Smarsh deployed Salesforce Agentforce across its customer support operations and cut internal handling time by 7.5 hours per case. The communications data company also hit a 72% deflection rate, resolving most inquiries without a human agent. For support teams under pressure to do more with less, those numbers put a hard metric behind the promise of AI-assisted service.
How the numbers break down
The Agentforce implementation produced three performance indicators that support leaders will recognize immediately. The 72% deflection rate means nearly three out of four inquiries never reach a live agent. A Resolution Confidence Score of 2.6 out of 3 reflects high reliability in the system's automated responses - high enough that teams can trust the AI to handle routine work without constant oversight.
Internal results matched the external gains. Smarsh reported a 65% adoption rate for its internal AI agent, named Emmy, with AI resolving 31% of all cases end-to-end. The 7.5 hours saved per case frees specialists for escalations and complex troubleshooting that actually require human judgment.
Internal adoption and the Emmy agent
Smarsh didn't limit Agentforce to customer-facing queues. The company rolled out Emmy as an internal support tool, and two-thirds of the organization adopted it. That 65% internal adoption rate suggests the AI agent cleared a common hurdle: getting employees to change how they work. When internal teams use the same tool they deploy for customers, the feedback loop tightens and tuning becomes faster.
The 31% end-to-end resolution rate for internal cases shows the AI isn't just triaging tickets - it's closing them. For support managers, that translates directly to reduced backlog and shorter wait times for the issues that still need human attention.
What the confidence score signals
A 2.6 out of 3 Resolution Confidence Score isn't a vanity metric. In production support environments, confidence scoring determines whether an AI response goes straight to the customer or gets routed for review. Scores in this range let operations teams set aggressive automation thresholds without spiking error rates. Smarsh's result suggests the model has enough training data and domain specificity to handle communications compliance inquiries with minimal hallucination risk.
Why this matters for customer support professionals
Deflection rate is the metric your leadership team will ask about first when evaluating AI support tools. Smarsh's 72% figure provides a real-world benchmark from a regulated industry - communications data and intelligence - where accuracy isn't optional. If you're building a business case for AI in your own support stack, pairing deflection data with time-saved-per-case numbers (7.5 hours here) gives you both the efficiency story and the capacity story in two data points.
For specialists looking to build skills around these tools, structured learning paths like AI for User Support Specialists cover the automation and agent-assist techniques that drive results like Smarsh's. Broader resources on AI for Customer Support can help teams evaluate which workflows benefit most from deflection-focused AI.
Your membership also unlocks:
