How EnaOps uses AI

One engine, many modules

EnaOps is a platform, not a point tool. Each module is a set of rules and outputs on the same engine, so the same safeguards, evidence and audit trail apply everywhere.

• Live (beta): Self Assessment, Bookkeeping & bank statements, Snap a bill, Questions for you, Month-end close, Invoicing & payment chasing, Mileage & driver earnings, QuickBooks connect, Ask EnaOps in ChatGPT & Claude.
• Coming soon: Work with an accountant.
• Later: AI receptionist & scheduling, Handover packs for trades.

The engine is not tied to one AI model or one kind of job, which is how the same approval layer will carry over to quotes, payment chasing and, later, calls and scheduling.

The engine: read, sort, check, draft, you approve

Every module runs the same five steps. The engine reads what you upload, sorts each item (for Self Assessment, into SA pages and boxes), checks for missing or conflicting evidence, and drafts the output with the working. The last step is always yours: nothing counts, is filed or is sent until you approve it.

What the AI reads and extracts

When you upload a document, a model classifies it (P60, P45, P11D, bank statement, invoice, receipt, dividend voucher, interest certificate, rental or mortgage statement, pension or student-loan statement, Gift Aid record, Child Benefit letter, CIS statement, SA302) and pulls out structured fields: employer reference, pay and tax, amounts and dates, bank lines with a suggested expense box.

Every extracted item carries a confidence score and points to where it came from — the page and the snippet of text.

Rules first, model second

Mapping to Self Assessment boxes is done by deterministic rules wherever possible (a P60 always maps to SA102 boxes 1 and 2). The model only fills gaps, such as suggesting which expense category a bank line belongs to. Tax rates, allowances and thresholds come from versioned tables keyed to gov.uk sources — never from the model.

The human approval layer

Governed, not autonomous. Nothing counts until you approve it. Each proposal lands in your in-tray where you can approve, edit, reject or undo. Your figures only ever use approved lines. Every change is logged, and undo restores the previous value.

The same layer applies to every module as it arrives: reminders and payment chasers will only go out on your say-so, the ChatGPT & Claude connector asks you to confirm any change it proposes, and the AI receptionist (later) will act only within rules you set.

ChatGPT & Claude connector (MCP)

You can connect EnaOps to your own ChatGPT or Claude as a connector (the open Model Context Protocol) and ask about your account in that chat: “what VAT do I owe this quarter?”, “show my unreviewed bills”, “how much am I owed?”. Set it up in Settings → Connect to ChatGPT / Claude.

• You sign in and choose. The assistant gets access only after you sign in to EnaOps and allow it: reading your figures, and (if you tick it) making changes.
• Changes need your yes. Approving lines, changing a category, creating a supplier rule or marking an invoice paid is shown to you first and only happens when you confirm in the chat. Each change is logged against the connector and can be undone here.
• Never filing or sending. The connector cannot file with HMRC, email customers or move money.
• Where your data goes. What the assistant reads from EnaOps (figures, line descriptions, invoice and document names) is sent to the AI provider you chose — OpenAI or Anthropic — and handled under your account and settings with them, not under EnaOps’ agreements. Check their data controls (for example whether chats are used for training) before connecting.
• Disconnect any time in your assistant’s connector settings, or ask us to revoke access.

No auto-filing, no HMRC contact

EnaOps does not file your return and does not talk to HMRC. It prepares a pack you can file from on gov.uk or hand to an accountant.

Cited evidence and working

Every figure in the draft shows “where this came from”: a document and page, or a rule with a link to the relevant gov.uk guidance. The calculation is shown step by step so you — or your accountant — can check it.

Model and provider

Extraction runs on a managed large-language-model service hosted in a UK or EU region under a business agreement that excludes use of your data for training. The specific provider and model version are recorded on each extraction and will be named here before public launch.

Your data is not used to train models

Documents and extracted figures are processed only to serve your account. We do not use them to train or fine-tune models, and our providers are contractually prohibited from doing so. See the privacy notice.

Known limits

• Low-confidence or out-of-scope items (capital gains, foreign income, partnerships) are flagged “ask an accountant”.
• CIS subcontractor deductions are captured but not yet used in the estimate (coming soon).
• Scottish rate bands are applied where you tell us you are a Scottish taxpayer; we do not determine residency for you.