Tallywise

Tallywise: AI reads the messy paperwork, plain code does the sums.

Tallywise is our own internal tool. It keeps our own books. An AI model reads the supplier paperwork that has no fixed layout, and plain code does the sums. A person fills any gap in the paperwork with a written reason, their name and the date. For seven filed VAT3 periods, its figures agree with what was filed on every box. It shows how we put AI to work on messy paperwork without letting it do the sums. Read on for what we built, what it measured and what it means for your business.

Problem

What needed solving.

Tallywise keeps our own books. It reads our purchase documents and bank statements, computes our bimonthly VAT3 figures, reconciles them against what was filed, and works out corporation tax.

Our purchase documents have no fixed layout. An AI model can read them. The rule we set was that it reads and reports, and never touches the sums.

Constraints

The limits the work had to fit.

  • Only what is printed. The AI model may only report what is printed on a document. A fact that is not on the page is left empty, never guessed.
  • No decisions for the AI. It decides no VAT treatment, no account and no currency conversion.
  • A person checks every AI reading. Each one is reviewed by a person and saved before any run of the books uses it.
  • Filed returns are history. Tallywise never rewrites a filed return. A change takes effect only when a named person accepts it with a date.

What we built

What we built, and how.

Of 88 purchase documents, 73 were read by fixed code and 15 by an AI model.

  • Fixed code first, AI for the rest. The fixed code is written for each supplier's layout. The AI model reads the documents that have no fixed layout.
  • The model stays out of the sums. No run that builds the books calls the model. Every later run reads the reviewed file and never asks the model again.
  • Gaps filled by a named person. Some documents left a gap, such as no invoice issued. A person filled each gap in a separate file with a written reason, their name and the date. That happened nine times, and the original reading stays on record.
  • Plain code does the sums. It computes the bimonthly VAT3 figures, reconciles them against what was filed and works out corporation tax. Two runs over unchanged inputs give identical output.
  • No catch-all account. A bank line it has no rule for stops the run and names the line, rather than going into a catch-all account.

What it measured

What we tracked, and what it showed.

As of 15 October 2026, Tallywise has read 88 purchase documents and 9 bank statements with 104 bank lines. For seven filed VAT3 periods, the computed figures agree with what was filed on every box. Nothing is outstanding: no difference is left unexplained.

We built it from 18 to 22 September 2026, in 25 commits.

A test is a small program that checks one thing the tool should do. On 15 October 2026 it had 606 tests, all passing.

Reconciliation surfaced one error repeated across six of our VAT3 returns. That was in September 2026. All six returns were amended as a self-correction, and €2.75 of interest was paid.

What this means for you

Messy paperwork, read

Fixed code, written for each supplier's layout, read 73 of our 88 purchase documents. An AI model read the other 15. A person reviewed each AI reading before the books used it. Order emails and supplier invoices in your business often have no fixed layout either.

AI kept from the sums

The model reports only what is printed on a document and leaves a missing fact empty. It decides no VAT treatment, no account and no currency conversion. Plain code does the sums. Your staff can draw the same line when they use ChatGPT for work.

Books that agree

For seven filed VAT3 periods, the figures agree with what was filed on every box. Reconciliation surfaced one error repeated across six returns, and all six were amended. A mistake shows up, rather than hiding in a total.

What we handed over

What was left in place at the end.

Tallywise prepares figures; a person files them. It is not tax advice, and it is not a product we sell.

If order entry and invoicing eat hours in your business, AI Automation and Agents automates one defined workflow end to end. Do your staff paste company data into ChatGPT without training or an agreed policy? The AI Literacy Workshop covers safe use with company data and drafts an AI use policy.

Where to start

Two ways in. Pick whichever suits you.

Talk to a founder.

Tell us what's eating your week. You leave with an honest answer on whether we can help, and where we would start. No preparation needed.

Prefer to write?

Tell us the problem in your own words: what takes the time, who it affects and what you have tried so far. A founder will reply within 24 hours.