Scripted support bots often handled a narrow set of questions, such as order status, while teams measured how many chats ended without escalation. That measure can hide unresolved issues. A customer leaving the chat does not necessarily mean the bot helped them.
More capable models give CX teams options beyond a standalone chat widget. The work now includes agent assistance, voice, and proactive outreach, with evaluation tied to issue resolution and customer experience.
Four applications across customer support
1. Measure resolution
Ask whether the issue was resolved within 24 hours as well as whether the chat escalated. In our experience, teams that switched their primary measure from deflection to first-contact resolution improved CSAT and reduced cost. The change encourages the system to hand off a case when a human can resolve it more effectively.
2. Support the human agent
An agent-assist tool can draft a reply, retrieve the relevant policy, suggest follow-up questions, and prepare a case summary. Measure its effect on AHT and CSAT while checking the accuracy of those suggestions. The human agent needs to be able to review and correct the output before it reaches the customer.
3. Voice as a first-class channel again
Lower voice latency, multilingual support, and escalation with context have expanded the options for phone support, including for customers who do not use chat. We have shipped voice agents for logistics, hospitality, and retail clients that handle 60-70% of inbound volume without a human, with measurable NPS gains over the prior IVR.
4. Proactive over reactive
Session and event data can help identify customers who may churn, struggle with onboarding, or encounter a billing issue. Models can also help draft outreach for those situations. Test the targeting and message together, using a holdout where possible, to establish whether the intervention helps.
Where teams over-invest
- arrow_rightBuilding a customer-facing chatbot before fixing the help center it would draw from. The bot inherits the help center's gaps and amplifies them.
- arrow_rightAdding tools that do not connect to the agent desktop. Staff need the information and actions in the workflow they already use.
- arrow_rightCreating sentiment dashboards without an owner or response process. Define what the team will do when a signal appears.
An example deployment
We shipped a deployment this quarter for a mid-market e-commerce brand, combining a retrieval-augmented assistant, a voice agent, an agent-assist sidebar in Zendesk, and a churn-risk model connected to lifecycle email. One operations lead owned the deployment, weekly reviews tracked first-contact resolution, and every model had a rollback plan. AHT dropped 28%, CSAT held flat in the first month and rose three points by month three, and the chat-to-call escalation rate halved.
Assess the deployment by whether it resolves the customer's issue and gives staff the context they need. Deflection, response speed, and model quality are useful measures only when considered alongside that outcome.
