Tax resolution / IRS-notice response

The IRS sends about 170 million notices to individual taxpayers every year (IRS Pub 4054) — and the agency itself concedes they’re long and confusing enough to warrant a “Simple Notice Initiative” to fix them (Treasury). For the recipient it’s a deadline-driven panic: a CP2000 underreporter notice gives 30 days to respond, and ignoring it escalates to a 90-day Notice of Deficiency after which the tax, penalties, and interest are simply assessed (TAS). Most people either overpay out of fear or hand the problem to a high-fee “tax relief” firm — even though one in three CP2000 notices results in no additional tax owed at all (H&R Block). The agent opportunity is to read the notice, check the IRS’s own math, pick the correct response, and file it on time — at software cost, without the fee-mill baggage.

Vitals: market: ~20M taxpayers owed ~$539B in back taxes (Jackson Hewitt) · deadline-driven, ~170M notices/yr · buyer: individuals / small businesses / their CPAs · model: low per-notice fee + subscription · whitespace: ★★☆

Market context — the distressed population and the resolution gap

• The population in distress is enormous. In 2019, ~20 million taxpayers owed ~$539 billion in back taxes (Jackson Hewitt); on top of that, ~170M notices a year flow to individuals (IRS Pub 4054), the bulk of them routine adjustments, balance-due, and math-error letters.
• The headline remedy barely functions at scale. Of those 20M who owed, only 54,225 even applied for an Offer in Compromise and just 17,890 were accepted (Jackson Hewitt); OIC acceptance was ~21% in FY2024 (Omni). The “settle for pennies on the dollar” pitch fits a tiny minority — yet it’s the industry’s main advertisement.
• The buyer is self-funding under duress. A taxpayer facing penalties and interest, or a small business with a payroll-tax notice, has acute willingness to pay for a correct, on-time response — today they overpay the IRS or overpay a firm.

The mess

• The notice is confusing and the clock is already running. Notices are long and often don’t even explain the reason for the change — the IRS launched a “Simple Notice Initiative” precisely because they’re hard to parse (Treasury); meanwhile a CP2000 is a 30-day letter (Jackson Hewitt) and a Notice of Deficiency a 90-day hard deadline to petition Tax Court (TAS).
• The IRS is often wrong, and silence is treated as agreement. One in three CP2000s results in no extra tax (H&R Block), but if the taxpayer doesn’t respond in time the proposed amount is assessed by default — so inaction converts an IRS error into a real, collectible debt.
• Choosing the right remedy is technical. Dispute the notice, request penalty abatement, set up an installment agreement, file an Offer in Compromise, or seek currently-not-collectible status — each has its own eligibility math and forms, and the wrong choice wastes the only window the taxpayer gets.
• The “help” is the second trap. The tax-relief industry has a documented fraud record — the FTC has repeatedly acted against firms that “bilked consumers… by falsely claiming they could reduce their tax debts” (FTC) — and regulators now warn about Offer-in-Compromise “mills” (Tax Notes).

Why now

Reading an IRS notice, reconciling its numbers against a taxpayer’s wage and income records, choosing among the statutory remedies, and drafting the response on the right form is a parse-records-and-apply-rules task — exactly what an LLM grounded in the IRS’s own rules can now do per-notice, cheaply, at the scale of 170 million letters a year. What used to require a paid professional’s hour per notice can be triaged in minutes.

The honest framing of the gap: AI is already arriving here, but as a copilot for tax professionals — Canopy’s “Coworker” automates notice workflows for firms, and tools like CPA Pilot help firms “stay fast” on notice response (Canopy, CPA Pilot). What’s missing is the taxpayer-facing version: an agent the distressed individual or small business points at their own notice, that acts honestly and at low cost rather than selling a “pennies on the dollar” dream. The incumbents’ trust deficit is the opening.

