CASE STUDY · AGENTIC AI
From ticket queues to resolution: a multi-agent support system for an online marketplace.
Anonymised · Sector: Consumer marketplace (buyers and sellers) · Solution partner: Brandsmashers Tech
- 50–60%Target share of routine requests resolved end to end, after 6–9 months
- < 1 minTarget first response for routine requests, down from hours
01 · PROJECT OVERVIEW
Support that resolves, not just responds.
A fast-growing online marketplace connects independent sellers with consumers across many categories. It handles thousands of orders, returns and seller interactions every day, and customer trust depends on how quickly and fairly problems get resolved.
As part of a customer experience modernisation initiative, the marketplace set out to build a support model that could scale with order volume without scaling headcount at the same rate.
Brandsmashers Tech was engaged to design and build a multi-agent support system.
- Designing the agent architecture and orchestration logic
- Integrating the agents with order, payment, logistics and help desk systems
- Establishing guardrails, audit trails and human escalation paths
- Building evaluation and quality assurance for agent behaviour before every release
- 1Resolving routine customer issues end to end, not just answering questions.
- 2Unifying order, payment, logistics and seller data behind one support experience.
- 3Keeping human agents focused on complex, high-empathy cases.
02 · THE CHALLENGE
When volume outgrows the support model.
As order volumes grew, the support function reached a point where response times, repeat contacts and agent workload all affected customer trust and seller satisfaction.
- PROBLEM 01Long waits and repeated explanations
Customers waited hours for a first reply, then had to repeat their story whenever a case moved between agents. Each handoff added delay and frustration.
- PROBLEM 02Routine requests that needed many systems
“Where is my order”, refund status or an address change meant checking three or four separate tools. The work was repetitive but too varied for a simple script.
- PROBLEM 03A chatbot that answered but could not act
An earlier FAQ chatbot provided information but could not take action, so customers still reached for a human and the real workload barely moved.
- PROBLEM 04Seller disputes and slow coordination
Issues involving a seller relied on email chains that could stretch over days, leaving customers without updates.
- PROBLEM 05Peak-season pressure
Seasonal spikes forced rapid hiring and training cycles that were costly and hard to sustain, while quality varied from agent to agent.
The problem was not a lack of answers. It was a lack of action across systems.
03 · THE AGENTS
One supervisor, five specialists.
A conceptual architecture: a supervisor plans the work and specialist agents act within narrow, auditable permissions.
Understands intent, plans the steps, assigns work and decides when to escalate.
- Order and logistics agent
Reads order status and carrier events, and detects delays and failed deliveries.
- Resolution agent
Applies policy to propose remedies such as reship, refund or credit, within preset value limits.
- Seller coordination agent
Contacts sellers, collects evidence and tracks response deadlines.
- Communication agent
Writes clear, on-brand updates in the customer’s channel and language.
- Quality and safety agent
Checks every proposed action against policy and tone rules before it executes.
CHECKS BEFORE ANYTHING RUNS
04 · THE APPROACH
Resolve the routine, keep people in control of the rest.
A balanced strategy: resolve routine issues autonomously, keep people in control of sensitive decisions, and measure real resolution rather than deflection.
- AOrchestrated multi-agent architecture
- A supervisor agent to understand intent, plan the steps and decide when to escalate
- Specialist agents for order and logistics, resolution, seller coordination and customer communication
- A quality and safety agent that checks every proposed action against policy and tone rules before it runs
IMPACTOne coordinated system working across tools, replacing manual hopping between them.
- BAction-capable integrations
- Secure connections to order management, payment, logistics and help desk platforms
- A unified, near-real-time customer and order context shared by all agents
- Remedies such as reship, refund or credit executed within preset value limits
IMPACTCustomers receive a real status or a real fix, not only information.
- CGuardrails and human-in-the-loop design
- Refunds above a threshold, fraud, safety and distress cases routed to a human with a full summary
- Every agent action logged for audit
- A person reachable at any point in the conversation
IMPACTFaster service without losing control, accountability or the human touch.
- DEvaluation and continuous improvement
- A library of realistic test conversations and policy-compliance checks run before each release
- Adversarial testing for edge cases and unusual requests
- Dashboards tracking resolution, escalation, cost and latency
IMPACTPredictable quality that improves with each release.
- 1→Customer message
- 2→Supervisor plans
- 3→Specialists act
- 4→Safety check
- 5→Systems updated
- 6Customer answered
05 · RESULTS
What changes.
Routine issues get a real answer or a real fix in one conversation, and the people on the team spend their time where judgement and empathy matter.
| MOMENT | BEFORE | AFTER |
|---|---|---|
| Customer reports a late order | Waits in a queue and explains from scratch | An agent recognises the order, checks the carrier and replies with a real status in one message |
| Remedy needed | An agent manually checks policy and the seller | The resolution agent proposes and, within limits, executes the remedy |
| Seller involvement | Email chains over days | The seller coordination agent runs a time-boxed workflow and keeps the customer informed |
| Complex or upset customer | Same queue as everyone else | Flagged early and routed to a senior human with full context |
Projected targets, not measured results. Targets are benchmark-based and set well below Gartner’s 2029 forecast of 80% autonomous resolution for common issues.
- 50–60%Routine requests resolved end to end without a humanBenchmark-based projection, well below Gartner’s 2029 forecast of 80% for common issues.
- < 1 minFirst response time for routine requestsFrom hours to under a minute.
- 20–30% lowerNet support cost per resolved issueIn line with the 20–35% range industry roundups describe as realistic.
- FewerRepeat contacts on the same issueCustomers get a final answer, not a deflection.
- RefocusedHuman agents’ timeShifted to complex, high-empathy cases.
- Resolution, not deflection
Customers leave with a status or a fix, so they don’t come back with the same issue.
- Faster first response
Routine requests are answered in under a minute instead of hours.
- Lower cost per resolution
Agents take on the repetitive work of checking three or four systems per request.
- A foundation that scales
A scalable support foundation for future growth, with less reliance on peak-season hiring.
06 · DELIVERABLES
How Brandsmashers built it.
- Agentic AI developmentThe supervisor and specialist agents, with tool permissions and escalation logic.
- API and system integrationsSecure connections to order management, payments, logistics and the help desk.
- Data engineeringA unified, near-real-time customer and order context.
- Quality assessmentReal-style test conversations, policy-compliance checks and adversarial testing before each release.
- Cloud runtimeMonitored, scalable deployment with cost and latency dashboards.
- AI
- Multi-agent orchestrationTool callingPolicy guardrails
- Integrations
- Order managementPaymentsLogisticsHelp desk
- Data
- Unified customer and order contextAudit logs
- Quality
- Test conversation libraryAdversarial testingRelease gates
- Operations
- Cloud runtimeCost dashboardsLatency dashboards
TAKEAWAYS
Lessons for consumer-facing leaders.
- 01Measure resolution, not deflection.
- 02Start with three or four high-volume intents and expand.
- 03Give agents narrow, auditable permissions.
- 04Treat the human handoff as a designed feature.
- E-commerce
- Retail
- Marketplaces
- Subscription businesses
- Consumer apps
YOUR TURN
Want support that resolves, not just responds?
Brandsmashers Tech designs and builds agentic AI systems with the guardrails, integrations and evaluation they need to run in production. Tell us which intents cost you most and we’ll show you where agents fit.