agent-side retrieval. a tool that sits behind the support agent, not in front of the customer. the agent asks a plain-english question and gets a cited answer in one query. no customer-facing bot. the agent stays in the ticket and sends the reply.
we shipped one for a mid-market e-commerce cx team. first-response time on complex tickets dropped 40%. this post is the receipt: what we built, what moved, and what didn't.
the client is anonymized. the 40% is the delivered outcome from the engagement, grounded in the customer-support work we describe on /ai-for/customer-support. one more thing up front: this was agent-assist only. no chatbot talked to a customer at any point.
#what was slowing the agents down?
the answers existed. finding them didn't scale. a complex ticket meant an agent opening zendesk history, a product doc, and a shipping-policy pdf, then stitching the answer by hand. four years of past tickets held the precedent. nobody could search four years of tickets in the time a customer waits.
so the first reply sat in a queue while the agent hunted. that wait is the number the support lead gets asked about. first-response time, not ticket volume.
can ai reduce first-response time in customer support without a chatbot?
yes. an internal retrieval tool lets agents ask plain-english questions across historical tickets, product docs, and policy pdfs and get one cited answer. the agent still writes and sends the reply. for one mid-market e-commerce cx team this cut first-response time on complex tickets 40%, with no customer-facing bot.
#what did we actually ship?
one internal retrieval tool over three sources: 4 years of historical tickets, the product docs, and the shipping-policy pdfs. an agent types a question the way they'd ask a senior teammate. the answer comes back with the source cited, so the agent can check it before it goes to a customer.
- search across 4 years of past tickets in one query, not tab by tab.
- cited answers, so the agent verifies the source instead of trusting a black box.
- answers grounded in the shipping-policy pdfs, where the edge-case rules actually live.
- the agent stays in the ticket. the tool never replies to a customer on its own.
training the support team was part of the build, not a handoff at the end. the kpi we agreed to move was first-response time on complex tickets. we measured it before and after. it fell 40%.
#isn't this the same as a deflection bot?
no, and the difference is the whole point. a deflection play tries to lower ticket volume by answering customers directly. this moved a different number. the metric was first-response time, delivered by putting retrieval behind the agent. the customer still talked to a person.
we are
stennir builds ai that sits behind your support agents: retrieval, triage, draft-and-review, measured against first-response time and resolution quality.
we aren't
stennir is not a customer-facing chatbot vendor and does not sell deflection as the goal; the agent stays in the ticket and your customers keep talking to humans.
tools like zendesk ai, intercom fin, or gorgias suggestions work inside their own walls: your macros, your help articles, your saved replies. we built retrieval across the rest. the historical tickets, the product docs, the shipping-policy pdfs, the runbooks. that's where the complex-ticket answers hide.
the kpis we move are first-response time and resolution quality on complex tickets, not headcount. we say that on day one and we measure it.
#would this work for our support team?
the fit test is simple. do your answers already live in places that are slow to search? four years of tickets, a stack of policy pdfs, product docs that shift. if agents lose minutes hunting before every complex reply, that's the minute retrieval gives back.
this build lives in our consultancy work, with the support-team training folded in from our training side. if you want to see how the pieces fit, the journal has more of these receipts.
if your agents are hunting for answers before every complex reply, that's the 40% we went after here. tell us what you're building and we'll tell you whether retrieval is the lever, or whether it isn't.