Key facts
| Agent runtime | Function calling loop on /v1/chat/completions (live) |
| Tool pattern | Typed JSON schemas for ledger, reporting and market-data lookups |
| Model routing | plugsky-micro and plugsky-lite for extraction; plugsky-pro and plugsky-frontier for analysis |
| Models | 30+ models behind one OpenAI-compatible endpoint |
| Controls | Scoped API keys, RBAC and SSO/SCIM; approval gates in your application |
| Audit | Tool calls, key lifecycle and inference metadata exportable to SIEM |
| Deployment | Region-locked planes, VPC, on-prem and air-gapped options |
| Roadmap endpoints | Assistants-style managed endpoints are coming soon; build the loop yourself today |
TL;DR
- Never let the model do arithmetic — call a calculator or query tool and let it narrate.
- Every figure in a draft must trace to a tool result your reviewers can replay.
- Keep posting, approvals and payments outside the agent behind human gates.
- Route data extraction to small models and variance analysis to frontier models.
- Log every turn so month-end output can be reconstructed and audited.
How it works, step by step
- Pick one bounded workflow — variance commentary, invoice coding or revenue reconciliation — and list its inputs, outputs and forbidden actions.
- Expose ERP, reporting and market-data systems as read-only tools with typed schemas and hard timeouts.
- Implement the loop: send the question with a tools array, execute each tool_call, append results, and stop when the model returns a draft.
- Move all arithmetic into tool code; the model may only cite computed values, never derive them itself.
- Add an approval gate where an analyst reviews the draft, edits it, and releases it into the reporting workflow.
- Log tool arguments, results and model metadata per turn, and export to your SIEM or data warehouse.
- Evaluate on historical periods with known outcomes before adding a second workflow.
Try it yourself
Tools, not mental arithmetic
The failure mode that kills finance pilots is confident, wrong numbers. Fix it architecturally: the agent never computes. Variance, growth rates, FX conversion and aggregations happen in tools that query the ledger or run deterministic code, and the model's job is to choose the tool, pass correct arguments, and write commentary around results it was given.
Plugsky's live function calling makes this a standard loop — pass a tools array on /v1/chat/completions, execute the returned tool_calls in your service, append results and call again. Because the API is OpenAI-compatible, the tools run inside your security boundary with the credentials and row-level access rules you already enforce.
Controls for regulated finance workflows
Finance reviewers need provenance, not just accuracy. Every draft should carry the tool calls that produced it so an analyst can replay the query and see the same numbers. Treat the agent as a drafting assistant: it prepares commentary, coding suggestions and reconciliations, while approvals, journal entries and disclosures stay with people.
- Identity: service keys scoped per workflow, RBAC for reviewers, SSO/SCIM for workforce access.
- Least data: pass only the periods and cost centres needed; keep PII and bank details out of prompts.
- Evidence: tool arguments, result summaries and model metadata logged per turn and exportable to SIEM.
- Residency: pin the workspace to a region-locked plane or deploy to VPC, on-prem or air-gapped.
Models, routing and evaluation
Route by task. Extraction of line items from documents suits plugsky-micro or plugsky-lite; variance narratives and multi-source synthesis suit plugsky-pro or plugsky-frontier. With 30+ models behind one endpoint, this routing is a model-string decision you can tune without re-integrating.
Evaluate on closed periods where the correct answer is known: numeric accuracy of every cited figure, completeness against analyst checklists, and refusal behaviour when source data is missing. Track edit distance between agent drafts and final analyst versions as a practical quality signal. Start with one workflow and one team, then expand only when replayable audits stay clean.
Honest comparison
| Concern | Plugsky finance agents | Spreadsheet macros | Managed vertical AI |
|---|---|---|---|
| Data access | Typed read-only tools into your systems | Local exports | Vendor connectors |
| Arithmetic | Deterministic tool code | Formula-driven | Varies |
| Review model | Draft, then human approval | Manual review | Vendor workflow |
| Audit | Per-turn tool and model logs exportable to SIEM | Cell-level history | Vendor-controlled logs |
| Residency | Region-locked planes, VPC, on-prem, air-gapped | On premises by nature | Usually cloud-only |
Frequently asked questions
Can the agent post journal entries?
No. It drafts and recommends; posting stays in your ERP behind existing approvals. Keeping write access out of the agent is the simplest control you can adopt.
How do we stop wrong numbers?
Force all arithmetic through tools and require every figure in a draft to reference a tool result. The model narrates computed values rather than deriving them.
Is function calling reliable enough for finance?
Function calling is live on Plugsky, but design for failure: validate tool arguments, reject malformed calls, and treat an unverifiable answer as a failed run rather than a draft.
What data goes into prompts?
Only what each step requires — periods, cost centres and line-item summaries. Redact bank details and personal data, and set prompt retention to match policy.
Can this run in our own environment?
Yes. Plugsky supports region-locked cloud planes, VPC, on-prem and air-gapped deployments, so ledger data can stay within your network.
How do we evaluate before rollout?
Replay closed periods, compare every cited figure against the ledger, score completeness against the analyst checklist, and track how much analysts edit each draft.
Where should a pilot start?
Variance commentary and invoice coding are read-heavy, bounded and easy to audit. Avoid anything that touches payments or external reporting while you build evidence.