Compressing the monthly bookkeeping cycle for sole proprietors is one of the design goals behind Qkwit: AI reads accounting documents, while a deterministic engine calculates taxes and social security contributions. That’s a design assumption, not a measured outcome for every business — here’s how the mechanism actually works.
What a bookkeeping month looks like today, without AI
For most sole proprietors, the monthly cycle looks fairly similar: gathering cost and sales invoices scattered across email, a banking app, and paper receipts, manually entering them into a spreadsheet or accounting software, sending everything to an accountant, waiting for the income tax advance and social security contributions to be calculated, and finally making corrections if something was misclassified or forgotten. None of this work is complicated on its own — it’s just spread out over time and requires keeping track of a lot of small things at once.
What changes with AI: who reads, and who calculates
In Qkwit, we split two roles that get tangled together in the traditional process. The first role is reading and understanding documents — that’s what AI (Claude) does: recognizing whether something is a cost invoice, a sales invoice, or a receipt, and extracting amounts, dates, tax rates, and expense categories. The second role is calculating what’s owed — the income tax advance, social security contributions, VAT due. That’s handled by a separate, deterministic engine built on explicit tax rules, not a language model. AI never “guesses” a tax amount — it supplies input data, and the engine calculates it using the same rules a person would apply with a calculator and the regulations in front of them.
This split has a concrete rationale: language models are good at recognition and classification, but they aren’t designed to be a deterministic tax calculator. If two people ask a model about the same tax amount phrased slightly differently, they might — though not necessarily — get a slightly different formulation of the answer. A calculation engine doesn’t have that problem, because it always returns the same result for the same inputs. We go into more detail on why we apply this split across all our AI implementations in document workflow automation with AI — the extraction and validation mechanism described there step by step is exactly what sits behind document reading in Qkwit.
What documents AI actually reads in Qkwit
In practice, this means the documents you’d have to deal with every month regardless: cost invoices from suppliers, sales invoices issued to clients, fiscal receipts, payment confirmations, and in some cases correction notes to previously issued documents. AI recognizes the document type, extracts the data needed for bookkeeping, and assigns it to the right expense or revenue category. The more standard a document — a typical VAT invoice in a typical format — the lower the chance it needs manual review instead of being processed automatically. Unusual documents, handwritten notes, or poor-quality scans increase the odds of escalation to a person — we go into the extraction and validation mechanism behind this, shared across all our document implementations, in more detail in the article on document workflow automation.
A compressed month as a design goal, not a promise
We want to be precise here, because it’s easy to over-read this claim. “A bookkeeping month in one day” is a design assumption we build Qkwit around — a target for the product’s architecture, not a measured, averaged result across every user or business type. The actual time it takes depends on how many documents need processing in a given month, how unusual they are, and how many require manual review instead of automatic processing. We treat it as the direction the product is heading — the mechanism described above (AI reads, the engine calculates) is what makes that goal achievable in some cases, not a guarantee of an identical result every time.
What’s live today, and what’s still “coming soon”
Qkwit currently runs in production for sole proprietors. Support for companies, payroll and HR, and accounting firms as separate user profiles is on the product roadmap — we describe it as “coming soon” without giving specific dates, since development priorities can still shift. If you run a sole proprietorship, the mechanism described in this article already applies to you today; if you manage a company or an accounting firm, it’s worth following the product’s progress before planning a migration.
Why AI shouldn’t calculate your taxes on its own
This question comes up often in conversations with clients interested in automating bookkeeping, so it’s worth naming directly: language models, even the best ones available today, aren’t a tool for calculating tax liabilities. They’re very good at reading, classifying, and recognizing patterns in documents — exactly what they do in Qkwit. The actual calculation is left to code that implements explicit, verifiable tax rules. That’s the same principle we apply across every one of our products that runs AI in production, not just in bookkeeping.
A deterministic engine has another advantage that matters if you’re ever audited: for the same inputs, it always reproduces exactly the same result, and you can trace which rules produced it. A language model, however good, doesn’t give you that guarantee of repeatability — and in bookkeeping, being able to reproduce and explain a result matters just as much as the number itself being correct.
FAQ
Does Qkwit really compress the bookkeeping month to one day? That’s a design goal we’re building toward in the product’s architecture, not a measured, guaranteed result for every user. The actual time depends on how many documents there are in a given month and what kind they are.
Does AI in Qkwit decide on its own how much tax is owed? No. AI reads and classifies documents, but the actual tax and social security liability is calculated by a separate, deterministic engine built on explicit rules — not the language model.
Does Qkwit support companies or accounting firms? Qkwit currently runs in production for sole proprietors. Support for companies, payroll, and accounting firms is on the roadmap, but without a set release date.
What happens if AI misclassifies a document? Documents that can’t be classified unambiguously, or that raise a question, go to manual review instead of being posted automatically — the same pattern we describe for any document automation implementation on this blog.
Do you need accounting knowledge to use Qkwit? The product is designed so a sole proprietor can use it without specialist accounting knowledge — AI reads and classifies the documents, and a rules-based engine handles the calculations.
Is the financial data AI processes kept secure? The layer that reads documents and the layer that calculates liabilities are kept separate, and access to data is limited to what a given task actually requires. That’s the same standard we apply across every one of our products running AI in production.
See how it works in practice
If you run a sole proprietorship and want to see the “AI reads, the engine calculates” split in practice, visit the Qkwit page. You can read more about our general approach to AI implementation, where we apply the same role split outside of accounting too, on the AI page — and if you have specific questions, get in touch.