AI for Accountants: Post-Tax-Season Reset for Solo CPAs

A CPA posted a story on X this week that every solo practitioner should read. A new client walked in with a 50-page retirement plan he’d generated himself. Roth conversions mapped out, withdrawal sequences, tax-loss harvesting schedules, the works. Confident, well-formatted, completely personalized to him. He’d been making real financial decisions based on it for seven months.

Every section looked right. Every section had something materially wrong. The withdrawal sequencing would trigger IRMAA surcharges every single year. The RMD projections used the wrong life-expectancy factors. And the whole thing was produced by a generic AI chatbot he’d trusted because it spoke his language.

The line in the CPA’s post that’s been rattling around tax-advisor circles all week: “The most dangerous financial advice isn’t obviously bad. It’s confident, well-formatted, completely personalized to your situation — and wrong in ways you have no way of knowing.”

That’s the backdrop to this blog. You’ve just closed another tax season. Your inbox is finally quiet. Your phone isn’t ringing at 9 PM. You’ve got a four-to-six-week breathing window before summer bookkeeping catches up to you, and you’re asking the same question half the solo CPAs on accounting Twitter are asking: where does AI actually help me, and where is it a liability?

The short answer: don’t use generic AI for compliance. Use it for everything else.

The Compliance Ceiling

Here’s the number that needs to anchor every AI decision you make this year.

When Column Tax ran their TaxCalcBench benchmark — a stress test of frontier language models on preparing full federal individual returns under simplified conditions — the best standalone model they tested hit 32.35% accuracy. That was Gemini 2.5 Pro. Claude Opus 4 came in at 27.45%. GPT-5 augmented with web search got to 41.67%. A specialized multi-agent tax system (Filed) got up to 72.5% strict accuracy. No generic chatbot, asked to prepare a return, gets more than about four in ten right (Filed / Column Tax benchmark).

For compliance — the actual return prep — that number makes the decision for you. You can’t use a tool that hallucinates depreciation schedules, picks the wrong form for a 338(h)(10), or swaps life-expectancy factors on RMDs. Your license is on the line. Your clients are on the line. The benchmark is doing you a favor by being public.

A practicing CPA summarized the day-to-day version of this recently: AI got a K-1 reconciliation wrong three times before getting it right. Another posts that in every daily Claude conversation she has, at least once it makes a mistake on a tax question. These aren’t edge cases. They’re the baseline.

So what’s left?

Plenty. Because compliance is only part of what a solo CPA does all year. The other part — the part that scales, that differentiates you, that actually compounds your book — is advisory. And advisory is where the AI-for-accountants story gets interesting.

The Advisory Arbitrage

Thomson Reuters surveyed tax and accounting professionals across 2025 and found 79% expect AI to have a high or transformational impact on the profession, while only 14% of firms have a defined AI strategy (Thomson Reuters). That gap is where individual solo practitioners have an asymmetric opportunity right now — the big firms are figuring out strategy committees. You can be running live workflows this week.

Sage’s own data puts 46% of accountants using AI daily in 2026, up from 18% in 2023 (Accounting AI Statistics 2026). CPA.com reported 58% of firms are actively piloting AI. But an AICPA data summary puts solo practitioners at only about 34% regular use — meaning two out of three solos are leaving the advisory arbitrage on the table, often because the first thing they tried was to use AI for return prep and got burned.

Here’s the frame that works: a practicing CPA put it almost exactly this way in February — “Most of us actually WANT the reporting part of our job to get replaced by AI so we can focus on advisory. We’re not scared of AI. We’re waiting for a product that can do it right.” That’s the arbitrage. Compliance isn’t ready. Advisory is. And advisory is where the margin is anyway.

The Four Post-Tax-Season Workflows

You have four weeks. Maybe six before you lose them to summer extensions and Q2 bookkeeping. Here are the four workflows to reset before you lose the window.

All four are built around the same rule: no real client PII goes into the prompt. No names, no SSNs, no EINs, no full addresses. Describe the client structure; don’t identify the client. Your facility’s enterprise AI (Microsoft Copilot with a BAA, ChatGPT Enterprise, or Claude Team with DPA) is different — those can handle PII if your firm agreement permits. The workflows below assume consumer-tier tools, which most solos use.

Workflow 1: The Client Meeting Prep from Last Year’s Return

Every client meeting is better when you walk in with a summary of what actually happened on their return. Most solos re-skim the 1040 five minutes before the meeting. The AI version:

I'm a solo CPA prepping a client-meeting summary. Below is a
de-identified summary of a 1040 return I filed for a client
(age 58, MFJ, two dependents in college, owns an S-corp
professional-services firm, rental property in a second state).

