Accounting SOP Learning Companion

Inspiration

Real accounting and process documentation often runs into hundreds — sometimes 1,000+ — pages, and is rarely read cover to cover by anyone, including the people whose job depends on following it.

At the same time, I noticed something in a recent conversation among Global Capability Centre professionals: most agreed that AI still isn't especially effective on the finance side. My own view is different — finance isn't like a purely automatable engineering discipline. It starts with real-world, external stakeholders (couriers, small vendors, non-finance staff) feeding data into systems, often incorrectly, and accounting principles (IFRS, GST, etc.) are universal, but how each organization actually implements them — which system, which approval chain, which local variation — is highly customized, even within the same industry.

That's the real gap: an experienced finance professional with 15–20+ years of accounting knowledge can still struggle at a new organization, not because they don't understand accounting, but because they don't yet know that organization's specific process and tools. As an ICAI examiner (I evaluated 254 CA Intermediate GST scripts in the May 2026 cycle) and a mentor to CA Final students, I see this exact gap repeatedly — the mistakes aren't about not knowing the principle, they're about not knowing the organization-specific implementation of it.

What it does

The Accounting SOP Learning Companion turns any internal SOP or process document into an interactive learning path:

1. Upload or paste an SOP (PDF or text) — including real, messy documents with headers, document-control metadata, and nested section structures, not just clean numbered lists.
2. GPT-5.6 structures the document into logical, actionable process steps — correctly excluding administrative boilerplate (purpose, scope, document control) and correctly handling nested step sequences inside larger sections.
3. For each step, the learner gets:
   • A brief principle refresher (the universal accounting concept — assumed already known).
   • A clear explanation of how this organization implements it — the specific system, approver, naming convention, or local rule.
4. The learner answers a scenario-based question for each step.
5. GPT-5.6 evaluates the answer live — not by keyword matching, but by genuinely checking whether the answer reflects the organization's actual required process or a generic, textbook-plausible assumption. It explicitly names what's correct, what's missing, and what the organization-specific difference is.

Real example, tested live

Using an actual SOP I wrote (a vendor-validation process addressing a real cross-border vendor mix-up scenario — an Italian vendor "WEC" versus a similarly-named "W.E.C." vendor based in Kenya, sharing an apparent tax ID):

• Informal, real-world-style answer: "saw 2 vendors similar name not sure which one just picked the italy one bcoz invoice says italy" → correctly flagged: choosing based on country alone doesn't satisfy the organization's required three-way identity check (legal, geographic, payment identity).
• Complete, organization-specific answer → correctly passed, with confirmation of every specific requirement met (escalation, written confirmation, correct identifiers).

Notably, the evaluation held up even against messy, ungrammatical input — important, since real users (junior staff, new joiners) don't type in clean, structured prompts.

How we built it

Built entirely in Codex (using GPT-5.6 Sol) in a single continuous session (Feedback/Session ID: 019f7463-aa8f-7c23-bc15-34185b3f6d3e):

• Frontend/backend: React + Vite client, Express API — scaffolded by Codex from a single natural-language brief.
• Document structuring: originally a deterministic, rule-based splitter; upgraded mid-build to a GPT-5.6-powered structuring step (via the Responses API) once testing showed the deterministic approach couldn't handle real, messily-structured SOPs — with the deterministic version preserved as an automatic fallback if the API is unavailable.
• Answer evaluation: GPT-5.6, via the Responses API, evaluates learner answers against the source SOP text, distinguishing organization-specific correctness from generic/textbook assumptions.

Challenges we ran into

• Scaling: a 41-page, 50-step synthetic test document initially exposed a hardcoded 12-step limit, incorrect subsection grouping, front-matter misclassification, and an inaccurate step-count summary — all fixed and regression-tested.
• Real-world document structure: my actual SOP (28 numbered sections, with nested internal "Step 1–7" sequences inside specific sections) broke the original deterministic splitter — title truncation, stray bullet artifacts, and mismatched principle refreshers (e.g., an "invoice" keyword incorrectly triggering an expense-recognition-timing principle unrelated to vendor validation). This is what motivated moving document structuring itself onto GPT-5.6, rather than patching more regex rules.
• A live integration bug a compile-only check couldn't catch: a legacy Express router fallback route incompatible with the current router version, caught only through an in-app browser smoke test.

Accomplishments that we're proud of

• The GPT-5.6-powered structuring step correctly handled my actual, unedited real SOP — 24–27 organization-specific steps identified, no manual cleanup required.
• The evaluation logic correctly distinguishes generic-but-plausible answers from organization-specific correct ones, verified on two independent real scenarios (a month-end close process and a vendor-validation process), and remained accurate even against deliberately informal, ungrammatical test input.
• A tested deterministic fallback ensures the app doesn't break if the API is unavailable.

What we learned

That the real bottleneck for AI in finance isn't a lack of automation capability — it's that most tooling assumes clean, already-correct data and documentation. The moment you test against a real, messy, human-written document, the gap becomes obvious — and closing it (as we did, moving structuring itself onto GPT-5.6) is where the real value is.

What's next for Accounting SOP Learning Companion

Today, this tool operates on SOP learning and onboarding. The same core mechanism — evaluating whether a stated decision reflects an organization's actual required process, or a generic, plausible-but-wrong assumption — extends naturally toward real-time, transaction-point guidance: catching an incorrect vendor-master selection before an invoice is posted, warning that a system "reset" doesn't mean cancellation before a duplicate payment goes out, or flagging a near-duplicate payment batch before approval. That's the direction I want to take this next — from static documentation and after-the-fact error detection, toward contextual, preventive finance intelligence.

Built With

• api
• express.js
• gpt-5.6
• javascript
• node.js
• openai
• react
• responses
• vite