths, and actively police search-result spoofing.

Hook: a single bad refund that tells the story

One traveler booked a cheap transatlantic fare, then needed a refund after a schedule change. The airline’s refund page failed. The website offered only chat; the chatbot closed the case. Desperate, the traveler Googled a support number and called—it turned out to be a scam. Multiple card numbers and weeks later, the customer was out hundreds, then thousands of dollars before reaching a senior executive and finally getting a refund.

That sequence—cheap fare, automated front door, broken refund flow, fake phone numbers in search results, financial harm—is the throughline of complaints filed against a single low‑cost, long‑haul carrier. It’s also a cautionary tale for any leader planning to scale AI agents for customer‑facing operations.

The rollout: Sprinklr, Odin, Freya—and a cost-cutting plan

Norse’s support stack evolved quickly. The airline used a unified customer‑service platform (Sprinklr) to consolidate channels, then deployed a Kindly-built chatbot branded “Odin,” and later switched to Delight.ai’s agent “Freya.” Vendors and Norse published metrics showing faster, cheaper resolutions; Delight.ai reported automated resolution rising from about 60% to roughly 80% within two weeks of Freya’s launch, and a company statement claimed Freya now handles the vast majority of inquiries—one executive even described the agent as managing roughly 99% of cases.

“Technology increases availability and customer support while helping keep fares low.” — Bård Nordhagen, Norse chief customer and communications officer.

Executives framed the shift as strategic: reduce admin headcount (a planned 35% cut), chase $50 million in savings under “Project Falcon,” and evolve support roles into “AI agent managers.”

“People will be managing and training AI agents and stepping in only when a human touch is needed.” — Alf Lim, Norse chief product officer.

The evidence: complaints, dollars, and recurring failure modes

WIRED’s public‑records request uncovered roughly 75 consumer complaints filed with the FTC. The complaints cluster around four failure modes:

Technical failures: refund pages that error or time out, broken forms that block refunds.
Hidden human help: missing or hard‑to‑find verified phone numbers and near‑total dependence on chatbots without clear escalation options.
Search-result fraud: consumers who looked for help found third‑party scam numbers surfaced by search engines and directories, and then shared payment or identity details with fraudsters.
Payment tangles: complex refund/payment flows involving third parties (names mentioned include Priceline, Xcover, Affirm) that automated agents couldn’t untangle.

Of 41 complaints that included dollar amounts, 21 alleged losses over $1,000. Dozens of people described giving multiple card numbers or sensitive personal data to fraudulent callers after failing to reach a verified Norse contact.

“This pattern of complaints is especially troubling,” said a consumer‑protection technologist, warning that automation had effectively placed customers directly in scammers’ path.

Regulatory options are murky. The FTC declined to comment on specific companies; consumer advocates have recommended state attorney‑general complaints as a practical route for affected customers.

Why automated metrics mislead

Vendors and companies often point to the “no‑human‑intervention resolution rate” as proof of success. That metric measures cases closed without escalation to a human. It does not measure whether the customer was satisfied, whether money was refunded, or whether the closure created exposure to fraud.

Two simple ways that metric can mislead:

Closure ≠ resolution: a chatbot can mark an inquiry resolved by sending an automated confirmation, even if the underlying refund never processed because a payment provider error needs human attention.
Scale hides edge cases: when most tickets are simple (itineraries, check‑in), automation shines. The costly, high‑risk cases—refunds, chargebacks, identity theft—are rarer but expensive. Masking those behind high percentages gives a false sense of safety.

Design risks specific to airline customer service

Airlines are uniquely risky for automation because they handle payments, identities, and time‑sensitive travel. Three systemic design gaps produce most of the harm:

Obscured escalation: If “speak to a human” is buried, customers will seek help elsewhere—often via search—exposing them to spoofed numbers.
Third‑party payment complexity: When bookings involve intermediaries or insurers, automated agents often lack access to cross‑platform accounting and policy nuance, stalling refunds.
Search and directory spoofing: Fraudsters exploit incomplete or slow updates to Google Business Profiles, travel aggregators, and SEO to surface fake support numbers above the official line.

What good AI-powered customer service looks like

Automation is a force‑multiplier—but also a magnifying glass for UX and fraud. A safer approach combines AI agents with explicit human-in-the-loop design, transparent contact points, and outcome‑oriented metrics.

Immediate checklist for leaders

Publish verified contact points: Prominently place verified phone numbers and an obvious “speak to a human” CTA on payments and refunds pages. Use schema.org markup and verify your Google Business Profile.
Guarantee escalation SLAs: Define time-to-human targets for high‑risk tickets (e.g., human within 4 hours for refund requests over $500) and publish them publicly.
Track safety KPIs: Add fraud incidents per 1,000 tickets, customer financial loss reported, time-to-reimbursement, and CSAT for escalations to standard dashboards.
Monitor search results: Weekly audits of organic search listings and third‑party directories; rapid takedown requests and legal escalation for spoofed listings.
Harden payment flows: Flag multiple card attempts, require additional verification for large refunds, and maintain human review checkpoints for borderline cases.
Offer easy remediation: A simple, visible reimbursement policy for customers who fell prey to fraud while trying to contact the company directly.

KPIs that matter (beyond “resolved without human intervention”)

Escalation ratio (percent of tickets routed to human agents)
Time-to-human for high‑risk categories
Fraud incidents per 10,000 tickets
Average financial loss reported per incident
Reimbursement speed (median days to reimburse victim)
Post-escalation CSAT for resolved high‑complexity tickets

For product teams

Design chat flows with explicit exit hatches: if a bot fails to confirm refund processing within X steps, immediately surface a phone callback or prioritize a human case.
Instrument end-to-end events: capture refund initiation, payment-provider response, and final settlement so the support system can detect and escalate failures automatically.
Run adversarial tests: simulate search‑spoofing and misdirected support flows to identify where customers are most likely to be led astray.

For legal and operations

Map liability: document where the airline’s responsibility begins and ends with third parties, and prepare clear customer remediation policies.
Coordinate with payments and fraud teams: ensure chargeback procedures and expedited investigations for suspected impersonation calls.
Engage regulators proactively: if automation changes access to human support, notify consumer‑protection authorities and publish a compliance plan.

Questions for your board

Has the company published verified human contact points for payments and refunds?

No contact points visible? That’s a red flag.
Do our KPIs include fraud incidence, customer financial loss, and time-to-reimbursement?

If the answer is no, leadership lacks the metrics to judge customer safety.
Is there a published SLA guaranteeing human escalation for high-risk tickets?

Without an SLA, automation can become a black box for vulnerable customers.
Do we monitor search and directory listings for spoofed numbers weekly?

Search management is operational security for customer support.

There will always be a trade‑off between cost and coverage. AI agents and chatbot automation deliver huge efficiency gains for routine issues, and businesses should use the technology aggressively where appropriate. But the Norse experience shows that when payments, identity, and financial redress are involved, automation must be paired with explicit human access, fraud safeguards, and outcome metrics—not just closure percentages.

Automation is a lever, not a destination—use it with guardrails.