Great support is about helping customers get the right answer at the right time, with as little friction as possible.
As Moss has grown, both the complexity of questions and expectations around speed and consistency in customer support have increased. Over the last 12+ months, we’ve experimented with tools, workflows, and operating models to reduce response and resolution times while improving customer satisfaction.
This post explains why chat was the first channel that we chose to transform with AI, what changed in the customer experience, what the AI does today (and what it doesn’t), the impact we’re seeing, and what we’re still working on.
TL;DR
- We changed chat support from a rigid, form-like workflow to natural-language conversations.
- Today, ~50% of chat conversations are resolved end-to-end by AI without human involvement.
- Time-to-respond and time-to-resolution have dropped by 85+%, while further increasing our already high CSAT (4.5+/5).
- Next milestones: expand AI patterns to other channels and connect AI to customer-specific data (when appropriate) to resolve more cases end-to-end.
Why we started with chat
Chat is one of our most-used support channels: it’s fast, it’s conversational, and customers expect immediate clarity.
- Share of requests via chat: 50–60% of support requests come via chat
- Qualitative feedback: a large share of customers prefer chat as the initial interface when they want a quick answer or confirmation
Before vs after: the support experience
Before: structured, predictable — but often frustrating
Previously, chat required customers to pick a category upfront and follow predefined workflows. It worked, but it also created friction: pick the wrong case type and you’re slowed down; follow-ups become a checklist; and the conversation starts to feel like a form.
If you’ve ever thought “can I just talk to a human?” after clicking through a support flow, you know the feeling.
After: natural conversation, powered by AI
Today, customers can start typing immediately — in any language. The AI infers intent, asks clarifying questions when needed, and responds based on our internal runbooks and public Help Center content.
If a request requires human judgement or customer-specific information, the AI hands over seamlessly.
What the AI does today (and its boundaries)
Capabilities
The AI can:
- answer questions using our Help Center and internal support runbooks,
- guide customers through Moss functionality and settings,
- resolve most informational requests end-to-end,
- improve over time by learning from past conversations and CS/Ops feedback loops.
Boundaries
At this stage, the AI does not retrieve customer-specific data or execute actions in Moss. Given the sensitivity of the data we handle (card details, spend history, employee info, accounting configurations, and more), conversations remain informational, and escalation is used when customer context or human judgement is required.
Human handover
When a human agent should take over, the handover is designed to be low-friction for the customer.
- The AI collects the necessary information for handover (e.g., context, clarifying details, relevant identifiers provided by the customer) so the human agent can start with the right context.
- Escalation triggers vary but are tied to the chatbot’s confidence thresholds, an explicit customer request, or policy constraints.
Impact (speed, quality, and scale)
We’re already seeing measurable impact:
- ~50% of chat conversations are fully resolved by AI, with no human agent involvement.
- First response time: reduced by 85+%
- Time to resolution: reduced by 85+%
- CSAT: a slight increase on an already high baseline (4.5+/5).
Our next target is to resolve >75% of conversations across all support channels with AI alone, while maintaining a high-quality customer experience.
Operating model: CS + Ops as the quality owners
In an AI-first model, Customer Support isn’t just “handling tickets.” Together with Ops, the team owns answer quality and the end-to-end customer experience.
They continuously improve the system by:
- refining Help Centre↗ content and internal runbooks,
- updating guidance and best practices,
- reviewing edge cases where the AI got stuck,
- tuning responses to match real customer intent.
Data privacy and GDPR
Because Moss handles highly sensitive data, building an AI-first support experience comes with strict requirements around data handling and permissions: what can (and cannot) be used, shared, or referenced. We’ve designed this experience to handle customer data and PII GDPR-compliantly.
