For enrolled agents and small tax resolution practices
The final notice is dated twelve days ago. Your new client just found it under the mail.
CP2000s to reconcile line by line, four years of transcripts to read, an abatement request for every client who filed late once, a financial statement built from three months of bank statements, and clients who forward a CP504 asking if the house is next. On flat-fee cases every extra hour comes out of your margin, and one missed 30-day window turns a hearing that holds off the levy into one that doesn’t.
The Small Firm AI Kit for Tax Resolution: one Claude Skill with 17 ready-made playbooks.
- CP2000 response letters
- Penalty abatement requests
- Collection financial summaries
- One-time payment
- 30-day money-back guarantee
- Works on free Claude
CP2000 response, Taxpayer A, tax year 2024 (excerpt)
Response due by December 7, 2026, the date printed on the notice. The enclosed response form is marked to agree in part, in the form’s own wording, and signed by the taxpayer.
Item 1, Form 1099-B, Brokerage 1, $18,400.00: the sale was omitted, but the proceeds are not all income. Cost basis is $15,950.00 (Exhibit A, basis statement, p. 2), so the gain is $2,450.00. We agree with $2,450.00 and disagree with $15,950.00.
Checked before you see it
- Checked: Every figure tied to a notice line or an exhibit page
- Checked: Agreed and disputed totals computed with code: $2,662.00 + $22,750.00 = $25,412.00
- Verify: Reply date read off the notice, never assumed from the notice type
Not a tax resolution pro? There’s a kit for your practice:
Compare allIf this sounds like your week, keep reading
- A CP2000 proposes $25,000 of income the client swears was reported, and the cost basis is somewhere in a 40-page brokerage PDF.
- A new client texts a photo of a CP504 at 10 p.m. asking if the IRS is taking the house, and the explanation gets written before anything else the next morning.
- Four years of account transcripts on the desk, and you are circling penalty lines to see whether the three clean years for First-Time Abate are really there.
- Three months of bank statements, two pay stubs and a car loan, and the 433-F still has to be built line by line against the new standards.
Where the hours go
Roughly 15 hours a week go to the same five jobs
The same week with Claude and these playbooks: about 7 hours back
- Today
- With The Small Firm AI Kit for Tax Resolution
Notice responses (CP2000 and balance-due notices)
Today: 4 h/wkWith The Small Firm AI Kit for Tax Resolution: 2 h/wkReading transcripts for new clients (Pro)
Today: 3 h/wkWith The Small Firm AI Kit for Tax Resolution: 1.5 h/wkCollection financial statements and plan or offer packages (Pro)
Today: 3 h/wkWith The Small Firm AI Kit for Tax Resolution: 1.5 h/wkClient letters explaining notices and case status
Today: 3 h/wkWith The Small Firm AI Kit for Tax Resolution: 1.5 h/wkPenalty abatement requests
Today: 2 h/wkWith The Small Firm AI Kit for Tax Resolution: 1.5 h/wk
Today: ≈15 h a week, ≈720 h a year on this work.
Back in your week with Pro, our estimate
≈7 h
With Basic: ≈4 h
Example starting value: $175. Move it to your own number.
- Pro pays for itself after
- ≈34 min
- Time back, each month
- ≈$4,900
How we estimated these numbers
Resolution work is often quoted by the case, so the extra hours come out of your margin, not the client’s bill. This is a conservative picture for a small practice: two notice responses a week, two new clients’ transcripts to read, a collection package every other week, an abatement request most weeks, and the client letters that go with all of them. If you take more collection cases, the financial statement hours grow.
- Notice responses (CP2000 and balance-due notices), 4 h/week: Our estimate: about 2 hours per response to reconcile the notice against the client’s records and write the letter, two a week. No published figure found.
- Reading transcripts for new clients, 3 h/week: Our estimate: about 1.5 hours per new client to read account and wage and income transcripts across years and write the case memo, two new clients a week. No published figure found.
- Collection financial statements and plan or offer packages, 3 h/week: Our estimate: about 6 hours to build a Form 433-F or 433-A summary from bank statements and pay stubs and prepare the payment plan or offer, one every two weeks. No published figure found.
- Client letters explaining notices and case status, 3 h/week: Our estimate: about 36 minutes a working day explaining notices, next steps and case status to clients. No published figure found.
