We use AI every day in our firm. It reconciles, it flags, it drafts, it summarizes. It has changed how we work, and I'm not going back.
But there's one thing I want every bookkeeper and accountant to hear clearly:
AI takes on the tedious tasks, not the brain power. It is not your intelligence, and it is not your common sense.
Matching hundreds of transactions, sorting through a statement line by line, finding the charge that showed up twice: that's tedious, and AI is great at it. Deciding what a one-sided transfer means for this client, knowing which question to ask, and recognizing when a number doesn't look right: that's brain power. That part is still yours.
AI can be wrong. It can be wrong in a calm, polished, perfectly formatted way that looks finished. The only protection is a person who knows what right looks like and asks the right questions.
So here are the rules we work by, why each one matters, and the drills we use to keep the human skills sharp.
Rule 1: You own the answer
Your name goes on the financial statements. Your name goes on the return. "The AI said so" isn't a defense to a client, and it won't be one to the IRS.
Why: AI does the work faster. It doesn't carry the responsibility. You do.
Rule 2: Know what right looks like before you ask
If you couldn't spot a wrong answer, you aren't ready to use AI for that task yet. You should know, roughly, what a restaurant's food cost should be, what payroll taxes should run on a payroll, and that a cash account shouldn't have a negative balance at month end.
Why: AI fills gaps with something that sounds right. If you don't know the field, you can't tell the difference between sounds right and is right.
Rule 3: Ask the right question
"Is this deductible?" is a weak question. "Is this deductible for an S corporation in Texas, for tax year 2026, on the cash method?" is a real one. The entity type, the state, the year and the method all change the answer.
Why: AI answers the question you asked, not the one you meant. The skill of asking the right question is the accounting skill. That part doesn't get automated.
Rule 4: Confident is not correct
AI writes a wrong answer in the same voice it writes a right one. No hesitation, no "I'm not sure." Tax rules, thresholds and limits change from year to year, and AI may hand you last year's number with total confidence.
Why: Fluent isn't the same as accurate. Before you use a rule, a rate or a threshold, check it against the source: the form instructions or the IRS publication.
Rule 5: Run the common-sense check
Before you accept any AI output, look at it like a seasoned reviewer would:
- Do the totals tie to the statement or the trial balance?
- Are the signs right? Is income showing as income?
- Does this make sense for this client and this month?
- Is the owner's draw sitting in an expense account?
- Did rent get booked 13 times this year?
Why: Most AI mistakes aren't subtle. They're the kind of thing that jumps out at someone who's paying attention. Common sense catches them. AI doesn't have any.
Rule 6: AI flags. A person decides.
Let AI do the line-by-line work: matching transactions, spotting duplicates, finding transfers on one side only. Then a person reviews every flag and decides what it means. (That's exactly how we run our bank recs now: If Your Bank Rec Takes 15 Minutes.)
Why: The matching is the tedious part. The judgment is the brain power. A flag is a question, not an answer. The decision, and the client conversation that usually follows, belongs to a person.
Rule 7: Always have AI review a second time
Once the work is done and a person has made the calls, run it back through AI for a second review. Start a fresh chat, give it the finished work, and ask: "What looks wrong, missing or inconsistent here?"
Especially when it's client facing. Financial statements, a client email explaining a tax change, a year-end planning summary, a monthly report to a business owner. Anything with your firm's name on it that a client will read gets a second AI review before it goes out.
Why: The first pass can miss things, and so can we. Every correction and adjustment is a new chance for an error. A second AI review catches what slipped through, and a person still decides what to do with anything it finds. Once a client reads it, you can't take it back.
Rule 8: Protect client data
No Social Security numbers, EINs or full bank account numbers go into an AI tool. Mask account numbers to the last four digits, and use only tools you've vetted.
Why: Speed is never worth your client's trust.
Rule 9: Keep learning the fundamentals
AI can make a junior faster. It can't make a junior good. They still need to understand debits and credits, why a transfer matters, and which small finding leads to a bigger question.
Why: If you skip the fundamentals, you lose the one skill that makes AI safe to use: knowing when it's wrong.
The drills: how we keep the human skills sharp
Rules are easy to agree with. Drills are what make them stick. Here are five you can start Monday:
- Estimate first. Before you run AI on anything, write down what you expect: roughly what net income should be, about how many transfers, what the payroll taxes should total. If the AI comes back far off, dig in before you accept it.
- Plant an error. Take a practice reconciliation and plant three problems: a transfer on one side only, a duplicate charge and an owner's draw booked as an expense. Run AI on it. Then have your team find them too. Talk about what each side caught.
- Prove it. Any time AI gives you a tax rule, a rate or a threshold, find the source before you use it. No source, no answer.
- Explain the flag. Before anyone clears an AI flag, they explain in their own words why it matters and what they'd ask the client. If they can't, that's the teaching moment.
- Ask it what could be wrong. After AI answers, ask: "What are you assuming? What could make this wrong?" You'll be surprised how often the second answer changes the first.
If you own a small business
Your bookkeeper using AI is a good thing. Ask them one question: "How do you check the AI's work?" A good answer sounds like these rules. If the answer is "it's usually right," that's worth a closer look.
And if you'd like a bookkeeping and tax firm that works this way, grab 15 minutes with me: Book a call
PICK A FREE AI AGENT
Browse our agents, pick the one that fits your work, and it's yours FREE: Pick your FREE agent
On the form, choose the agent number you want. Works with any LLM (ChatGPT, Claude, Gemini, Copilot). And yes, the rules above still apply.
Want agents shaped to your practice?
The Legacy membership, $29 a month (founding rate, first 20 seats), gives you a FREE agent to start, a 15-minute call with me every month, and one custom AI tool built around a real task in your business every month. You don't have to become an AI expert. We have the agents, and we help you mold them to the way you work. See the Legacy membership
- Browse all the AI agents: https://aiaccounting.legacysbc.com/
- Join our FREE community: https://www.facebook.com/groups/1024923770013348
So tell me: what's your rule for checking AI? I'd love to add the best ones to our list. Share yours in the comments.
Things are changing. Be part of it, and benefit from it.
Yvonne






