Automating customer returns processing with an extraction agent, from 8 minutes to 40 seconds
A mid-market UK home & garden retailer (engagement 2024) processing ~15,000 returns/month, each requiring an agent to read the customer's reason, look up order history, check return eligibility against 12 policy rules, decide refund vs. exchange, and initiate logistics. Straightforward cases took ~8 minutes. The 22-agent team was at capacity with a 30% volume-growth forecast looming. Discovery & Blueprint categorised returns into 6 case types by complexity, 87% fell into the top three (straightforward enough for full automation), 13% required human judgment (fraud signals, policy exceptions, escalations). Our LangGraph extraction-and-decision agent reads the request, retrieves order context, applies the 12 policy rules deterministically (NOT via the LLM, auditable), and either completes autonomously or routes to a human with a pre-filled decision summary. 87% fully automated; processing time for those: ~40 seconds. Annual ops cost −$1.4M.
87%
Cases fully automated
8min→40s
Processing time
−$1.4M
Annual ops cost
The Challenge
15,000 returns per month. Each case required an agent to: read the customer's reason, look up order history, check return eligibility against 12 policy rules, decide on refund vs. exchange, and initiate the logistics process. Straightforward cases took 8 minutes. The team of 22 agents was at capacity with no headroom for the 30% volume growth forecast.
Our Approach
Discovery & Blueprint categorised returns into 6 case types by complexity. 87% fell into the top 3 types, straightforward enough for full automation. 13% required human judgment (fraud signals, policy exceptions, escalations).
We built a LangGraph-based extraction and decision agent that reads the customer request, retrieves order context, applies policy rules deterministically, and either completes the return autonomously or routes to a human with a pre-filled decision summary.
Outcome
87% of returns cases processed end-to-end with no human involvement. Processing time for automated cases: 40 seconds. Annual operational cost reduced by $1.4M. The 22-agent team now handles 13% of cases that genuinely require human judgment, and does it better because they're not exhausted by volume.
What We Learned
01
Categorise by case type before architecting, the automation opportunity is rarely uniform.
02
Deterministic policy rules are more reliable and auditable than asking the LLM to apply them.
03
The goal is not to replace all humans, it's to let humans focus on the cases that actually need them.
Stages Engaged
Feasibility Call
Discovery & Blueprint
Concept Validation
Production Build
Total Duration
5 months total
Artifacts Delivered
PRD
Agent Architecture
Policy Rules Specification
WBS
SOW
Start with a Feasibility Call
2 hours. No cost. We'll tell you honestly whether AI makes sense for your case.