TaxFigure

Ask this one a section 1031 question and it comes back with IRC section 1031(a)(3), Treasury Reg. section 1.1031(k)-1(b) and, if a state is in scope, something like C.R.S. section 39-22-303.6 sitting inline in the answer. That is the whole proposition. TaxFigure is a research assistant whose output is meant to be pasted into a client memo, and it is built so a firm can configure guardrails requiring citations on every response and blocking any answer that lacks primary authority. Around the research sits TaxScan, which turns W-2s, 1099s, K-1s and completed returns into structured data, a five-stage automated return review, and entity-scoped workspaces that track a client's position across years.

Pricing is unusually transparent for this corner of the market and unusually easy to misread. A seat is $8 a month billed annually, with volume discounts starting at five seats and an example on the vendor's page of five seats for $450 a year. But every AI action spends credits from a monthly allowance of 100 per seat: a research chat is 4 to 8, document analysis 2 to 5, and a full return review 60 to 100. One seat therefore covers one return review a month, or roughly fifteen research questions, before extra credits at $0.08 each. The $8 headline is the floor of a metered bill, not the cost of using the product in March.

Underneath the citations there is no licensed editorial layer. TaxFigure reads primary law directly, which means it is not paying RIA or CCH and can price accordingly, and also that nobody has written an explanation of the Code section for a human editor to stand behind. Currency is thin on evidence: the platform stamps answers with a date, and it stamped one 2026-09-03 while being examined, but the update methodology is not published. The August 2026 product notes list a citation overhaul and verified citations as new work, which is reassuring about direction and a signal that clicking through was not always reliable. Nothing on the site describes what happens when the model cannot support an answer, only that firms can block uncited responses.

The security page deserves a specific warning, because it contradicts the homepage. The marketing page claims SOC 2 Type II among its compliance badges. The security page says the SOC 2 Type II examination has an observation period of 1 August to 31 October 2026 with a report expected in early 2027. There is no report yet. The page is otherwise commendably open: AWS us-east-1, TLS in transit, KMS-encrypted storage, MFA enforced on administrative accounts, and a named subprocessor list that includes OpenAI, Anthropic, Google, Perplexity, Groq, Cohere and several inference gateways. A firm with a client-data policy should read that list before signing.

TaxGPT is the closest competitor and has an issued SOC 2 Type II and no published price, which is exactly the reverse trade, while Blue J is the more expensive and more established option offering predicted outcomes rather than metered credits. Against either, one-to-watch is confirmed. The product is real, the pricing is honest, and the security page is more candid than most. A tool that is still bringing its own attestation and citation verification to completion is not one to run a busy season on alone.