Stop Answering 'Where's My Order?' Sixty Times a Day
The cheaper AI support project your agents can ship this afternoon
Klarna's OpenAI assistant handled 2.3M tickets in month one — then in 2025 the CEO quietly walked the AI-only stance back. The direction held; the project shape didn't. Most support teams are still building the wrong thing.
LiveChatAI's 2025 50-industry teardown puts SaaS support at $25-35 per ticket and self-service deflection at $1-4. The shippable AI play hides in that gap — and it isn't a chatbot.
The Math You're Not Letting Yourself Do
Look at this week's queue. Roughly 200 of every 312 tickets say one of five things: where's my order, how do I return this, can I get a refund, what's your shipping policy, does this fit. Salesmate and Desk365's 2026 benchmarks land in the same range — 60-80% of inbound tickets are repeat FAQ traffic across industries. The customer whose order arrived destroyed and is threatening to dispute the charge waits in the same queue as "what's your return window?"
The cost-per-ticket nobody puts on a slide
At $25-35 per SaaS ticket or $5-7 retail, the math is brutal. A 500-ticket week with 70% repeat traffic costs $8,750-12,250 in agent time — to answer questions your help center already answers. Guideflow's 2026 ticket-reduction analysis found only 36% of self-servable tickets actually get self-served. Customers won't read the help center. So your best agents copy-paste from it.
The tax that doesn't show up on the dashboard
Your best agent didn't take the job to copy-paste from a knowledge base. They took it for the 20% of tickets that need judgement — the destroyed-in-transit, the angry-but-fixable, the edge case that becomes a process improvement. Support leads who've made this trade explicitly tend to redesign the queue before they redesign the team.
Why the Chatbot Project Keeps Failing — And the Cheaper Thing That Doesn't
In February 2024, Klarna's OpenAI-built assistant ran 2.3M conversations in month one — two-thirds of total volume, the work of 700 full-time agents, response time from 11 minutes to under 2. An 82% speedup. A 25% drop in repeat issues. Then in 2025 CEO Sebastian Siemiatkowski walked it back and rehired humans for complex tickets.
Cost was a predominant evaluation factor in organizing support — resulting in lower quality. — Sebastian Siemiatkowski, CEO, Klarna (2025, on the AI-only walk-back)
The direction held. AI carries the repetitive volume; humans carry the hard tickets. The shape that failed was AI-as-replacer. The shape that ships is AI-as-drafter, human-as-reviewer — and it doesn't need a customer-facing widget, a six-month build, or an engineering team. Your agent pastes the ticket. The drafter writes the reply from your docs and cites the source line. Your agent verifies in five seconds and sends. Zero customer-facing risk because the customer never sees the model. The agent still owns the send button.
What the Drafter Actually Needs Before It's Safe to Ship
Grounded in your docs, not the model's training data
A drafter answering from the model's training data hallucinates your refund policy. A drafter grounded in your help center, FAQ, return policy, refund policy, shipping policy, and 2-3 internal SOP notes answers the same questions your best agent does — in your tone, with a citation. The citation is the unlock: agents trust a draft they can verify in five seconds. They don't trust a draft they have to re-research from scratch.
The four guardrails that keep humans on what matters
In-scope topics get a draft. Always-escalate triggers — legal threats, chargebacks, refund-outside-policy, account-specific data, anything emotional or sensitive — return "ESCALATE" and a one-line summary, no draft. Never-promise rules (specific delivery dates, custom discounts, refunds outside policy) are wired into the system prompt. Anything unclear comes back as one or two clarifying questions the agent asks the customer — not invented answers.
The reframe collapses the project from six-month chatbot build into an afternoon. The deeper read: "AI strategy" at the support-team level isn't about replacing humans. It's about pulling the boring 60-80% off the people you can't afford to lose, so they're on the desk when the destroyed-in-transit ticket lands.