Key facts
| Agent stack | Chat completions, function calling and embeddings are live |
| Use cases | KYC document review, alert triage, credit memo drafting, customer service support |
| Human decisions | Analysts and officers remain accountable for clearances, alerts and credit |
| Tools | Read-only access to document stores and case systems |
| Deployment | Plugsky cloud, your VPC, on-prem and air-gapped options |
| Models | 30+ models; small aliases for classification, stronger ones for synthesis |
| Pricing model | Flat monthly self-serve plans with unlimited fair-use usage |
| Roadmap | Batch, fine-tuning and assistants endpoints are coming soon |
TL;DR
- Use agents to assemble evidence, not to make determinations.
- Keep read-only tools and human gates for KYC, alerts and credit.
- Log prompts, sources, model versions and reviewer identity for every case.
- Prefer VPC, on-prem or air-gapped deployment for customer data.
- Measure analyst time saved and false-positive handling, not automation rate.
How it works, step by step
- Pick one process such as alert triage or KYC document review.
- Expose read-only tools for documents, case systems and policy retrieval.
- Ground every summary in retrieved passages and require citations.
- Add a structured output schema so analysts review fields, not free text.
- Route decisions and any customer-impacting action to a human reviewer.
- Log everything, then measure time saved, quality and escalation rate.
Try it yourself
Banking workflows suited to agents
Banking has high-volume, evidence-heavy review work:
- KYC document review: extract and organise entity, ownership and identification fields for an analyst to verify.
- Alert triage: enrich transaction alerts with customer context and prior history, and rank them for an investigator.
- Credit memo drafting: assemble financial spreads, covenant notes and risk factors from approved sources.
- Customer service support: retrieve policy and product answers with citations for an agent to confirm.
None of these make a determination; they prepare one.
Architecture for controlled decisioning
Design for auditability first:
- Retriever: embeddings over policy, procedures and case history.
- Extractor: strict schema for document fields with per-field references.
- Read-only tools: case lookup, customer context and document search.
- Ranker: priority scores for alerts and reviews with reasons attached.
- Human gate: an analyst or officer confirms every clearance, alert and credit decision.
Never expose write tools for account state, limits or payments, and log each tool call with its inputs and outputs.
Data, model risk and deployment
Customer data, transaction records and credit information are highly sensitive, and banks usually have model risk and outsourcing requirements of their own. Keep identifiers out of prompts where references will do, choose the processing region deliberately, and set retention to match records policy. Use scoped keys, RBAC, SSO and audit logs, and pin model versions so behaviour is reproducible for review. Where the risk function prefers isolation, deploy in a VPC, on-prem or air-gapped environment; the compatible API means agent code stays unchanged and your evaluations remain valid.
Pilot and measurement
Prototype on plugsky-micro or plugsky-lite with sanitised cases, then use the 14-day full-access trial to test a frontier model on the hardest reviews. Track analyst minutes per case, extraction accuracy, alert prioritisation quality and escalation rate, and compare against the current process rather than an absolute target. Self-serve plans are flat monthly with fair-use usage, so volume spikes do not create variable bills. Start with one team, keep prompts versioned, and re-run the evaluation before extending to adjacent processes.
Honest comparison
| Capability | Plugsky | Typical agent framework | Building in-house |
|---|---|---|---|
| Tool calling | Live function calling over read-only internal APIs | Framework-specific adapters | You build and maintain |
| Audit | You log prompts, sources, model version and reviewer | Partial or add-on | You build all of it |
| Deployment | Cloud, VPC, on-prem and air-gapped | Usually cloud-only | You operate GPUs and serving |
| Pricing | Flat monthly self-serve, fair-use usage | Varies by framework and model | GPU plus operations cost |
| Model choice | 30+ models behind one API | Varies by provider | You host every model |
Frequently asked questions
Can the agent approve KYC or credit decisions?
No. It prepares evidence and drafts; an analyst or officer remains accountable for clearances, alert closures and credit decisions.
Which endpoints do banking agents use?
Chat completions, function calling, JSON mode and embeddings are live. Batch, fine-tuning and assistants endpoints are coming soon.
How do we keep data inside the bank?
Deploy in your VPC, on-prem or air-gapped. The compatible API means the agent code, prompts and tests do not change.
Is there a free plan?
Yes. The free plan includes plugsky-micro and plugsky-lite with no credit card, which suits a sanitised prototype.
How does pricing work?
Self-serve plans are flat monthly with unlimited fair-use usage. See the live pricing page for current plans and enterprise options.
How should we handle model risk requirements?
Pin model versions, log prompts and sources, keep evaluations reproducible, and treat the agent as an assistant to a documented human process.
How do we measure value?
Track analyst minutes per case, extraction accuracy, prioritisation quality and escalation rate against the current process.