How do you build AI agents for accounting?

Key facts

Agent stack	Chat completions, function calling and embeddings are live; you own the loop
Extraction	Structured JSON output for invoice, receipt and statement fields
Retrieval	plugsky-embed plus RAG over policy and prior-period records
Tools	Any internal API can be exposed via function calling or MCP
Models	30+ models; small aliases for extraction, stronger ones for reconciliation reasoning
Approval	Design a human gate before posting entries or changing ledger state
Pricing model	Flat monthly self-serve plans with unlimited fair-use usage
Roadmap	Assistants and batch endpoints are coming soon

TL;DR

• Extract first, reconcile second, post only after human approval.
• Give the agent tools for reading records, not for writing to the ledger.
• Log every tool call and cited document for the audit trail.
• Use small models for extraction and stronger ones only for exceptions.
• Track exception rate and reviewer minutes as the core metrics.

How it works, step by step

1. Map one close process and identify the extraction, matching and exception steps.
2. Expose read-only tools for documents, ledger records and policy retrieval.
3. Build an extraction agent that returns a strict JSON schema with source references.
4. Add a reconciliation step that compares extracted fields against ledger and bank records.
5. Route exceptions and any posting action to a human approver with the evidence attached.
6. Log tool calls, model version and approvals, then measure exception rate and reviewer time.

Try it yourself

Open the AI agent cost calculator →

Accounting use cases that work as agents

Agentic patterns fit processes with a clear evidence trail and a review step:

• Invoice and receipt extraction: turn PDFs into structured fields with a source reference per field.
• Reconciliation: match payments, invoices and bank lines, and list unmatched items with reasons.
• Close checklist: track which tasks are complete, chase missing documents and summarise status.
• Anomaly review: flag unusual amounts, duplicates or vendor changes for a controller to inspect.

The agent prepares; an accountant decides.

Architecture of a safe accounting agent

Keep the loop small and the permissions tight:

• Retriever: embeddings over policy, contracts and prior-period records for context.
• Extractor: strict JSON schema with per-field confidence and document references.
• Matcher: deterministic rules where possible, model reasoning for ambiguous cases.
• Approval gate: a human confirms before any ledger write or payment release.
• Logger: records tool calls, inputs, model version and the approver identity.

Never give the agent write access to the ledger; it should propose actions through an approval queue.

Guardrails and audit

Segregation of duties still applies. The agent should not be able to create a vendor, approve its own exception and post an entry in one path. Split capabilities across tools, require an approver for any state change, and cap monetary thresholds so large items always escalate. Log every request and tool result so an auditor can reconstruct what happened, and keep the model version pinned for reproducible results. Where policy or client contracts require it, run the workload in your VPC, on-prem or air-gapped while keeping the same compatible API.

Evaluation and rollout

Build a labelled set of documents and matched transactions, then measure field-level extraction accuracy, match precision, exception rate and reviewer minutes per item. Prototype with plugsky-micro or plugsky-lite, and use the 14-day full-access trial to test whether a stronger model reduces exceptions enough to justify the switch. Self-serve pricing is flat monthly with fair-use usage, so month-end peaks stay predictable. Roll out one process at a time, keep prompts versioned, and re-run the evaluation whenever a form or template changes.

Honest comparison

Capability	Plugsky	Typical agent framework	Building in-house
Tool calling	Live function calling over any HTTP API or MCP tool	Framework-specific adapters	You build and maintain
Retrieval	plugsky-embed plus RAG over accounting documents	Add-on or external store	You integrate everything
Approval gate	You own the gate; the API stays read-only by design	Varies	You design controls
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

Frequently asked questions

Can the agent post entries to the ledger?

No. Give it read-only tools and an approval queue; a human should confirm any ledger write, payment release or vendor change.

Which Plugsky endpoints do accounting agents use?

Chat completions, function calling, JSON mode and embeddings are live. Assistants and batch endpoints are coming soon.

How do we keep extraction accurate?

Use a strict JSON schema with per-field references, sample-check results against the source documents and re-test whenever a form template changes.

Is there a free plan?

Yes. The free plan includes plugsky-micro and plugsky-lite with no credit card, which is enough to prototype extraction and matching.

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.

Can we run it in our own environment?

Yes. VPC, on-prem and air-gapped deployments are available with the same compatible API, so the agent code does not change.

How should we measure success?

Track field-level extraction accuracy, exception rate and reviewer minutes per item. Those metrics show whether the agent is reducing work or moving it.