Give me a 1-page meeting-prep document:
1. 4-bullet recap of the return's tax picture (plain English)
2. 3 concrete advisory questions I should ask them this meeting
   (e.g. "did your S-corp salary still make sense this year?")
3. 3 opportunities to suggest they haven't done yet (SEP, Roth
   conversion window, HSA, 529, backdoor IRA)
4. One thing to flag that could bite them next year if nothing
   changes (e.g. estimated payment shortfall, IRMAA threshold,
   AMT exposure)

Keep it to one page. Warm, direct, no tax-code jargon.

--- RETURN SUMMARY (de-identified) ---
[paste your de-identified summary: filing status, income
sources and ranges, major deductions, credits, state, any
unusual items]

Walk in with that. The conversation goes from “did we file everything” to “here are three things I’ve been thinking about for you.” That’s the billable hour in advisory.

Workflow 2: The Monthly Variance Memo from a QuickBooks Export

Bookkeeping clients pay the same fee every month. The ones who stay are the ones who get real insight, not just a reconciled P&L. A monthly variance memo — if it lands in their inbox before they ask — is worth more than the reconciliation itself.

I'm writing a monthly variance memo for a bookkeeping client
(de-identified: a retail business, ~$60K monthly revenue,
small team). Below is a CSV export of this month's P&L
compared to last month and the prior-year-same-month.

Write me a 4-paragraph variance memo:
Paragraph 1: The one-line summary (is this month ahead or
behind, by how much, and why in plain English)
Paragraph 2: 2-3 line items that moved materially, with
specific percentages, and one likely business-side reason for
each
Paragraph 3: One thing to watch heading into next month based
on the trend
Paragraph 4: Two questions I should bring to our next call

Write it in my voice — direct, helpful, not alarmist. Avoid
accounting jargon. Treat the reader as a small-business owner
who is smart but not trained in finance.

--- P&L DATA ---
[paste the CSV text, with the client name removed from headers]

Do that once, and you’ve built a template you’ll reuse every month for every bookkeeping client. The first time it takes 15 minutes. The second takes 5.

Workflow 3: The Late-Payer Letter That Preserves the Relationship

Every solo CPA has the same pile in April and May: invoices 60 and 90 days past due. The letters most practitioners send are either too stiff (they damage the relationship) or too soft (they don’t work). Here’s the prompt:

I'm a solo CPA and a client is 72 days past due on a $3,400
invoice for their 1040 preparation. I don't want to damage the
long-term relationship, but I do need to get paid. Write me a
single email, under 150 words, that:

1. Opens warmly (we've worked together for years)
2. References the specific invoice clearly without being
   aggressive about the age
3. Offers two specific paths to resolve (pay in full, or
   two-installment plan ending by [date])
4. Includes one light reason it matters ("I'm clearing the
   books before summer")
5. Ends with a direct, professional ask

No guilt. No threats. No "per my last email." I want the tone
of a trusted professional reminding a friend, not a collector.

Send that email, then send a duplicate to yourself. You’ll have a template that works across every stuck invoice, and you’ll save the mental weight of figuring out the tone each time.

Workflow 4: The Scenario Modeling Walkthrough for an Advisory Conversation

This is where GPT-5.5’s dense-document strength — the capability OpenAI shipped on April 23 — actually changes the work. You can paste a complex multi-entity scenario and get back a cleanly-structured client-facing walkthrough.

I'm a solo CPA preparing for a tax-strategy conversation with
a client (de-identified: owns an S-corp, spouse is W-2,
household MAGI around $310K). They've asked about converting
$80K of a Traditional IRA to a Roth this year.

Walk me through a scenario comparison:
1. If they convert $80K in 2026 at their current marginal rate
2. If they convert $40K in 2026 and $40K in 2027
3. If they wait entirely

For each scenario, estimate:
- Additional federal tax this year
- Medicare IRMAA implications if any
- ACA subsidy implications if relevant
- 5-year projected Roth balance (assume 7% average return)
- The one-sentence "why this might be right for them"
- The one-sentence "why this might be wrong for them"

Include any assumptions you're making. Flag which numbers I
need to verify against current tax tables before I present this.
Finish with 3 questions I should ask the client before
finalizing any recommendation.

Critical note: the numbers in the output are a draft. You verify every tax-table figure against the actual current tables before presenting anything. The AI writes a structure. You confirm the math. That’s the division of labor that protects your license.

Why Not Just Use Juno?

Juno raised a $12M seed on April 10 from Bonfire VC on exactly the thesis that AI can replace 90% of the tax-prep busy work. Their product is built for SMB accounting firms — not self-prep consumers, not BigFour. Early numbers cited in the funding-round announcement: ~500 customers in year one, mid-seven-figure ARR in eight months, processes that typically take 2-3 hours shrunk to 7-10 minutes.