The money

Signal	Figure	Basis
Notices to individuals per year	~170 million	IRS Pub 4054 / Treasury (irspub, treasury) — VERIFIED
Taxpayers owing back taxes (2019)	~20M, ~$539B	Jackson Hewitt (jh) — VERIFIED
CP2000s with no added tax	1 in 3	H&R Block (hrblock) — VERIFIED
CP2000 / deficiency deadlines	30-day / 90-day	Jackson Hewitt, TAS (jh, tas) — VERIFIED
OICs accepted (2019)	17,890 of 54,225 applied	Jackson Hewitt (jh) — VERIFIED
OIC acceptance rate (FY2024)	~21%	Omni (omni) — VERIFIED

The differentiator is honesty + cost, not a clever billing model

There’s no clean contingency: you can bill on penalties/interest avoided or tax wrongly proposed (the 1-in-3 case), but the natural model is a low per-notice fee plus a subscription. The wedge against a distrusted, high-fee industry is precisely the opposite of that industry — transparent pricing, no “pennies on the dollar” promises, and a correct answer fast.

How it works today

A notice arrives, the clock starts, and the taxpayer takes one of two bad paths: freeze and overpay (or miss the deadline and get assessed), or hand it to a tax-relief firm that charges a large upfront fee against a remedy most people don’t qualify for.

Mermaid source

flowchart LR
  classDef human fill:#fdecec,stroke:#e0564f,stroke-width:1.5px,color:#0f172a;
  classDef pro fill:#eef0fe,stroke:#6366f1,stroke-width:1.5px,color:#0f172a;
  classDef gov fill:#e8f1fd,stroke:#2563eb,stroke-width:1.5px,color:#0f172a;
  classDef ext fill:#fff7ed,stroke:#d97706,stroke-width:1.5px,color:#0f172a;

  IRS("IRS sends ~170M notices/yr<br/>CP2000, math-error, balance-due<br/>often long & confusing"):::gov
  Taxpayer("Taxpayer / small business<br/>panics, often overpays or ignores<br/>(1 in 3 CP2000s owe nothing)"):::human

  subgraph Paths["Two bad paths"]
    direction TB
    Ignore("Ignore it →<br/>30-day CP2000 lapses →<br/>90-day Notice of Deficiency →<br/>assessed + penalties + interest"):::ext
    Hire("Hire a 'tax relief' firm →<br/>big upfront fee, 'pennies on<br/>the dollar' OIC pitch<br/>(FTC actions, OIC 'mills')"):::pro
  end

  Court("Tax Court petition<br/>or full assessment"):::gov
  OIC("Offer in compromise:<br/>rarely filed, ~1 in 5 accepted"):::gov

  IRS --> Taxpayer --> Ignore
  Taxpayer --> Hire
  Ignore --> Court
  Hire --> OIC

Where an agent fits

Collapse the panic into a fast, correct response. The taxpayer uploads the notice; the agent parses it (type, tax year, amount, deadline), reconciles the IRS’s figures against the taxpayer’s own records — catching the one-in-three case where nothing is actually owed — selects the right move (dispute, penalty abatement, installment agreement, OIC, or currently-not-collectible), and drafts and files the response inside the 30- or 90-day window. The taxpayer e-signs; a CPA or EA gates genuine disputes. The product is the response plus an audit-ready file of notice → basis → response.

This is the playbook in miniature: an agent that acts on the deadline rather than waiting to be asked; the IRS’s notice types, deadlines, and remedy-eligibility rules encoded as a domain layer; and reading IRS transcripts and a taxpayer’s own records that mostly lack a clean API.

Mermaid source

flowchart LR
  classDef human fill:#fdecec,stroke:#e0564f,stroke-width:1.5px,color:#0f172a;
  classDef agent fill:#eafbf1,stroke:#16a34a,stroke-width:1.5px,color:#0f172a;
  classDef ext fill:#fff7ed,stroke:#d97706,stroke-width:1.5px,color:#0f172a;
  classDef store fill:#eef0fe,stroke:#6366f1,stroke-width:1.5px,color:#0f172a;

  Taxpayer("Taxpayer / small business<br/>uploads the notice"):::human

  Agent("IRS-notice agent<br/>acts inside the 30/90-day window<br/>honest, low-cost, not a fee mill"):::agent

  subgraph Owns["What the agent owns"]
    direction TB
    Triage("Parse the notice:<br/>type, year, amount, deadline"):::agent
    Check("Verify the IRS math vs.<br/>the taxpayer's records<br/>(catch the 1-in-3 wrong)"):::agent
    Remedy("Pick the response/remedy:<br/>dispute · abatement · IA · OIC · CNC"):::agent
    Draft("Draft + file the response<br/>before the deadline"):::agent
  end

