Quick Answer
Agentic AI tools can draft a return, sort receipts, and flag a plausible deduction in seconds, but current IRS rules stop them well short of filing anything on your behalf. Three walls hold no matter how capable the underlying model gets: only a human taxpayer, through a signed Form 8879 or a Self-Select PIN, can authorize an e-filed return; a human preparer, not software, must complete and retain the Form 8867 due-diligence file behind credits like the Earned Income Tax Credit; and no agent yet reliably resolves incomplete source data — a blank cost-basis field, a few out-of-state workdays, a late Schedule K-1 — without guessing wrong or needing a person to close the gap. Treat an agentic tax tool as a fast first draft for the 2026 filing season, not a filer, and budget a human pass on anything touching basis, multistate income, or refundable credits.
Why Agentic Tax Filing Is Suddenly Everywhere in 2026
Tax software stopped being a glorified form-filler sometime around 2025. What used to be a wizard that asked “did you sell any stock this year?” is now a system that reads a PDF brokerage statement, categorizes forty types of freelance expenses, drafts a Schedule C, and explains its reasoning in plain English before a human even opens the return. Vendors call these systems agents rather than assistants because they chain several steps together on their own initiative: pull the document, classify the line items, populate the form, cross-check for obvious errors, and propose a filing package. That is a meaningfully different capability than the autofill and OCR tools of a decade ago, and it arrived at exactly the moment the government stepped back from its own free-filing experiment.
The IRS notified state partners in late 2025 that Direct File, its free federal e-filing pilot, would not return for the 2026 season. That decision left a gap in the market that commercial AI-driven tax products moved quickly to fill, and it pushed a larger share of ordinary filers toward chatbot-style tax help — some built into established software, some just a general-purpose model asked to “check my return.” The same year, the IRS Office of Professional Responsibility issued its first dedicated guidance on artificial intelligence inside Circular 230, the rulebook that governs enrolled agents, CPAs, and attorneys who practice before the agency. The message to professionals was blunt: using an AI tool does not lower the duty of due diligence, and a practitioner cannot point to a model’s output as an excuse for an error. That single sentence is really the spine of this entire article. Nothing about how fluent or confident a model sounds changes the legal architecture built around a human filer signing, certifying, and standing behind a return.
For a look at where automation already works cleanly — mostly around straightforward, single-state gig income — our companion piece on AI tax automation for freelancers is a useful counterpoint to everything that follows here, because it maps the territory where agentic tools genuinely save time without hitting a wall.
The Human-Signature Wall: Why No Agent Can Actually File Your Return
Every individual return transmitted electronically to the IRS needs a signature equivalent, and the agency has never built an exception for software acting on its own. If you self-prepare through an online product, you authorize the filing with a Self-Select PIN tied to your prior-year AGI. If a paid preparer transmits on your behalf, the mechanism is Form 8879, IRS e-file Signature Authorization: the preparer must show you the completed return, you review the actual numbers, and you sign — on paper or through an approved electronic signature process — before anything goes out the door. The Electronic Return Originator is required to retain that signed authorization for three years and produce it on request.
An agentic system can draft every line of that return. It cannot be the “you” in that sentence. There is no legal category under which a piece of software attests, under penalty of perjury, that a return is true, correct, and complete to the best of its knowledge. That attestation has to trace back to a natural person who reviewed the numbers, which is precisely why the most aggressive “fully autonomous tax filing” pitches quietly still insert a human click-to-confirm step right before transmission. It is not a UX choice. It is the law forcing a stop.
This wall also shapes what a genuinely well-built agentic product looks like in practice. The better implementations treat the model as a drafting and QA layer, then route the finished package through an explicit, unambiguous human sign-off screen that shows the actual figures being transmitted, not a summary. Products that bury or auto-check that step are cutting a corner the IRS has not agreed to let anyone cut.
The Due-Diligence Wall: Form 8867 and the Credits an Agent Can’t Self-Certify
Refundable and partially refundable credits carry a separate layer of friction that has nothing to do with signatures. Any paid preparer who completes a return claiming the Earned Income Tax Credit, the Child Tax Credit or Additional Child Tax Credit, the Credit for Other Dependents, the American Opportunity Tax Credit, or Head of Household filing status must also complete and keep Form 8867, the Paid Preparer’s Due Diligence Checklist. That form is not a rubber stamp. It requires the preparer to document that they asked follow-up questions when something looked inconsistent, that they did not rely on information they knew or should have known was incorrect, and that they can show their work if the IRS ever asks.
