AI Bookkeeping for Solopreneurs: Where It Gets Tax Wrong

Our previous version of this article told you to paste a CSV into ChatGPT with a prompt that said "Categorize these transactions using these IRS categories: [list]" — and then never gave the list.

That bracket is the whole problem. We went and got the list, and the exercise turned up something more useful than a bookkeeping workflow: five things a solopreneur routinely pays for that an AI will confidently file on Schedule C, and that don't belong there at all. Every one of them errs in the direction that makes your return look better than it is.

The tools below still work and still cost under $30 a month. What's changed is everything around the categorization step — plus three outright errors in the old version, corrected at the bottom.

This is not tax advice. It's a description of where a specific tool fails, with the primary sources so you can check them. Your facts and your state are yours.

The list our old prompt didn't have

Schedule C Part II is a closed list. Every deductible business expense lands on one of these lines, or gets itemized in Part V:

Read line 15 and line 14 again. The IRS printed the two most common miscategorizations onto the form itself, as exclusions. Insurance "other than health". Employee benefits "not yourself".

And notice line 30 sits below the total on line 28. Home office is not one of the operating expense lines. That matters more than it sounds.

The five your AI will get wrong

A language model asked to "categorize business expenses" produces plausible accounting buckets — Software, Subscriptions, Meals & Entertainment, Insurance, Retirement. Most remap onto Schedule C lines fine. These five don't.

1. Health insurance is not a Schedule C expense. It goes on Schedule 1, Part II, line 17, as an adjustment to income. Line 15 of Schedule C literally says "other than health."

Why the IRS is strict about it: on Schedule C it would reduce your income tax and your self-employment tax. On Schedule 1 it reduces income tax only. That's the entire reason for the split, and it's why this error always runs in your favour. Eligibility has three conditions — net self-employment income, no access to an employer-subsidised plan (including through a spouse), and the plan established under your business — and the eligibility test applies month by month.

2. Your own retirement contribution is not line 19. Schedule C line 19 is "Pension and profit-sharing plans", and it is for plans you provide employees. Your own SEP-IRA or Solo 401(k) contribution goes on Schedule 1.

A model matching "retirement contribution" to a line labelled "pension plans" gets this exactly backwards. And the IRS notes the amount itself is not straightforward: your deduction depends on your net earnings, which depend on your deduction — "this is a circular calculation." It points you to Publication 560's worksheets. No model reading a CSV is doing that.

3. An owner's draw is not an expense. A transfer from your business account to your personal account is a distribution of equity. It appears nowhere on the return. You're taxed on net profit regardless of what you withdrew.

This is the easiest error to produce mechanically: a recurring transfer, same amount every month, to an account whose description means nothing. To a model it looks like rent.

4. Home office is line 30, not lines 20b and 25. If you work from home, your rent and electricity aren't operating expenses to be split across the rent and utilities lines — they're inputs to a home-office computation that sits below the total.

The simplified method is $5 per square foot, capped at 300 square feet — a $1,500 maximum — entered directly on Schedule C. The regular method needs Form 8829. Either way it requires regular and exclusive use of that space, and it's limited by the gross income from the business.

5. A business meal is 50% deductible, and entertainment is 0%. Schedule C line 24b is labelled "Deductible meals" for a reason. A model that correctly identifies a client lunch and enters the full amount has still produced a wrong return — this one is arithmetic, not classification.

For completeness in 2026: employee recreational events stay 100%; client meals and travel meals are 50%; entertainment is nil; and from 1 January 2026 meals provided on premises for the employer's convenience became fully nondeductible under IRC §274(o). That last one is irrelevant if you have no employees — but it means any model trained before 2026 will state the old rule.

The prompt, with the list actually in it

Here's the replacement for what we published before. The point isn't the tool; it's that the rules live in the prompt rather than in your memory.

Categorize these transactions to IRS Schedule C lines. Use ONLY these:
8 Advertising | 9 Car and truck | 10 Commissions and fees |
11 Contract labor | 13 Depreciation/Sec179 | 15 Insurance (NOT health) |
16 Interest | 17 Legal and professional | 18 Office expense |
20a/20b Rent or lease | 21 Repairs | 22 Supplies | 23 Taxes and licenses |
24a Travel | 24b Meals (50% only) | 25 Utilities | 27a Other (name it)

Do NOT assign these to a Schedule C line. Put them in a separate
EXCEPTIONS list instead:
- Health insurance premiums (Schedule 1, not Schedule C)
- My own SEP/Solo 401k contributions (Schedule 1, not line 19)
- Transfers to my personal account (owner's draw, not an expense)
- Rent, power, internet if I work from home (line 30 computation)
- Anything that looks personal, or mixed personal/business

For each transaction give: line number, confidence (high/med/low),
and one sentence of reasoning. For meals, note that only 50% is
deductible. Never guess a business purpose I did not give you.

