Why We Built Our Own Tax Retrieval System Instead of Buying One

That date is dead. Ministerial Decision No. 66 of 2026 moved it to 30 October. But the old figure still sits in the Ministry's own published guidelines, so the model finds it, quotes it, and sounds certain.

There is the problem in one line. Not that these models are stupid. They are fluent, plausible and quietly out of date, and they give you no way to tell which.

We could not put that in front of a client. So we built our own.

What we built

A retrieval system over primary sources, with nothing else in it.

The tax corpus holds 708 documents. 627 come from the FTA and the Ministry of Finance, covering VAT, excise, Corporate Tax, procedures, ESR and EmaraTax. The other 81 are Big 4 alerts, kept apart and labelled as opinion, because commentary is not law and should never be dressed up as it.

A second corpus holds 83 UAE legal instruments. Company law, free zone regimes, AML, DIFC, employment, the audit profession. Between them the two run to 791 sources and just over 15,000 searchable passages. You can browse the same documents in our legislation library.

Every one is the official text. Not a summary. Not a firm's note on it. The decision itself.

It updates every night. New FTA and Ministry material is pulled automatically, so a decision published today is searchable tomorrow morning. That matters more than it sounds: 46 official documents landed in 2026 by the end of August, and July alone produced 13, five of them a kind of instrument the FTA had never issued before. Refresh a corpus quarterly and you miss all of it.

Nobody scraped this together and walked away. Our team decides what goes in and what stays out, document by document. Arabic texts we cannot verify in translation are excluded rather than guessed at. The system was trained and tested on real client questions, and it is retested every time the corpus changes.

The best thing it does is refuse

Ask it something the sources do not cover and it replies "Not found in the supplied sources", then stops.

That sounds like a limitation. It is the most important thing in the build. A system that cannot say I do not know will invent something, and an invented tax answer is worse than no answer. Ten refusals beat one guess.

When it does answer, the law reference, the article number and the quoted wording come with it. Checking takes about as long as opening the PDF.

Two desks that cannot see each other

Tax questions go to the FTA corpus. Legal questions go to the legal corpus. Neither can reach the other's sources, so a tax question asked on the legal desk gets a refusal rather than a half-right answer.

That separation was not free. Our first build missed one retrieval path, and FTA sources leaked into legal answers. We found it by deliberately attacking the boundary instead of trusting the design, fixed it, then tested again in both directions.

Worth saying out loud, because nobody publishes this sort of thing. A system can look perfectly correct in ordinary use and still be wrong at the edge.

Try them yourself. AskCALX Tax and AskCALX Legal, both from the AskCALX hub.

The problem nobody has solved

Finding the right provision is the easy half. Knowing it was amended last year is the hard half.

Repealed text reads exactly like live text. Nothing in the wording tells you it is dead. Put the original and the amending decision in the same corpus and a retrieval system can return the original, quote it perfectly, and be completely wrong.

Ministerial Decision No. 229 of 2025 repealed Ministerial Decision No. 265 of 2023, retroactively to June 2023. Ministerial Decision No. 66 of 2026 replaced a single paragraph inside Ministerial Decision No. 244 of 2025. A machine reading either pair, with no sense of which came second, has no way to choose.

Amendments and effective dates have to be checked against the facts and the tax period. We handle it by showing the source and its date on every answer. Answers from the public desk are automated research, not advice. Anything we act on for a client is read and signed off by a person first.

What it is not

Advice. It finds the provision and quotes it. Working out what that provision means for your business, on your facts, is the part we are licensed for and the part we have not automated.

It is not a product we sell either. We built it because we needed it.

Two things follow for you. Answers come back the same day with the article number attached, because finding it is no longer the slow part. And they come from the text in force now, not from a note written in 2023 or from whatever a model happened to remember.

Ask it something specific. If your question turns on your own facts, send it to us instead.