TaxFlowby CrewCircle
For practitioners, by design

How TaxFlow actually works

Written for accountants who've been burned by AI before. No black box - every mechanism below is the real one, including the parts that sometimes fail and the parts we haven't built yet.

01 · Your question, end to end

Four stages, each inspectable - not one opaque model call

Step 1

Retrieval

Find the real source material first

Your question searches a knowledge base of real ATO rulings, determinations, and legislation - hybrid semantic + keyword search, re-ranked, not one vector lookup.

  • Weighted toward the right state/jurisdiction, without hard-excluding others
  • One long Act can't crowd out a shorter, more on-point ruling
  • Your firm's own precedents compete in the same ranked pool, not bolted on after
  • Superseded law is retrieved separately, down-weighted, and labelled historical
Step 2

Routing

The model is chosen by the question

A fast, cheap model handles most questions. TaxFlow escalates to a stronger model automatically when retrieval signals say the question is hard - you never pick a "mode."

Step 3

Generation

Drafted strictly from what retrieval found

Every factual claim carries a numbered citation. When the sources don't cover the question, TaxFlow is instructed to say so - not fill the gap from general training knowledge.

Step 4

Verification

A second, independent pass checks the first

A separate model call - different prompt, acting as a senior reviewer, given only the draft and the sources - checks every claim and returns a graded verdict: critical, warning, or note.

02 · What verification is actually checking for

Three specific failure modes - not a generic "fact check"

Factual accuracy

The reviewer is given the exact source text and the exact claim, and asked whether it's supported - graded critical (wrong rate/section/test), warning (misleading or incomplete), or note (minor).

Topic mismatch

A fluent answer can cite real documents that are simply about the wrong area of law - a payroll-tax ruling backing an FBT question. TaxFlow classifies the topic of the question and every cited source separately, and treats a mismatch as an automatic trigger regardless of how confident the answer sounds.

When it runs

Verification is gated to answers with real risk signals - low confidence, few citations, an explicit "sources don't cover this" admission, or a topic mismatch. A confident, well-cited, on-topic answer skips the slower check.

When something critical is found: it triggers a corrective regeneration - TaxFlow attempts to fix the answer with a widened source pool before you ever see it. What you see is the result after that correction loop has already run once.
03 · A real answer, not a demo one

Taken from a live session - including where it stopped short

Question asked

"One of our partner entities is a small business entity with turnover under $2M - does it qualify for the 50% CGT discount plus the small business 50% reduction on the sale of its interest in the development?"

Verified against sourcesUsed 5 sources

"Based on the provided sources, I cannot give you a complete answer to this question because the sources do not address capital gains tax (CGT) concessions or the 50% CGT discount and small business 50% reduction you mention.

The provided documents (18) focus exclusively on the definition of 'small business entity' under Subdivision 328-C... they do not contain the CGT-specific rules that govern whether a small business entity qualifies for the 50% CGT discount."

The answer then states plainly what the sources do confirm, and recommends which legislation (Division 152 ITAA 1997) and action (a Private Binding Ruling, if complex) to pursue instead.

Retrieval found real, correctly-classified sources - they just weren't the right ones for this question, because the knowledge base's CGT-concession coverage had a real gap. TaxFlow didn't paper over that gap with a fluent guess. It said so, and named where to look instead. We'd rather you see this answer than a more impressive-looking one that got the legal test wrong.

04 · Human sign-off is structural

Enforced server-side - not a UI convention a determined user could route around

Staff

  • Ask questions, draft documents

Cannot approve or finalise anything

Reviewer

  • Everything Staff can do
  • Approve and finalise documents & ATO responses

Owner

  • Everything Reviewer can do
  • Manage firm knowledge, staff, and billing

A firm can never remove its last active Owner

A junior team member can draft an advice memo at any hour. It doesn't become the firm's work product until a Reviewer or Owner explicitly approves it - the same discipline a paper-based practice already runs, enforced in the software instead of left to informal habit.

Where your data actually goes

Client records, documents, and firm knowledge are stored and hosted in Australia, isolated per firm - never blended into another firm's answers. Generating an answer sends the text of your question to Anthropic's and OpenAI's standard API infrastructure, which is not Australia-only - that's genuinely how this category of product works today, and we'd rather say so than let you assume otherwise.

Read the full data residency page →
05 · What this doesn't do

Stated as plainly as everything above

A "how it works" page that only describes strengths isn't transparency, it's marketing with extra steps. These are real, current limitations.

Not a substitute for professional judgement

Every draft carries an explicit limitation note. The ATO has stated relying on AI-generated advice isn't a reasonable excuse for an error - this is built with that standard in mind.

Knowledge base coverage has real gaps

Coverage grows over time; it isn't complete today. TaxFlow tells you when a question falls outside it, rather than guessing.

Verification isn't infallible

It's a second model call checking a first one, not a formal proof of correctness. It substantially reduces the risk of an unreviewed error reaching you - it doesn't remove the need for review.

Depends on third-party AI providers

Answers are generated by external LLM providers. A provider-side outage can cause a question to fail outright - error handling for that is active, ongoing work.

Test it on your own hardest question

That's the actual point of this page - not to be convinced by us, but to go check.