Tax technology company
Versus template-based OCR or single-form agents, Magnetic is touted as an end-to-end agentic solution, meaning that the system reasons across the whole return. The technical core is an orchestrator agent with a team of specialist sub-agents — one per schedule, with smaller, focused context windows — that can handle any number of individual tasks, and delegate to each other. It pulls a number off a K-1, ties it to a prior-year basis schedule, applies the relevant section of the Tax Code, and writes the result into the firm's software.
More specifically, the user begins by preparing a starting file: For a returning client, the firm rolls forward the prior year's client file in their tax software (UltraTax, CCH Axcess, Lacerte, Drake, or ProConnect) to a blank current-year version, preserving carryovers and historical data, while for a new client, they create a blank file with the taxpayer's basic information. They export that file out of their tax software.
From there, they upload the starting file and the source documents to Magnetic (in any format: PDF, Excel, JPG, Word, scanned paper, phone photos, even handwritten notes). The software then scans every document to confirm it can be opened and parsed, with the user notified about anything that needs their attention before processing begins.
Next is the document parsing stage, where the system reads not just printed fields but handwritten notes, margin annotations, circles, underlines, and any other metadata a human preparer would notice such as file names, number of documents, file types, etc. The documents are flattened, rotated, scaled, categorized (W-2s with W-2s, 1099s with 1099s, and so on), and named by payer and account number ("W-2 — Magnetic" "1099-Consolidated — Vanguard 1234"). Duplicates, documents for non-taxpayers (e.g., a child or a trust), and other anomalies are flagged. The output is a single bookmarked PDF.
This is where the logic layer comes in, which Magnetic pointed to as a major differentiator. The orchestrator agent takes stock of all the documents and plans how to approach the return, using tools to list documents, view their scan results, look up the relevant Tax Code, run calculations, and delegate to specialist sub-agents, one per schedule (Schedule E has a sub-agent, 1098 has a sub-agent, and so on). The sub-agents work with smaller, focused context and specialize in the rules and edge cases of their domain. The output of this phase is an internal representation of the return, not yet written to the tax software, but fully reasoned through.
After that, Magnetic runs the user's tax software on its own infrastructure at the operating system level, automating keystrokes and navigation the same way a human preparer would, meaning there is no need for a vendor API and the user does not have to run anything on their own machine. Every return is then reviewed by an experienced in-house U.S.-based Enrolled Agent or CPA who will confirm the return is correct, fix anything the agent missed, fill in edge cases the system did not handle, and add any preparer notes the system did not generate on its own. The human professional uses an internal tool that surfaces the system's own notes and confidence scores so they know where to look first.
Finally, the firm receives three artifacts: the completed tax software client file, the bookmarked workpaper PDF, and a notes file documenting anything that needs taxpayer follow-up, missing documents, or items flagged for additional review. The humans do a 20-30 minute high-level review and send the return on. Outputs are cross-checked across models to produce a per-value confidence score that downstream agents and human reviewers can rely on.
"This step is how the accuracy guarantee actually holds up at the limit — it takes us from 'really good' to near-perfect. This is also the method by which we receive rapid feedback that allows us to continually tweak the system to account for all of the quirks and details of each tax engine and niche Tax Code, which has been key to the pace of product development," said Magnetic co-founder Thomas Shelly in an email.
Shelly strongly emphasized the accuracy guarantee: If a Magnetic-prepared return contains an error that results in a change of $10 or more to the final tax liability, the entire return is free. Magnetic said it chose $10 as its threshold as it is small enough that firms should feel incentivized to alert them to any mistake they catch with the assurance that it benefits them and Magnetic, as feedback is a major driver of accuracy improvements. In the Spring 2026 tax season, only 3% of users claimed the guarantee.
"This is the first unconditional accuracy guarantee any tax prep service has put in writing. The purpose is total alignment with our customers: Their pain is our pain. We need to hear about every mistake, no matter how small, so we can build the best tax agent on the market," said Shelly.
While they boast of a 99.9% field-level accuracy rate on all common forms, Magnetic felt this was not the best metric to focus on, as accuracy is more than just entering every field correctly (for example, mortgage interest can be entered perfectly on Schedule A, "Itemized Deductions," but if it should have been attributed to the home-office, the return is still incorrect).
"Our product goes beyond strict data input, as we have built an agentic model that can contextualize information and make logical choices on each tax return. A return-level guarantee is the only way to measure that success in the way that matters most to the firm," said Shelly.
Magnetic also guarantees that tax software files will have a turnaround time of three days, with the average delivery time during the 2026 spring season being about 2.5 days.
Right now, Magnetic supports 1040 tax filing only. They do not handle business or entity returns. The agent supports a broad range of tax situations commonly found on individual returns, including Schedules A, B, C, D, E, K-1 income, multistate, and basis tracking.
Shelly also noted that Magnetic is IRC Sec. 7216 compliant. Further, all user data stays in the U.S., all processing happens on U.S. infrastructure, and all human reviewers are U.S.-based EAs and CPAs. In addition, Magnetic does not train on customer data. Its reviewer feedback loop is human-in-the-loop edge-case detection, not model training on customer data. The company also publishes a written information security plan,
