Live ops24/7 supportPeak concurrency
Gaming
Player support that survives a launch weekend
A live-service studio routes player chat through the gateway. Routine password and billing questions never reach a human, while suspected account compromise is escalated within seconds.
- Workload
- Player support, NPC dialogue, live moderation
- Routing
- Support intent → Customer Service AI; risk → Risk Engine → human agent
- Constraint
- Traffic spikes 9× at patch release
Result — 78% of player tickets resolved without a human, and median first response fell from 4 minutes to 11 seconds.
RegulatedAudit trailRead-only scopes
Finance
Research summarisation with an auditable trail
Analysts summarise filings and market research through the gateway. Every request is attributable to a named person, and models with external egress are unavailable to regulated workloads.
- Workload
- Research summarisation, KYC review, internal Q&A
- Routing
- Confidential class → Private LLM; public research → frontier models
- Constraint
- MAS technology risk guidelines, 7-year retention
Result — Audit preparation went from three weeks of log collection to a single console export.
PHI redactionIn-regionClinical
Healthcare
Clinical documentation that never leaves the region
Clinicians dictate notes that are structured by AI. Detected identifiers are stripped at the gateway before any request leaves the hospital network.
- Workload
- Clinical documentation, discharge summaries, coding support
- Routing
- All PHI → Private LLM in-region; de-identified research → frontier
- Constraint
- No PHI may leave asia-southeast-1
Result — Documentation time per encounter down 32%, with zero PHI egress events across 1.4M requests.
MultilingualEdgeMaintenance
Manufacturing
Maintenance copilots on five factory floors
Line technicians query equipment manuals in their own language. The gateway routes translation to whichever model performs best for that language pair.
- Workload
- Maintenance copilots, defect triage, line instructions
- Routing
- Korean → Claude; Japanese → GPT; Bahasa → Gemini
- Constraint
- Intermittent connectivity at the edge
Result — Unplanned downtime reduced 18% in the first quarter after rollout.
Per-cohort budgetsGuardrailsScale
Education
Tutoring assistants with hard budget ceilings
Each faculty gets a monthly cap. When a cohort approaches its ceiling the gateway throttles non-urgent traffic instead of issuing a surprise invoice.
- Workload
- Tutoring, grading support, curriculum drafting
- Routing
- Economy tier by default; frontier only on escalation
- Constraint
- Fixed annual budget across 12 faculties
Result — Delivered the full academic year 9% under the allocated AI budget.
DeflectionAgent assistCaching
Customer Service
Deflection across chat, email and voice
Repeat questions are answered from the semantic cache at near-zero cost. Anything ambiguous is handed to an agent with a drafted reply already prepared.
- Workload
- Deflection, agent assist, quality review
- Routing
- Cache → Customer Service AI → human handoff
- Constraint
- Cost per contact must fall year on year
Result — 41% cache hit rate, cutting cost per contact by roughly USD 1,940 per month.
Code reviewAgentsCost tiering
Software Development
Agent workloads routed by task complexity
Trivial refactors go to an economy model; architectural review goes to a frontier model. Engineering sees the cost of each choice in the console.
- Workload
- Code review, refactoring, agent pipelines
- Routing
- Complexity score decides tier; secrets scanned before egress
- Constraint
- No source code may reach a model without scanning
Result — Review turnaround down 46% while spend stayed flat against the prior quarter.