  Data("IRS transcripts/account · wage docs<br/>taxpayer records, mostly no clean API"):::ext
  Gate{{"Taxpayer e-signs the response;<br/>CPA/EA gate on disputes (human gate)"}}:::human
  Trail[("Audit-ready file:<br/>notice + basis + response")]:::store

  Taxpayer --> Agent --> Owns
  Owns -. "reads" .-> Data
  Owns --> Gate
  Agent -. "evidence trail" .-> Trail

Whitespace & incumbents

The field is occupied at the edges, but the honest, autonomous, taxpayer-facing middle is open:

• High-fee human “tax relief” firms — Optima and peers are the loudest incumbents, advertising “pennies on the dollar” settlements; the category carries a documented fraud record (the FTC has forced tax-relief scammers to pay millions for false debt-reduction claims — FTC) and regulators warn of OIC “mills” (Tax Notes). High upfront fees, low trust.
• DIY consumer tax-prep — TurboTax/H&R Block-style products handle filing and bolt on limited audit/notice support, but don’t autonomously diagnose and answer an arbitrary notice for the user.
• AI copilots for tax pros — Canopy’s “Coworker” automates notice workflows and transcript summaries for firms; CPA Pilot and IRS Solutions serve practitioners (Canopy, CPA Pilot). These make the professional faster; they aren’t a self-serve agent for the 20M distressed taxpayers.

Why ★★☆, not higher

This isn’t greenfield — AI notice tooling exists, just pointed at tax professionals, and consumer tax-prep brands have distribution. The defensible wedge is the autonomous, honestly-priced, taxpayer-facing agent that the fee-mill incumbents can’t credibly become and the pro-copilots aren’t built to be.

Hard problems

Problem	Why it’s hard here	Signal	Likely approach (speculative)
Getting the answer right under penalty	A wrong response can forfeit the only window or create real liability; the IRS itself is wrong ~1/3 of the time on CP2000	1 in 3 CP2000s owe nothing (hrblock); silence = assessment (tas)	Probably reconcile IRS figures against pulled transcripts/records, with a confidence threshold routing disputes to a CPA/EA before filing
Accessing the taxpayer’s data	The case needs IRS transcripts, wage/income records, and the taxpayer’s own docs — fragmented, sensitive, mostly no clean API	IRS notices reference data the taxpayer must assemble; AUR is an income-mismatch engine	Likely transcript retrieval + document ingestion under the taxpayer’s authorization (e.g., Form 8821/2848), normalized per notice type
Choosing the right remedy	Dispute vs. abatement vs. IA vs. OIC vs. CNC each have distinct eligibility math; OIC fits very few	OIC accepted ~21% and filed by a tiny fraction (jh, omni)	Probably an eligibility model that recommends the highest-value qualifying path and refuses to oversell OIC
Earning trust in a scam-stained category	The incumbent brand is “tax relief scam”; users are primed for distrust	FTC enforcement; OIC “mills” warnings (ftc, taxnotes)	Likely transparent flat/low pricing, no settlement promises, and a verifiable evidence trail as the anti-fee-mill posture

Sources

• IRS Publication 4054 · U.S. Treasury — Simple Notice Initiative — ~170 million notices a year; the IRS’s own admission they’re confusing
• Taxpayer Advocate Service — 90-Day Notice of Deficiency — the deadline chain
• H&R Block · Jackson Hewitt · Omni — CP2000 error rate, back-taxes population, OIC reality
• FTC — press releases · Tax Notes — the tax-relief fraud record and OIC “mills”
• Canopy · CPA Pilot — AI notice tooling aimed at tax professionals

Reconstructed from public sources; claims are tier-labeled (VERIFIED / INFERRED / SPECULATIVE) — see how to read the tiers. Supporting quotes live in this repo’s evidence map (evidence/opp-tax-resolution-evidence-map.md).