How should accounting teams use an AI API?

Key facts

API compatibility	OpenAI-compatible /v1/chat/completions; drop-in base URL change
Models	30+ models behind one API; open-weight options for private deployment
Typical patterns	Invoice extraction, ledger coding, reconciliation triage, policy Q&A
Integration path	Middleware connects the API to ERP, AP and practice-management systems
Pricing model	Flat monthly self-serve plans with unlimited fair-use usage; no per-token billing — see the live pricing page
Free tier	plugsky-micro and plugsky-lite on the free plan, no card required
Deployment	Plugsky cloud, your VPC, on-prem and air-gapped options
Identity and audit	SSO/SCIM, role-based access, SIEM audit log export and BYOK via KMS or HSM

TL;DR

• Start with extraction and retrieval, not autonomous posting or filing.
• Keep your OpenAI SDK code; change the base URL and model name.
• Run in the cloud, your VPC, on-prem or air-gapped as client contracts require.
• Scope keys per client and keep an audit trail for workpapers.
• Start free with plugsky-micro and plugsky-lite, no card required.

How it works, step by step

1. Pick one bounded workflow, such as invoice extraction or policy Q&A.
2. Classify the data in scope and confirm what may lawfully leave the firm's boundary.
3. Build a labelled sample of real documents with expected outputs and a reviewer.
4. Ground answers with embeddings and RAG over the firm's own templates and manuals.
5. Build against the OpenAI-compatible endpoint and keep prompts and evals in version control.
6. Choose the deployment tier and region your client contracts require.
7. Measure accuracy and reviewer time, then expand to the next workflow.

Try it yourself

Open the AI ROI calculator →

Where an AI API fits in accounting

Accounting is a document discipline, so the highest-value use cases are retrieval and extraction rather than open-ended generation. The API should read the source, propose a structured result, and leave approval to a qualified person:

• Invoice and receipt extraction: turn PDFs and scans into structured JSON for accounts payable, with low-confidence fields flagged for review.
• Ledger coding suggestions: propose account, tax and cost-centre codes from historical patterns, then let the accountant accept or override.
• Reconciliation triage: summarise why a bank line and ledger entry do not match so staff start from a hypothesis instead of a blank screen.
• Workpaper drafting: assemble checklists, queries and draft notes from engagement documents for the reviewer to edit.
• Policy Q&A: answer staff questions from the firm's own accounting manual, templates and prior-year memos with citations.

Security, privacy and data handling

Client financial records are confidential and often covered by professional-secrecy duties, so the API should process them inside a boundary the firm controls and can evidence:

• Exclude client content from model training in the contract and verify how the exclusion is implemented.
• Issue one project-scoped key per client or internal service so a leak is contained and revocable.
• Keep an audit trail of prompts, outputs and approvals so workpapers remain defensible.
• Redact bank details and identifiers the workflow does not need before text leaves your middleware.

Deployment options and model choice

Firms typically start in cloud for internal drafting and move to a private endpoint when client contracts require it. The same OpenAI-compatible API runs across Plugsky cloud, a private endpoint in your VPC, on-prem and air-gapped, with region-locked planes for residency. One key reaches 30+ models, including open-weight options for offline deployment, and migration is a base URL change. Chat, streaming, JSON mode, function calling, embeddings, RAG and agents are live; audio, images, moderation, files, batch, fine-tuning, assistants and the responses API remain coming soon.

From pilot to production

Most accounting pilots fail on data quality and review discipline rather than model quality. Watch for these:

• Rolling out extraction without a labelled accuracy baseline, so nobody can tell if it improves.
• Letting a draft journal entry post without a qualified reviewer's approval.
• Sending full client files when a redacted excerpt would answer the question.
• Forgetting retention rules for prompt logs that now contain client data.
• Treating a pilot on tidy digital PDFs as proof for poor-quality scans.

Start with one bounded workflow such as AP invoice extraction, score a labelled sample of real client documents, and expand only when accuracy and reviewer minutes are measured. Keep prompts and evaluation sets in version control.

Honest comparison

Capability	Plugsky	Typical per-token API	Building in-house
API compatibility	OpenAI-compatible chat, embeddings and tools	Usually compatible	Full rewrite
Deployment	Cloud, VPC, on-prem and air-gapped	Mostly cloud-only	You operate GPUs and serving
Data residency	Region selection and sovereign options	Limited regions	You control fully
Pricing	Flat monthly self-serve, fair-use usage	Per-token, harder to forecast	GPU plus operations cost
Model choice	30+ models behind one API	Varies by provider	You host every model
Industry fit	Invoice extraction, ledger coding, reconciliation triage, policy Q&A	Generic API, you adapt it	You build every workflow

Frequently asked questions

Can we keep our existing OpenAI SDK code?

Yes. Plugsky exposes an OpenAI-compatible API, so you change the base URL and model name and keep your SDK, prompts and middleware.

Is there a free plan?

Yes. The free plan includes two free models, plugsky-micro and plugsky-lite, and does not require a credit card.

How does pricing work?

Self-serve plans are flat monthly with unlimited fair-use usage; there is no per-token billing on self-serve. See the live pricing page for current plans.

Can we deploy in our own environment?

Yes. Plugsky supports a private endpoint in your VPC, on-prem and air-gapped deployment for enterprise setups, with region selection on cloud plans.

Would client data be used to train models?

Plugsky processes customer data to provide the service. Training exclusion belongs in your DPA; confirm the current terms with the enterprise team before sending client records.

Which endpoints are live today?

Chat completions, streaming, JSON mode, function calling, embeddings, RAG and agents are live. Audio, images, moderation, files, batch, fine-tuning, assistants and the responses API are coming soon.

How do we evaluate quality?

Build a labelled set of real documents with expected outputs, compare two or three models, track field-level accuracy and reviewer minutes, then pin the model version you approve.