- Penalty abatement requests, 2 h/week: Our estimate, starting from the IRS’s own burden estimate of about 1 hour 34 minutes for Form 843 (Instructions for Form 843, Rev. Dec 2024), plus checking the 3-year history on the transcripts, about one request a week.
These are our estimates, not measured results. No published study times tax resolution tasks, so we started from the conservative hours above and assumed every task still needs you: checking each figure against the notice and the statements, choosing the strategy, clearing every [VERIFY], getting the taxpayer’s signature and signing off yourself, as the IRS OPR’s June 2026 AI guidelines expect. So we cut each task by half at most. For comparison, the IRS itself estimates about 1 hour 34 minutes to complete a Form 843 and about 2 hours for a Form 2848, and in Karbon’s State of AI in Accounting 2026 report firms said AI saves an average of 60 minutes a day per employee.
Week: A standard 40-hour week (our basis). We found no published time study of tax resolution practices, so every figure below is our estimate.. Yearly figures use 48 working weeks, monthly ones 4.
One Skill · 17 playbooks inside
Describe the case. Get the actual draft.
One Skill in Claude: you say what you need in plain words, its router opens the right playbook, and back comes a draft with every figure sourced. These are excerpts of what comes back, built on fictional, de-identified facts. The forms themselves stay in your resolution software.
Every notice dated, explained and answered, with the reconciliation shown.
You ask
“Taxpayer A’s CP2000 for 2024 proposes $25,412 of income. Here are the notice, the basis statement, the Schedule C and the ledger. What do we agree with?”
You get
Each proposed item next to the client’s records, agreed and disputed amounts computed.
Playbook: cp2000-response.md · in every plan
tax-resolution-pros/
├─ SKILL.md
the router: reads your request, opens the right playbook
├─ your profile, saved at setup
└─ playbooks/
├─ citation-guard.md
├─ confidentiality-preflight.md
├─ client-notice-letters.md
└─ … 14 more
See all 17 playbooks
In every planBasic · $49
- Every IRC, IRM and regulation cite flagged before anyone else sees it
citation-guard.md - Taxpayer data de-identified before any work starts
confidentiality-preflight.md - Plain-English letters that tell the client what the notice means
client-notice-letters.md - A record of AI use and verification for every file
ai-use-log.md - Any IRS notice read, dated and explained in a minute
notice-triage.md - A CP2000 response with a reconciliation table
cp2000-response.md - A First-Time Abate request that proves the clean history
first-time-abate-request.md
Added in ProPro · $99
- A case memo from the account transcripts
transcript-analysis.md - A 433-F-ready financial summary from the statements
collection-financial-statement.md - The right payment plan, with the numbers worked out
installment-agreement-planner.md - An offer in compromise workup with the RCP computed
offer-in-compromise-workup.md - A CDP hearing request with the 30-day date computed
cdp-hearing-request.md - A currently-not-collectible request built on the numbers
cnc-hardship-request.md - A reasonable-cause statement tied to the IRM factors
reasonable-cause-statement.md - A deadline log for every open case, with no invented dates
resolution-deadline-log.md - The right authorization, Form 2848 or 8821, filled in correctly
poa-form-check.md - A resolution engagement letter with an AI clause
engagement-letter-ai-clause.md
Why not just ChatGPT?
Why not just paste the notice into ChatGPT?
Most tax professionals already use AI: 90% of the 1,000 US accountants in the 2026 Blue J and CPA.com survey use general platforms such as ChatGPT, Gemini, Perplexity or Claude. But in Karbon’s 2026 report only 21% of firms have an AI policy. A plain chatbot is fast, but it doesn’t know the 2026 rules, it can’t tell you where a figure came from, and the IRS has now said what it expects from practitioners who use one.
Generic AI
- The IRS Office of Professional Responsibility’s Introductory Guidelines for Responsible AI Use in Federal Tax Practice (June 24, 2026) say practitioners must verify AI-produced facts, citations and calculations and cannot rely solely on AI (Circular 230 §10.22), and that firms should document AI use and verification (§10.36). A chat window gives you neither the checks nor the record.