The penalty for skipping this is not trivial: IRC Section 6695(g) imposes a per-failure, per-credit penalty on the preparer that runs well over $600 and is adjusted for inflation most years, so a single return with two or three qualifying credits handled carelessly can generate multiple penalty hits at once. Crucially, the obligation sits with the human preparer of record, not with any tool they used to get there. An agent can pre-fill the checklist boxes based on the data it was given, but it cannot supply the judgment call the IRS actually wants: did anything about this taxpayer’s situation look off, and did someone competent chase it down? A model trained to be agreeable is structurally bad at the specific job of pushing back on a client’s story, which is exactly the job Form 8867 exists to force.
DIY filers claiming these same credits without a paid preparer face a lighter documentation burden on paper, but the IRS’s own audit-selection systems lean heavily on machine scoring to flag EITC and dependent-related claims for extra scrutiny. An agentic tool that fills in a qualifying child’s residency months without probing for the messy custody or multiple-household situations that actually determine eligibility is not committing a paperwork failure so much as setting up its user for a notice eight months later.
The Data-Completion Wall: Missing Cost Basis, Multistate Income, and K-1s
The least discussed and most expensive failure mode in agentic tax filing has nothing to do with law and everything to do with incomplete source documents. Three patterns show up constantly.
The first is cost basis on older or transferred brokerage lots. Since 2011, brokers must report cost basis to the IRS on “covered” securities, but plenty of positions remain “noncovered” — shares acquired before the reporting mandate took effect, restricted stock that vested years ago, or lots moved between custodians without a matching basis transfer. On the 1099-B, that shows up as a blank Box 1e. A well-designed system flags the gap and asks the human for the original vesting statement or transfer confirmation. A poorly designed one treats a blank field as a zero, which converts what might be a modest gain into a fully taxable one, sometimes by tens of thousands of dollars on a single lot.
The second is multistate sourcing for people who work across state lines without a matching set of state-coded 1099s or W-2s. A remote consultant based in a no-income-tax state who spends several weeks physically working inside a state that does tax income — California and New York are the two most aggressive about this — can trigger a nonresident filing obligation purely from the location where the work was performed, regardless of which state issued the 1099 or where the client is headquartered. Nothing on the standard tax documents tells an agent where the taxpayer’s body was sitting on any given workday. Absent a travel log or calendar the taxpayer manually supplies, the tool has no way to know an obligation even exists, so it simply doesn’t surface one.
The third is Schedule K-1 timing and complexity. Partnerships, S corporations, and many trusts are permitted to issue K-1s as late as mid-September under extension, well after most individual returns are drafted. An agent working from whatever documents happen to be available in February will either omit K-1 income entirely or work from a prior-year estimate, and K-1 basis tracking — at-risk limitations, passive activity loss rules, prior suspended losses carrying forward — involves multi-year continuity that most consumer-grade agentic tools were never given the data to reconstruct in the first place.
None of these three failures require the AI to be “wrong” about tax law. The model can know every relevant rule correctly and still produce a bad return because the input data it was handed was incomplete, and it filled the gap with a plausible-sounding default instead of a flag.
The Knowledge-Drift Wall: Hallucinated Rules and Stale Thresholds
Tax law changes every year, sometimes mid-year, and large language models have a documented tendency to answer confidently even when their underlying training data is stale or when a plausible-sounding rule simply does not exist. Bloomberg Tax and several professional publications have tracked cases of generative tools citing outdated deduction limits, misquoting phase-out thresholds, or inventing a code section that sounds authoritative but is not real. For a filer who cannot independently check the citation, the confidence of the answer is not a signal of its accuracy.
This is a big part of why the IRS Office of Professional Responsibility’s 2026 Circular 230 guidance puts the verification burden squarely on the human practitioner. A professional who uses an AI research tool is still expected to confirm any cited authority against the actual current-year statute, regulation, or IRS guidance before relying on it in a filed position. There is no professional-responsibility carve-out for “the AI told me the threshold was X.” The same logic should apply, informally but just as seriously, to anyone using a consumer chatbot to sanity-check their own return: a confident answer about a specific dollar threshold, phase-out range, or filing deadline is worth exactly nothing until it is checked against the IRS’s own current-year instructions or a recently updated professional source.
A Worked Example: How Two Missing Data Points Cost Danielle Real Money
Danielle is a single marketing consultant who lives full-time in Texas, a state with no personal income tax. In 2025 she billed $210,000 for consulting work, all reported on 1099-NEC forms, and she also sold a block of restricted stock units that had vested back in 2016 and were later transferred to a new brokerage without a matching cost-basis record. The sale generated $18,000 in proceeds. Her old vesting statements, which she still had, showed an actual basis of $6,800, meaning her true gain was $11,200. Separately, nine weeks of that consulting year involved an embedded engagement physically performed inside California for a California-headquartered client, generating $54,000 of the total fee.