Two things make this better than what we had. It names the lines, so the output maps onto the form. And it creates an exceptions list — the transactions the model is instructed not to categorize — which is where the expensive errors were hiding.

Where "approximately right" is worth nothing

There's a doctrine called the Cohan rule, from a 1930 case about a Broadway composer who claimed travel and entertainment expenses without records. It lets a court estimate a deduction where you can prove the expense happened but can't document the exact amount.

It's weaker than people think. It's permissive, not mandatory — courts may estimate and the IRS has no obligation to accept one. It goes to amount, not existence: you must first prove the expense occurred. And the Tax Court has refused it where a taxpayer could have substantiated and simply didn't.

Then there's the part that matters here. Section 274(d) of the tax code supersedes Cohan entirely for travel, meals, gifts and vehicles. The regulations say no deduction "shall be allowed a taxpayer on the basis of such approximations." For those categories you need amount, time, place, business purpose, and the business relationship of anyone else involved.

So the overlap is exact and unlucky: the expense types a model is most likely to mishandle are precisely the ones where being roughly right buys you nothing. A CPA firm we read put it more bluntly than we would: "Keep good records, please. Do not rely on the Cohan rule."

Practically: documentary evidence is generally expected at $75 and above for those categories, lodging away from home always needs it regardless of amount, and business gifts are capped at $25 per recipient per year.

What a CPA review actually buys you

Our old version said to keep a $200-400 year-end CPA review "because that's where human judgment still wins." That's true and incomplete, and the gap matters once AI is doing your categorization.

If the IRS proposes an accuracy-related penalty — 20% of the underpayment under §6662, which applies to negligence or to a substantial understatement (for individuals, the greater of 10% of the tax due or $5,000) — you can defend on reasonable cause by showing you relied on a professional. The Tax Court's test has three prongs:

1. The adviser was competent and had sufficient expertise to justify reliance
2. You provided necessary and accurate information to the adviser
3. You actually relied in good faith on their judgment

Prong two is where AI bookkeeping lives. If your books were categorized by a model and you never checked them, it is genuinely arguable you didn't provide accurate information. And good faith doesn't mean handing over a shoebox and signing what comes back — it means reading the return and asking about anything unfamiliar or suspiciously favourable.

The limit on top of that: reliance abates penalties, not tax. In the case the test comes from, the taxpayer escaped the penalty and still owed the money.

None of that is a reason to skip the review. It's a reason to do the part of the review that AI makes easiest to skip.

Your return has a shape, and AI changes it

The IRS's first screen is an automated score — the Discriminant Information Function — which, as one accounting platform describes it, "assigns a numeric score to every return based on how likely it is to result in a tax adjustment." It works by comparison: "The IRS compares your deductions to averages for taxpayers in your income bracket and industry."

The example given is a sole proprietor earning $75,000 claiming $25,000 in vehicle expenses and $15,000 in travel — flagged not because any single item is wrong but because the proportions are unusual.

Here's the connection nobody draws. An AI that dumps every ambiguous transaction into one line — "Office expense", or an unlabelled "Other" — doesn't just misfile individual items. It reshapes the distribution the screening model is scoring. And Part V won't take a lump sum anyway: you must itemize each other-expense with a description and amount. "Miscellaneous, $4,200" is the worst possible entry.

This matters more for you than for a company. The GAO reports that the IRS attributes roughly $80 billion in unpaid tax a year to sole proprietors underreporting income — about 16% of the entire $496 billion tax gap. The structural reason is plain: your income isn't withheld and only some of it is reported to the IRS by third parties. Schedule C is where the IRS knows least, so it looks hardest.

The 2026 number that proves the point

If you want a ten-second demonstration of why AI is unreliable on tax facts, use this one.

The IRS set the 2026 business standard mileage rate at 72.5 cents per mile from 1 January. Then on 13 July, after fuel prices rose 38%, it raised the rate to 76 cents for 1 July through 31 December, in Announcement 2026-11. It was the first mid-year adjustment since 2022.

So a correct 2026 mileage deduction needs two rates split by date. A model trained before July 2026 will apply 72.5 to the whole year with total confidence. One trained on 2025 data will use 70.

That's the whole argument in one figure you can verify yourself.