- Older material still says payment plans run 72 months. Current IRS pages describe simple payment plans for individuals who owe $50,000 or less, with no financial statement and up to the collection statute date, normally 10 years from assessment. A chatbot trained on old pages quotes the old rule.
- A CP504 is a notice of intent to levy, but it carries no Collection Due Process rights: those come with the final notice (LT11, Letter 1058 or CP90), and the 30 days start there. Mix the two up in a client letter and the client waits for the wrong deadline.
- The IRS hasn’t said whether entering tax return information into an AI tool is a disclosure under §7216, and its last formal §7216 guidance dates to 2013. The AICPA’s advice is to be safe and get a signed client consent. The civil penalty under §6713 is $250 for each improper disclosure or use, up to $10,000 a year.
The Small Firm AI Kit for Tax Resolution
Why this works where a plain chatbot didn’t
- 1
De-identified before anything starts
Confidentiality Pre-Flight runs first on every task: it swaps names for Taxpayer A and EIN 1, masks Social Security, EIN and account numbers, asks whether a signed §7216 consent is on file, and reminds you which Claude plan settings keep client data out of training.
- 2
Every figure sourced, every total computed with code
CP2000 Response Builder ties each line to the notice and an exhibit page. Monthly disposable income, the reasonable collection potential and payment plan months are computed with code, inputs shown. Anything without a source is marked [NO SOURCE].
- 3
Deadlines read off the notice, and your signature last
Every deadline starts from the date printed on the notice you upload and is marked [VERIFY]. Citation Guard never generates an IRC section or IRM provision. Every output ends with the checks the IRS OPR’s June 2026 AI guidelines ask for: facts, citations and calculations, plus a line for your sign-off.
How it works
Set up once. Then just ask.
No coding, no prompts to copy. You install it the way you would add a document to Claude.
- 1
Download the ZIP
Right after checkout you get one ZIP, ready to download on the page: a single Claude Skill for your practice with every playbook in your plan inside. No account, no login, no subscription.
- 2
Upload it to Claude
Add it once in Claude’s settings. After that you just ask for the work: the Skill reads your request and opens the right playbook, so you never pick a file yourself. It runs on every Claude plan, including Free with code execution turned on. For real client files, use a Team or Enterprise plan, which doesn’t train on your data.
- 3
Answer the 5-minute Practice Profile interview
Your credential, the states you practice in, letterhead and signature block, how you price cases, the notices you see most and whether your clients sign a §7216 consent. Saved once and read by every playbook, so no letter starts without your rules.
Compare your options
What the other options cost
These are the real alternatives a solo or small resolution practice has today, at the prices published or reported for them in November 2026. None of them is wrong, and the kit doesn’t replace your resolution software. They are priced for a different job, or bill you every month.
- Tax resolution software (PitBullTax, IRS Solutions, Canopy)PitBullTax $135 a month for one user on an annual plan; IRS Solutions $378 a month for a 3-user license; Canopy’s Tax Resolution add-on $50 a user a month on top of its plan (vendor pricing pages)Built for forms, transcripts and case management, and good at it, so keep it. The response letters, the abatement narratives and the client explanations around the forms are still your hours.
- Tax AI research assistants (Blue J, CoCounsel Tax Essentials)Blue J $1,498 per user a year; CoCounsel Tax Essentials about $1,300 per user a year (as Blue J’s comparison page reports it)Built to answer tax research questions and billed every year. They don’t reconcile a CP2000 against a basis statement or build a 433-F summary from bank statements.
- An AI tax assistant subscription (CPA Pilot)$29 to $199 a month, by message allowance (vendor pricing page)Research, return review and client emails, capped by messages on the lower plans, and it renews every month.
- Prompt packs for tax professionals$37 to $149, one time (Payhip and Gumroad listings)Lists of prompts: no procedure, no de-identification step, no verification block and nothing computed with code.
- Your own hours, or a staff member’sYour hourly rate, on every notice and every financial statementDone again on every case, and on a flat case fee those hours aren’t billed.
- A generic chatbotFree or a consumer subscriptionNo notice procedure, no exhibit references, no verification, and it will quote a payment plan rule that sounds right. SSNs pasted into a consumer account raise a §7216 question you don’t want.
