AI didn't replace our support team. It made them dangerous.
Nikolay Tsanov28 May 2026 6 min read Everyone is asking whether AI will replace customer support. After a year of running AI-assisted support for our clients, we think that's the wrong question entirely.
Every week someone asks us the same question: "Should we just replace our support team with an AI chatbot?"
Our answer surprises them: no. But you should absolutely stop running support the way you did in 2023.
The chatbot trap
Pure-AI support fails for a predictable reason: the conversations that matter most are exactly the ones AI handles worst. Password resets and shipping queries? Automate them, nobody will miss the human touch. But the angry enterprise customer threatening to churn, the confused user who can't articulate their problem, the edge case your docs never imagined: those conversations decide your retention curve, and they need a human with judgment and authority.
Companies that went all-in on chatbots learned this the hard way. Deflection rates looked great in the board deck. Then NPS cratered, and the support inbox became a complaint inbox about the support itself.
What actually works: the centaur model
The setup we run for our clients pairs human specialists with AI at every step:
- AI triages everything first. Intent, urgency, sentiment, and language, all classified in seconds and routed to the right place. No human reads a queue cold anymore.
- AI drafts, humans decide. Agents get a suggested reply with sources from the knowledge base. They edit, personalise, and own the final word. Response time drops by half; quality goes up, not down.
- The boring 40% disappears. Refund status, plan changes, and simple how-tos get fully automated, with a human escape hatch that's always one message away.
- Every conversation feeds the product. AI clusters tickets into themes, so instead of "users seem confused about billing" your product team gets "23% of this month's tickets trace to the new invoice screen."
The result isn't a smaller team doing the same work. It's the same-sized team doing work that was previously impossible: proactive outreach, quality reviews on every conversation, real product insights instead of gut feelings.
The numbers that convinced us
Across our CX engagements over the past year, the centaur model has consistently delivered first-response times under two minutes during covered hours, CSAT holding above 90% even during volume spikes, and (the one clients care about most) support insights directly shaping the product roadmap every single sprint.
That last one is the quiet superpower. Support stops being a cost centre and becomes the cheapest user-research programme you'll ever run.
Where to start
If your team is drowning, don't start by buying tools. Start by mapping one week of conversations: what's repetitive, what's risky, what's revealing. Automate the first, protect the second with humans, and build a pipeline so the third actually reaches the people building your product.
Or let us do it for you. Running customer experience as a service, AI included, is literally our job.