TaxGPT

Two products share this name and a firm evaluating it should treat them separately. The first is research: a question in plain English returns an answer where every material claim carries a citation to primary authority, drawn from the Code, Treasury regulations, federal and Tax Court decisions, revenue rulings and procedures, private letter rulings, and state codes across all 50 states, with Canadian authority and treaties alongside. Tax Writer turns a conclusion into a memo or an IRS letter. Document analysis reads a return or a notice and reports what it found. The second is workflow: Cowork assigns an agent to a process a firm currently runs by hand, and Agent Andrew reviews a completed return.

On currency, TaxGPT makes a specific claim that distinguishes it from the licensed libraries. The database ingests new legislation, IRS and Treasury guidance and court decisions as they are published rather than on an editorial update cycle, and the vendor names OBBBA, SECURE 2.0 and the IRA clean-energy credit phase-outs as examples. Sources are curated and vetted by the company's CPAs and tax attorneys. That is a reasonable answer to the currency problem and a partial one to hallucination, and the site is careful not to claim absolute accuracy, framing the output as something a professional traces and verifies. What is missing is any stated behaviour when the model cannot support an answer. Nothing describes a refusal threshold, and a research tool that will always produce something is a different risk profile from one that says it does not know.

The workflow side is where the real questions sit. Cowork agents are described as autonomously logging into and operating a firm's existing tax and practice management software, with connections to Drake, ProConnect, UltraTax, CCH Axcess and Lacerte and to storage in Google Drive, Dropbox, OneDrive, SharePoint, SafeSend, SmartVault and Canopy, surfacing items needing human judgment. No migration, which is the appeal. It is also a piece of software holding credentials to the systems of record and driving them, and whether that is permitted under each tax vendor's own terms is a question to put to Drake or Wolters Kluwer, not to TaxGPT. Agent Andrew is the safer half: upload a finished 1040, 1041, 1065, 1120 or 1120-S, get a colour-coded report of errors, audit-risk patterns and missed deductions in five to ten minutes.

Security is better documented than most tools in this tier. The vendor states SOC 2 Type II, AES-256 in transit and at rest, multi-factor authentication, role-based access, US hosting, and that customer data is never used to train the model with no opt-out required because the practice does not exist. Pricing is the opposite: three named plans, Pro, Pro Plus and Enterprise, and not one number on the page. A firm cannot compare this against TaxFigure at $8 a seat or TheTaxBook at $429 without booking a call, and the marketing figure of cutting research time by up to 99% does not help.

One-to-watch stands. The security posture and the integration list argue for more, but a product spanning cited research and credential-holding autonomous agents, sold without a published price and without a stated refusal policy, is not yet something to build a busy season around.