The stack, corrected

Wave (free, or Pro $19/month, $190/year) is still the foundation. Starter gives unlimited invoices, estimates, bills and bookkeeping records at $0. Pro adds automatic bank import, automatic categorization, unlimited receipt capture and late-payment reminders. Receipt scanning is otherwise an $11/month add-on ($96/year).

One oddity worth flagging: Wave's pricing page describes Pro as including unlimited receipt capture and separately lists a discounted Pro rate for the receipt add-on ($8/month, $72/year). Both can't be straightforwardly true, and we're not going to invent an explanation — check what your account actually shows before budgeting.

Payment processing is 2.9% + $0.60 (3.4% + $0.60 Amex). Payroll is separate, from $25/month.

QuickBooks Self-Employed is gone, and we were still recommending it. Intuit closed it to new customers and pulled the mobile apps. Its own comparison page says plainly: "New customers will be able to choose from our other products, including QuickBooks Solopreneur." Existing subscriptions still run; no new features are being added.

The replacement, QuickBooks Solopreneur, is $20/month or $215/year, and includes automatic separation of business and personal transactions plus Schedule C filing. If you're migrating from Self-Employed, note that reports of the automatic transfer are mixed and that receipts and past invoices don't come across — re-upload before you cancel anything.

Make (Core: $9/month billed annually, $10.59 monthly) still ties it together: Stripe charge to Wave entry, monthly export with the categorization prompt pre-filled, quarterly P&L for your estimated tax. We covered the patterns in our client follow-up automation guide.

ChatGPT's free tier handles the categorization step. Since 6 August 2026 it has unlimited text chats, and this is a text task — we covered the change here. Don't subscribe for this.

Realistic total: $9/month billed annually ($10.59 monthly) if you run Wave Starter and a free AI tier. With Wave Pro and a $20-a-month assistant such as ChatGPT Plus, it's about $50 a month billed monthly, or about $45 with Make and Wave on annual billing. (We previously put the top of the range at $48. That added Make's annual price to monthly prices for everything else.)

On what that replaces: we previously said "$150-300/month" for a bookkeeper with no source. What can actually be sourced: Wave's own bookkeeping service, Wave Advisors, starts at $149-199/month, and Intuit's own guide puts subscription bookkeeping services at $300-$1,500/month and the median bookkeeper rate near $24/hour. The honest statement is that the range is wide and our old figure sat at the bottom of it.

Three things this article had wrong

The 1099 threshold is $2,000 now, not $600. We told you to issue a 1099 to any contractor you paid $600 or more. The One Big Beautiful Bill Act raised the Form 1099-NEC and 1099-MISC threshold to $2,000 for payments made after 31 December 2025 — that is, for the tax year this article is about. The $600 figure had stood since 1954. Separately, the 1099-K threshold went back to $20,000 and 200 transactions. The income is taxable either way; only the reporting duty moved.

We withdrew Dext and then kept using it. We correctly pulled Dext as the receipt pick — it's $31.50/month with a five-user minimum, an accounting-practice product, the wrong shape for one person. Then it stayed in the automation examples, the stack table, the monthly routine and the audit section. All four now say Wave.

Two figures were invented. "Most solopreneurs lose 15-25% of their legitimate deductions" had no source. Neither did "mileage tracking + quarterly estimates often save more than $240/year" — a number that happened to equal the annual cost of the product it was justifying. Both are gone.

What still needs a human

Unchanged from before, and reinforced by everything above: entity structure, retirement plan choice, state-specific rules, audit defence, and anything where the answer depends on facts a bank feed doesn't contain.

Add one more. Bookkeeping quality is itself evidence. If you ever have to show a profit motive — the hobby-loss question, which turns partly on "the manner in which the taxpayer carries on the activity" — businesslike records are the argument. That's a reason to keep good books that has nothing to do with saving time.

For the wider stack see the 2026 solopreneur AI stack and what a solo AI stack really costs. The same pattern we found for licensed real estate agents applies here: the tool doesn't carry the liability, you do.

---

Tax material quoted from primary sources where they could be opened on 4 September 2026 — Schedule C and its related IRS pages, the IRS tangible property regulations, the IRS retirement plan guidance, 26 U.S.C. §6662, the IRS mileage announcements and the GAO tax gap report — and attributed to secondary sources where they could not. Prices re-verified at each vendor's own page. This is not tax advice; your facts, your entity and your state govern, and the rules cited change. Links marked "Paid link" earn us a commission if you sign up through them, at no extra cost to you; no other link here pays us — how we make money.