%20--%3e%3cstyle%20type='text/css'%3e.cls-0%20{fill:%23FF9900;}%3c/style%3e%3cpath%20class='cls-0'%20d='m117.3%2084h-14.1l39.2-79h-24.8l-20.3%2041.7-20.5-41.7h-3.8l-20.3%2041.7-20.8-41.7h-24.3l41%2083.3h8.2l18.2-37.7%2018.3%2037.8h23.7v2.8h10.7v-10.5h-10.6v3.3h0.2zm3.3-74.7h15.2l-28.9%2057.1-7.2-14.9%2020.6-42.2h0.3zm-69.3%2074.9-37-75h15.1l20.8%2042.4-5.8%2011%204.1%201.6%2016.9-33.9%207.2%2015.7-18.4%2038.1-2.9%200.1zm44.5-0.1-28.2-58.9%207.4-14.8%2029.6%2060.8-6.1%2012.9c0%200.1-2.7%200-2.7%200zm27.9%200.6v2.8h-3.1v-2.8h3.1z'/%3e%3c/svg%3e)
73% ticket deflection from a RAG support agent, live in four weeks
Support automation usually fails in one of two ways. Either it answers confidently from the wrong source and erodes trust faster than it saves time, or it is hedged so heavily that customers route around it and the deflection rate never moves.
- 73% ticket deflection
- RAG-powered support agent across 12 product lines
- Four-week rollout, running in production
What made it hard
Support automation usually fails in one of two ways. Either it answers confidently from the wrong source and erodes trust faster than it saves time, or it is hedged so heavily that customers route around it and the deflection rate never moves.
Twelve product lines makes that sharper: an answer correct for one product can be actively wrong for another, so retrieval has to respect which product the customer is actually asking about.
How it was done
Retrieval before generation
The difficulty in a RAG system is almost never the model. It is whether the right passage is retrieved, scoped to the right product line, before the model is asked anything.
Evaluated against real questions
Quality was measured against questions customers actually ask, not a curated set. A demo passes on the curated set; production does not.
Handover as a feature
Knowing when to stop and pass to a human is part of the product. An agent that cannot decline is the one that produces the confidently wrong answer.
Shipped in four weeks
Delivered as a working system rather than a pilot that needed a second project to make it real.
Where it landed
73% of tickets deflected across 12 product lines, running in production after a four-week rollout.
Deflection at that level only holds when the answers are trusted; the number reflects retrieval quality more than model choice.
Bring us the version of this you are facing
A 45-minute call with the engineer who would run it. We will tell you what we would look at first, and whether it is a project at all.
Talk to an Engineer