- The Small Firm AI Kit for Tax Resolution$49 or $99, once$49 or $99, paid once. Even Pro costs less than one month of PitBullTax for a single user, and it never renews.
Pick your plan
One payment. Yours to keep.
Less than one month of tax resolution software, with no renewal after it. One payment, yours to keep.
Pro costs, once
$99
Basic: $49 once
It pays for itself after
≈34 min
of time won back, at $175 for a practitioner hour
Get back just 1 hour a month
$175
a month, against a one-time $99. Our estimate is ≈7 h a week.
Break-even = plan price ÷ the hourly value set in “Where the hours go” above.
Basic
The core kit plus the three notice essentials.
$49one-time
7 ready-made playbooks
Try it on real work. Not for you? Full refund within 30 days.
- 4 core playbooks: Citation Guard, Confidentiality Pre-Flight, Client Notice Letters, AI Use Log
- Notice Triage: what the notice is, where the client stands and the date that matters
- CP2000 Response Builder with a reconciliation table
- First-Time Abate Request, the 3-year history checked against the transcripts
- Practice Profile interview: set your credential, letterhead and house style once
- Verification block at the end of every output
Pro
The full resolution workflow, from the first transcript to the hearing request.
$99one-time
All 17 playbooks
Only $50 more than Basic: +10 playbooks and the Pro bonuses below.
Try it on real work. Not for you? Full refund within 30 days.
- Everything in Basic
- Transcript Analysis with each period’s collection statute date
- Collection Financial Statement, ready for Form 433-F or 433-A
- Installment Agreement Planner with the 2026 limits
- Offer in Compromise Workup with the RCP computed
- CDP Hearing Request with the 30-day date computed
- Currently-Not-Collectible Request
- Reasonable Cause Statement
- Resolution Deadline Log, no invented dates
- POA Form Check: Form 2848 or 8821
- Engagement Letter + AI Clause
- Word templates for your CP2000 response, abatement request and client notice letter
- Firm Style Guide: your house style, applied by every playbook
- Sample outputs from every playbook, to compare against your own
- 12 months of fixes and adjustments to the included playbooks
30-day money-back guarantee
Run the kit on a real, de-identified case. If it doesn’t earn its place in your practice within 30 days, email us and we refund every cent. No conditions.
- One-time payment, no subscription
- Works on free Claude
Run a different kind of practice? CPA firmsTax preparersBookkeepingTell us yours →
No Claude account yet?
No Claude yet? You’re 2 minutes away
Claude is Anthropic’s AI assistant. Its free plan runs The Small Firm AI Kit for Tax Resolution: you create the account, switch on one setting and upload the kit, in about two minutes and without a credit card.
- The free plan works: no subscription and no credit card to start
- About two minutes from sign-up to your first request
- Paid plans only add more usage, for when you want it
Client confidentiality comes first. Every task starts with Confidentiality Pre-Flight, which offers anonymized mode and swaps names for placeholders like Taxpayer A and EIN 1. Try the kit free on anonymized facts, then run real files on a Claude business plan, which doesn’t train on your data.
- 1
Create a free account
Go to claude.ai and sign up with your email, Google or Apple account. It’s free and ready right away.
- 2
Turn on one setting
In Settings → Capabilities, switch on “Code execution and file creation”. Skills need it, and it’s free.
- 3
Upload your package
In Customize → Skills, click “+”, choose “Upload a skill” and select the ZIP you received. Then just ask.
Fair questions
What tax resolution pros ask before buying
Is it allowed to use AI on client matters under Circular 230?
The IRS OPR’s Introductory Guidelines for Responsible AI Use in Federal Tax Practice (June 24, 2026) don’t forbid it. They ask you to verify AI-produced facts, citations and calculations (§10.22), understand the technology you use (§10.35), document AI use and verification (§10.36), not rely blindly on AI for written advice (§10.37), and reflect AI efficiencies fairly in your fees (§10.27, a point the AICPA has pushed back on). The kit gives you a verification block on every document, an AI Use Log line for your file and de-identification by default. It is aligned with the guidelines; it doesn’t certify compliance.
What about §7216?