An agentic tax tool imported her 1099-NEC totals cleanly and produced an accurate Schedule C. On the brokerage statement, it encountered a blank Box 1e for the noncovered lot and defaulted the cost basis to zero rather than prompting her for the original vesting documentation, reporting the full $18,000 as gain instead of the real $11,200. It also had no visibility into where she had physically performed the work; because every 1099-NEC she received was issued without any state code and her home address was Texas, the tool built a Texas-only federal return and never generated a California nonresident filing.
The table below compares what the agent produced against what a corrected filing looks like once a human filled in the two gaps.
| Line item | Agent’s filing | Corrected filing |
|---|---|---|
| Reported capital gain on the noncovered lot | $18,000 (basis defaulted to $0) | $11,200 (basis of $6,800 applied) |
| Federal tax on that gain (15% long-term rate) | $2,700 | $1,680 |
| California nonresident return (Form 540NR) filed? | No | Yes, on $54,000 of CA-source income |
| California tax on that sourced income (approx., mid-bracket) | $0 paid, liability still owed | Roughly $4,300 paid |
Two things stand out. First, the basis error actually cost Danielle money she was owed back — she overpaid the IRS by about $1,020 because the agent’s default assumption ran in the taxpayer’s disfavor, not in her favor, which cuts against the common assumption that AI errors always benefit the filer. Second, and far more seriously, the unaddressed California obligation does not go away; it sits as an unfiled, unpaid liability that keeps growing through interest and, once the Franchise Tax Board eventually cross-references her client’s California-sourced payment records, potential failure-to-file penalties on top of the original tax. A single missing document field and a few weeks of unlogged travel turned into more than $5,300 of mishandled tax exposure that the agent’s confident-looking summary never flagged as a problem.
Dollars Left Unresolved by the Agent’s Filing
Correct federal gain
Agent-reported gain (basis defaulted to $0)
Unfiled California nonresident tax exposure
Bar widths are scaled to dollar value within this example; the dashed line marks the zero baseline the agent’s own summary implied for the state liability.
Where Agentic Tools Stand Today: A Task-by-Task Breakdown
Not every part of a return carries the same risk. The table below sorts common tax-prep tasks by how much an agent can currently be trusted to handle without a close human review, based on the governing rule or form behind each one.
| Task | Agent-safe today? | What still needs a human | Governing rule |
|---|---|---|---|
| Importing and categorizing W-2/1099 income | Largely yes | Spot-check totals against source documents | N/A — data entry |
| Standard deduction vs. itemizing decision | Largely yes | Confirm itemized receipts actually substantiate claims | Schedule A instructions |
| Cost basis on noncovered securities | No | Locate original purchase/vesting records | Form 8949 instructions |
| EITC, CTC, AOTC, or Head of Household claims | No, if a paid preparer is involved | Preparer must independently document eligibility | Form 8867 / IRC 6695(g) |
| Multistate or nonresident sourcing | No | Supply workday location logs the agent can’t infer | State-specific nonresident sourcing rules |
| Partnership/S-corp K-1 basis tracking | No | Reconstruct multi-year at-risk and loss carryforwards | Passive activity and at-risk rules |
| Final e-file transmission | Never unsupervised | Taxpayer must personally sign the authorization | Form 8879 / Self-Select PIN rules |
| Responding to an IRS notice or audit | No | Licensed representation (POA under Form 2848) | Circular 230 practice rules |
Common Mistakes Filers Make When Trusting Agentic Tax Tools
- Assuming a clean import means clean data. A tool that successfully reads a PDF has confirmed the file opened, not that every field inside it is complete or correct.
- Letting the tool auto-transmit without a real signature review. A pop-up you click through in three seconds is not the review Form 8879 was designed to require.
- Skipping backup documentation for refundable credits. If a return claims EITC or the Child Tax Credit, someone needs a paper trail showing eligibility was actually checked, not just claimed.
- Ignoring out-of-state workdays until a notice arrives. States that tax nonresident income rarely announce themselves; they wait for a data match and then send a bill with interest attached.
- Treating a chatbot’s cited rule as gospel. A specific dollar threshold or code section needs to be checked against the current-year IRS instructions before it goes anywhere near a filed return.
- Filing before every K-1 arrives. Guessing at pass-through income to hit the April deadline usually costs more in amended-return work than requesting a routine extension would have.
A Practical Checklist Before You Let an Agent Touch Your Return
- List every account, employer, and platform that issued you a tax document, and confirm the agent actually ingested all of them, not just the ones it found first.
- For every 1099-B line, manually check whether Box 1e is populated; if it’s blank, go find the original purchase or vesting record before accepting any basis figure.
