Case study · AI & Automation

An AI support agent that resolves 72% of tickets on its own

HelpDeck's support team was drowning in repetitive tickets. We built an AI agent that reads, resolves, and escalates with guardrails, so humans only touch the cases that need judgment.

Client

HelpDeck

Industry

B2B SaaS

Timeline

8 weeks

Year

2026

0%

tickets resolved end to end

0%

answer accuracy on eval set

0min

median first response time

01 · The challenge

Where things stood

HelpDeck's five-person support team was handling 3,000+ tickets a month, and 70% of them were variations of the same forty questions. Response times had slipped past nine hours, churn surveys started mentioning support speed, and hiring more agents only bought a few months.

Previous chatbot attempts had failed badly. The off-the-shelf bot invented refund policies that did not exist, and the team turned it off within two weeks. Trust in AI inside the company was near zero.

02 · The solution

What we built

We built a retrieval-grounded support agent that answers only from HelpDeck's verified knowledge base, with citations attached to every reply. If confidence drops below a threshold, or the topic touches billing and refunds, the agent drafts a reply and routes it to a human queue instead of sending.

Every action is logged, every answer is scored against a 250-question evaluation set we built with their team, and a weekly quality report goes to the head of support. The agent got smarter and the team could prove it.

03 · The approach

How we did it

01

Audit the ticket history

We clustered 12 months of tickets to find what actually gets asked, then built the evaluation set from real questions with verified answers.

02

Ground it in real docs

Help center articles, internal runbooks, and policy pages were chunked, embedded, and indexed with a reranking step for accuracy.

03

Wrap it in guardrails

Scoped tools, confidence thresholds, forbidden topics, and a human escalation queue with full context attached to every handoff.

04

Ship, measure, tune

Two weeks in shadow mode answering silently alongside humans, then a gradual rollout with dashboards for resolution rate and accuracy.

04 · The stack

Built with

Chosen for this project, not from a default template. Every build gets the stack its problem deserves.

OpenAILangGraphPineconeNext.jsFastAPIPostgreSQLRedisAWS
Our last chatbot invented refund policies. This one cites its sources, knows when to hand off, and my team finally trusts an AI tool.
James CarterHead of Support, HelpDeck

Want results like this?

Tell us what you are building. We will reply within one business day with an honest take on scope, timeline, and cost.