Section 7216 limits how a preparer uses and discloses tax return information. The IRS hasn’t said whether entering it into an AI tool is a disclosure, and its last formal §7216 guidance dates to 2013. The AICPA’s advice is to be safe and get a signed client consent; the civil penalty under §6713 is $250 per improper disclosure or use, up to $10,000 a year. The kit works in de-identified mode by default, and when identifiable data is needed it asks for a consent in the Rev. Proc. 2013-14 format first (a ready form is in the AI Compliance Kit add-on). We don’t claim that using Claude is §7216-compliant; nobody can yet.
Is my clients’ data safe?
Every task starts with Confidentiality Pre-Flight, which offers de-identified mode, swaps names for placeholders like Taxpayer A and EIN 1 and masks Social Security, EIN and account numbers. On Claude’s Free, Pro and Max plans you choose whether chats are used for training, and you can turn it off in the privacy settings; Team and Enterprise plans don’t train on your data by default. The OPR guidelines point to secure, enterprise-approved tools, so use a Team or Enterprise plan for real client files, and list Claude among the approved tools in your written information security plan, the plan IRS Pub 5708 shows you how to write.
Will it invent IRM cites or deadlines?
Citation Guard never generates an authority: it extracts every IRC, IRM and regulation cite, marks it [VERIFY] and tells you whether it appears in your uploads. Deadlines come from the date printed on the notice, and any date counted from it is marked [VERIFY]. We don’t promise zero errors. We promise every citation, figure and deadline is flagged for you to check.
I already have PitBullTax, IRS Solutions or Canopy. Why would I need this?
Keep it. The kit doesn’t replace your forms, transcript downloads or case management. It works on the writing around them: the CP2000 response and its reconciliation, the abatement request, the reasonable-cause story, the cover letters and the plain-English explanation every worried client needs.
Doesn’t Anthropic already do this?
Anthropic’s public plugin library includes a tax-season organizer that gathers preparation material for a CPA. As far as we found, nothing in it reads IRS notices, drafts abatement requests or builds collection packages. These playbooks fill that gap and run on Claude’s Free plan. We are independent and not affiliated with Anthropic.
Built for your rules
Built for the case in front of you
Tell Claude what kind of case it is, or let Notice Triage read it off the notice, and the playbooks switch forms, rules and the steps they remind you of. Every deadline starts from the date on the notice and stays marked [VERIFY].
Your setup
Notices and underreporting
CP2000s and other proposed adjustments, answered item by item before the date on the notice.
- Response form choice marked in the form’s own wording (agree, or disagree with some or all), signed by the taxpayer
- Form 1040-X marked “CP2000” only when there are other items to correct, as the IRS instructs
- Every disputed line tied to an exhibit, with the exhibit index built for you
- Sending options from the IRS page: Document Upload Tool, fax to the campus on the notice, or mail with proof of mailing
Guardrails
Verification isn’t a disclaimer. It’s the product.
Every output ends with the same verification block, built around what the IRS OPR’s June 2026 AI guidelines ask practitioners to check: the facts, citations and calculations, the source of each, the items marked [VERIFY], a line for your AI use record and a line for your signature. We never promise zero errors. We promise every citation, figure and deadline is flagged for you to check.
Sample verification block
From: CP2000 response, Taxpayer A, tax year 2024 (excerpt)
- Checked: Every figure tied to a notice line or an exhibit page
- Checked: Agreed and disputed totals computed with code: $2,662.00 + $22,750.00 = $25,412.00
- Verify: Reply date read off the notice, never assumed from the notice type
- Flagged: Income ledger page for March not in the file, marked [MISSING]
- 01
Citation Guard
No IRC section, IRM provision, regulation or case is ever generated. Every citation is marked [VERIFY] and checked against your uploads: found in your documents, or NOT FOUND, verify before sending.
- 02
A source on every figure
Every amount carries its notice line, transcript page or exhibit. Anything without one is marked [NO SOURCE], and a missing statement is listed as [MISSING], never filled in.
- 03
Numbers computed with code
CP2000 reconciliations, monthly disposable income, the reasonable collection potential, payment plan months and collection statute dates run as code with the inputs shown. No tax is recomputed unless you supply the figures.
- 04
Deadlines read off the notice
Reply dates come from the date printed on the notice. A deadline counted from a notice date, like the 30 days for a CDP request, uses the letter date, the safer reading, and is marked [VERIFY]. No date is ever assumed from the notice type alone.