- Write down every state where you physically performed work this year and roughly how many days, then compare that against the states the software actually generated returns for.
- If any refundable credit is being claimed, keep a folder of the documents that support it — school enrollment records, custody agreements, residency proof — even if no one asks for them yet.
- Read the actual return line by line before signing, not just the software’s plain-English summary of it.
- Verify any cited dollar threshold, deduction limit, or deadline against the current IRS instructions for the specific tax year, not last year’s numbers.
- If your income involves a K-1, confirm whether it has actually arrived; if not, file an extension rather than estimating.
- Keep the final signed authorization and a copy of the transmitted return for at least three years, matching the retention period preparers themselves are held to.
Key Takeaways
- Agentic tools can draft, categorize, and cross-check a return, but only a human taxpayer can supply the signature that legally authorizes e-filing.
- Refundable credits carry a separate due-diligence obligation under Form 8867 that falls on the human preparer of record, with real per-failure penalties for skipping it.
- The costliest failures usually trace back to incomplete source data — missing cost basis, unlogged multistate workdays, late K-1s — not to the model misunderstanding tax law.
- A single blank field, like a missing Box 1e cost basis, can swing a filing by thousands of dollars in either direction, so silent defaults deserve more suspicion than confident summaries.
- IRS guidance issued in 2026 puts the verification burden for AI-assisted work squarely on the human professional, and the same discipline is worth applying to DIY filers using consumer AI tools.
Frequently Asked Questions
Can an AI agent actually file my taxes for me, or do I still need to sign it myself?
You still need to sign it yourself. Every electronically filed individual return requires a signature equivalent, either a Self-Select PIN for self-prepared returns or a signed Form 8879 when a paid preparer transmits on your behalf. No agentic tool can supply that attestation, because it has to trace back to a natural person who reviewed the actual figures before transmission.
Why did an AI tax tool assume my cost basis was zero?
When a broker doesn’t report cost basis on a 1099-B, usually because the security is “noncovered” (bought before mandatory basis reporting began or transferred between custodians without a matching record), the field is left blank. Some tools default a blank basis field to zero rather than prompting for documentation, which overstates your taxable gain. Always check Box 1e on every 1099-B line before accepting the software’s number.
Do I need to file a state return if I only worked there temporarily?
Possibly, yes. States that tax income, including California and New York, can require a nonresident return based on where you physically performed work, not where your employer or client is located or where you live. Standard tax documents don’t capture your workday location, so an agentic tool has no way to know an obligation exists unless you supply that travel information separately.
Can an agentic tool complete the due-diligence checklist for tax credits like the EITC?
It can populate the checklist fields, but it cannot supply the substantive judgment Form 8867 exists to require. A paid preparer who claims the Earned Income Tax Credit, the Child Tax Credit, the American Opportunity Tax Credit, or Head of Household status must personally document that they asked follow-up questions and didn’t rely on information they had reason to doubt. That obligation sits with the human preparer, not with any software they used.
Are AI-generated tax citations reliable?
Not automatically. Generative tools have a documented history of citing outdated thresholds or referencing rules that don’t actually exist, especially for provisions that change from year to year. Any specific dollar amount, phase-out range, or deadline an AI tool cites should be checked against the current-year IRS instructions before it influences a filed position.
What should I still hire a human for, even if I use an agentic tax tool?
Keep a licensed preparer or your own careful review in the loop for anything involving refundable credits, multistate or nonresident income, partnership or S-corp K-1s, and the final signature step. Those are exactly the areas where the current rules require human judgment or human attestation, regardless of how sophisticated the underlying software becomes.
References
- Internal Revenue Service: Form 8879, IRS e-file Signature Authorization, and Instructions.
- Internal Revenue Service: Form 8867, Paid Preparer’s Due Diligence Checklist, and Instructions.
- Internal Revenue Service, EITC Central: Due Diligence Law, Regulations, and Requirements; Consequences of Filing EITC Returns Incorrectly.
- IRS Office of Professional Responsibility: 2026 Guidance on the Use of Artificial Intelligence Under Circular 230.
- Internal Revenue Service, Instructions for Form 8949 (2025): Reporting Sales When Cost Basis Is Not Furnished by a Broker.
- Federal News Network / Nextgov: Coverage of the IRS’s 2025 Notice to States That Direct File Would Not Return for the 2026 Filing Season.
- Bloomberg Tax: AI Hallucinations in Tax: The Risks and How to Mitigate Them.
- California Franchise Tax Board: Guidance on California-Source Income for Nonresident Service Providers.
- Journal of Accountancy: IRS Outlines AI Risks and Circular 230 Duties for Tax Practitioners.