That’s real. Juno and similar specialized tax AI systems (Filed, Blue Dot, Canopy’s AI) consistently beat generic chatbots on benchmarks — Filed hit 72.5% strict accuracy on TaxCalcBench versus Gemini 2.5 Pro’s 32.35%. If your firm’s ready to budget $200-500/month per seat for specialized tax AI, you should.

What a solo CPA can replicate DIY, with GPT-5.5 or Claude or Gemini in the advisory workflows above, is about 60% of Juno’s value for $20 a month. The 40% you’re giving up is the automated calculation side — the part where hallucinations matter. That’s the honest split.

Use specialized products for compliance. Use general-purpose AI for everything around it. Don’t pretend they’re interchangeable.

What AI Can and Can’t Do for a Solo CPA Right Now

Being honest about boundaries is the whole point.

It can: summarize last year’s return into an advisory-ready brief, draft variance memos, write client emails in your voice, generate scenario comparisons for you to verify, explain tax concepts to clients in plain language, draft intake forms, rewrite policy letters, produce first-draft operating-agreement language for you to refine with an attorney, translate complex tax scenarios into presentation decks, and in general handle the 60% of your week that is writing.

It can’t: prepare a correct 1040 (only 23-42% accuracy on generic models), apply current-year tax tables reliably without verification, catch jurisdictional nuances (state rules, local adders, IRC conformity), tell you what changed in the tax code this year without a dated citation, or substitute for your professional judgment on ambiguous facts.

It will: confidently give you wrong answers in the exact same tone it uses to give right ones. That’s why a practicing CPA needs to stay in the loop on every client-facing output. You are the last line, always.

Which Model for Which Job

Quick field guide from CPAs actually using these daily:

Task	Best tool	Why
Scenario walkthroughs, dense-document parsing	GPT-5.5 (via ChatGPT Plus)	April 23 launch is strong on long financial docs
Research, explaining concepts, teaching clients	Claude Opus 4.7	Polished writing, careful reasoning, better bedside manner
Tax-code research + citation lookup	GPT-5 with web search	Only tool in this tier with current sources
Quick back-office drafting (intake forms, emails)	Any of the above	The quality gap is small here
Actual return calculation	None of the above	Use Intuit ProConnect, Drake, Lacerte, or CCH

A note on the model choice: one CPA on X in January flagged that Claude 4.5 was specifically weak on tax and legal advice versus GPT-5.2 Pro and Gemini 3. The models improve on a weekly cadence — what’s true this month may not hold in August. Retest quarterly.

What This Means for You

If you’re a solo practitioner: You have a four-week window while client phone calls stay quiet. One hour with Workflow 1 gives you a meeting-prep template. One hour with Workflow 2 gives you a variance memo template. Two more hours with 3 and 4, and you’ve built the four templates you’ll use for every client for the rest of the year. The ROI in retainer-ready advisory work is clear.

If you’re a bookkeeper: Workflow 2 alone justifies the effort. A monthly variance memo, delivered before the client asks, changes bookkeeping from commodity to advisory. Your pricing can follow.

If you run a 3-5 person firm: Build the templates as a team. Have one person draft, one review, one integrate into your firm’s voice. The templates become firm IP. Don’t let the solo practitioner down the street move first.

If you’re a new advisor: Learn the boundary first. Use AI for the advisory workflows above. Do NOT use it to prepare a return. Keep your compliance stack professional (ProConnect, Drake, Lacerte) and your advisory stack AI-assisted. The best new advisors you’ll compete with in three years are doing exactly this split.

If you’re skeptical: Fair. The TaxCalcBench numbers are real — no generic AI should be touching compliance right now. But the advisory work is where the margin is, and that’s exactly where the tools are strongest. The CPA whose client made 7 months of wrong decisions isn’t a story about AI being useless. It’s a story about the client skipping the advisor. That advisor — you — is the whole point.

The Bottom Line

The next tax season is 9 months away. The next four weeks are the most concentrated advisory-template-building opportunity you’ll have all year. AI is not your CPA — nothing in TaxCalcBench or any real-world benchmark suggests otherwise. But it is the sharpest advisory assistant solo practitioners have ever had access to, and it costs $20 a month.

Build the four templates this week. Save them where your future self will find them. See which three clients would benefit from a proactive touchpoint before summer. Run the meeting-prep prompt on their returns tonight. Watch what happens to your retention and your referrals by September.

Want the full advisory system — all four prompts above plus the 6 more that close the loop (roll-forward memo, year-end planning brief, FBAR/FinCEN reminder letter, client onboarding packet, quarterly CFO-lite report for your best bookkeeping clients, and the referral-request template) plus the de-identification rules in a printable checklist? The AI for Accountants Workshop is built around exactly this post-tax-season reset. Two hours. Every prompt tested against the workflows above. If you’ve been meaning to pivot from pure compliance to an advisory mix, this is the four-week window the course is designed for.

Sources: