An agent demonstration built around triage, enrichment and routing — with escalation treated as a feature rather than a failure.
- Client
- Relay
- Industry
- Logistics & Operations
- Year
- 2026
- Services
- AI Agents & Workflow AutomationSoftware & Product EngineeringConversational AI
Overview
What this project is.
Relay models an operations team receiving a high volume of unstructured inbound requests. It demonstrates how we scope, bound and supervise agent automation so it can be left running without becoming a source of new problems.
Challenge
The problem to solve.
Open-ended agents are difficult to trust and harder to debug. Given broad autonomy they take unexpected routes, and when something goes wrong there is no clear record of why. Operations teams will not adopt automation they cannot audit.
Insight
What we noticed.
The reliable pattern is narrow. An agent with a defined remit, a small set of tools, explicit limits and a clean handover is genuinely useful. Autonomy is not the goal — removing the mechanical handling between two judgement calls is.
Strategy
The approach taken.
Map the process as it actually runs, including the exceptions people handle without noticing. Give the agent a bounded remit and a specific toolset, log every decision with its reasoning, and route anything below a confidence threshold to a person with the context already gathered. Run it alongside the manual process before widening it.
Creative Direction
How it looks and moves.
Interface design for supervision rather than novelty. A queue view showing what was handled and what was escalated, decision trails readable at a glance, and clear visual separation between what the system did automatically and what needs a human.
Execution
What the work involved.
Process mapping including exception paths
Agent design with an explicit tool boundary
Integration with CRM, mail and ticketing systems
Confidence thresholds driving automatic escalation
Full decision logging with reasoning retained
Supervisor queue for review and override
Deliverables
- Process map
- Agent service
- System integrations
- Supervisor interface
- Audit logging
- Monitoring & alerting
Stack
- Python
- FastAPI
- LangChain
- PostgreSQL
- Redis
- React
- Docker
What it sets out to achieve
The outcome the work is built for.
This is a capability demonstration, so the outcomes below describe design intent rather than measured performance. We publish numbers only where a client has verified and approved them.
Bounded by design
A narrow remit and explicit tool set make behaviour predictable and debuggable.
Escalation as a feature
Low-confidence cases reach a person with the context already assembled.
Auditable end to end
Every action carries a decision trail, which is what makes it adoptable in operations.