- 05
De-identify first
Before any task, de-identified mode is offered: names become Taxpayer A and EIN 1, Social Security, EIN and account numbers are masked. The OPR guidelines say never to upload sensitive data to unsecured sites, and consumer plans are for trying the kit, not for client files.
- 06
§7216 kept in view
The IRS hasn’t ruled on AI tools under §7216. Confidentiality Pre-Flight asks whether a signed consent is on file before identifiable return information is used, and stays in de-identified mode when there isn’t one. We never claim that using Claude is §7216-compliant.
- 07
AI use documented
AI Use Log adds a line to the file for every document: what was AI-assisted, what was checked and who reviewed it, the record the OPR guidelines tie to Circular 230 §10.36.
- 08
Nothing goes to the IRS or the client unreviewed
Responses, requests and client letters carry a “Practitioner review required before sending” banner, and response forms are signed by the taxpayer. Nothing goes out without you.
FAQ
Everything else
Didn’t find your answer? Email us:hello@summonskills.com
What exactly is in the ZIP?
One Claude Skill for your practice. Inside: a router file (SKILL.md) that reads each request and opens the right playbook, one Markdown playbook per job (7 in Basic, 17 in Pro), each with its instructions, the dated IRS rules it relies on and its verification block, and the Practice Profile interview every playbook reads. Pro also includes the Firm Style Guide inside the Skill, plus a bonus pack: Word templates for the CP2000 response, the abatement request and the client notice letter, and sample outputs from every playbook. 12 months of fixes and adjustments to the included playbooks are included (new playbooks for future rule changes are not part of that).
Does it work on Claude’s Free plan?
Yes. The Skill runs on every Claude plan, including Free, as long as code execution is turned on. Free is fine for trying the kit with de-identified or fictional facts. For real client files, use a Team or Enterprise plan, which doesn’t train on your data by default; on Free, Pro or Max, turn training off in the privacy settings.
I use ChatGPT. Will this work there?
No. The kit is built as a Claude Skill, and ChatGPT has no equivalent way to install it once and have it open the right playbook for each request. Claude’s free plan is enough to try it, an account takes a minute, and you can keep ChatGPT for everything else.
Who is it for?
Enrolled agents, CPAs and attorneys who handle IRS notices, collections, penalty abatement, payment plans, offers and CDP hearings. The collection and hearing playbooks assume unlimited representation rights. Annual Filing Season Program participants can represent only clients whose returns they prepared, and only before revenue agents, customer service representatives and the Taxpayer Advocate Service, so for them the notice and letter playbooks are the useful part.
Is it up to date with the 2026 rules?
It is written to them: simple payment plans up to $50,000 for individuals, the Collection Financial Standards effective June 29, 2026 (you load the current tables), the $205 offer fee and the April 2026 offer booklet, and the IRS OPR AI guidelines of June 2026. Every rule in a playbook carries its source and date, so when the IRS changes one you can see what to update, and anything that depends on a table is marked [VERIFY].
Does it pull transcripts or file anything with the IRS?
No. You pull transcripts through the Transcript Delivery System with your 2848 or 8821 on file and upload the PDFs. The kit drafts and organizes; you send responses yourself through the Document Upload Tool, by fax or by mail, and the taxpayer signs what the IRS needs signed.
Does it handle state notices?
The structure works the same way: the notice read, the items reconciled, the response drafted. State deadlines and procedures come only from the rules you enter in your Practice Profile, each marked [VERIFY].
What if it isn’t right for my practice?
Email us within 30 days and we refund the full price. No conditions.
Is this tax advice?
No. The kit is for credentialed practitioners and their staff. It drafts and organizes; you review, decide and sign. You remain responsible under Circular 230 for every response, request and letter that leaves your office.
Your next CP2000, abatement request and 433-F, every figure sourced before it leaves the office.
Notices, transcripts and statements in; every figure tied to its page, totals computed with code, deadlines read off the notice, and your signature last. About 7 hours a week back, by our estimate. $49 or $99, paid once. Use it on a real, de-identified case, and if it doesn’t earn its place within 30 days, we refund every cent. Not tax advice; for credentialed practitioners.
One payment, yours to keep · 30-day money-back guarantee