A firm partner I know paid a boutique consultant $4,000 to review a new client’s financials and write the engagement risk memo. It came back three pages long. It said “moderate risk, standard procedures apply.” It could have described any client at any firm.
I ran a version of the same exercise with Claude in under 20 minutes. I’m not the one who makes the risk call, that’s not my lane. But I noticed something about how the AI got asked the question, and it’s the same thing that was wrong with the $4,000 memo.
Most people prompt AI like they’re asking a coworker a quick question. This month, Anthropic’s own guidance on prompt engineering said the fix plainly: give the model a role, a required reasoning path and a defined output format, or you get the same generic guess a search engine gives you.
Anthropic’s prompt engineering guidance is specific about what separates a sharp answer from a template one. A vague request gets a vague structure back. A prompt that assigns a role, forces the model to reason step by step, and demands a specific output shape gets something closer to what a senior reviewer would actually hand you.
The Journal of Accountancy found the same pattern from the other direction. In a study of 277 accountants, the ones using structured AI workflows for client advisory work closed books 7.5 days faster and saw a 55% jump in weekly client support. Not because the AI got smarter. Because the requests got sharper.
Read Anthropic’s prompt engineering guidance → https://claude.com/blog/best-practices-for-prompt-engineering
Wrong read: “AI just needs a better prompt template. Copy, paste, done.”
Right read: a template gets you the same shape every time. What changes the output is forcing the model to commit to a conclusion and name a stopping point, the same discipline a reviewing partner applies without thinking about it. A generic prompt gets you description. A structured one gets you a decision, with a built-in answer to “when do I walk this back.”
That last part is the piece almost nobody adds. Ask AI for a risk assessment and it’ll describe risk all day. Ask it to also name the exact threshold where the engagement gets declined or escalated, and it has to take a position.
Here’s what’s on the line if you skip that step:
Client advisory memos that read like boilerplate and don’t build the trust they’re supposed to build
Engagement risk assessments that describe a problem but never name what would make you walk away from it
Firm decisions that stall because the AI output explains the situation but never recommends the call
Staff taking a first-draft AI answer at face value because nothing in the prompt forced a second layer of reasoning
Don’t: paste a request and accept the first draft, whatever role you happened to type at the top.
Do: assign a specific role, force a structure with named sections and require a stopping condition, the point where the answer changes or the engagement gets a second look.
ACTIONABLE TOOL: Five Prompts Built Around Role, Structure, and a Stopping Point
Copy any of these, fill in your details, run it.
Variance Narrative Memo
You are a senior controller reviewing this month’s variance report for CLIENT/DEPARTMENT.
Output: what moved and why, not just that it moved. Which variance is noise versus signal. One recommended action. The specific number that would change the recommendation.
Engagement Risk Memo
You are an audit partner scoping a new client, CLIENT TYPE/INDUSTRY.
Output: the three risk factors specific to this client, not generic industry risk. The procedures each one actually requires. The exact condition that would move this from “standard” to “decline.”
Client Advisory Memo
You are a fractional CFO preparing a quarterly review for CLIENT.
Output: the plain-English story behind the numbers. Two scenarios for next quarter. One clear recommendation. The metric that would make you reverse it.
New Service Pricing Memo
You are a practice growth partner scoping a new advisory offering, SERVICE.
Output: who it’s actually for. What it costs you to deliver. Three pricing tiers with the reasoning behind each. The client profile you’d turn down even at full price.
AI Vendor Evaluation Memo
You are your firm’s technology reviewer evaluating a new AI tool’s claim, VENDOR CLAIM, e.g. “99% accuracy”.
Output: what the claim actually measures versus what it implies. The one question the vendor’s demo didn’t answer. The specific test you’d run before trusting it with a real client file.
If a section doesn’t fit your situation, tell the model to skip it. If the first pass is good but not sharp, ask it to go deeper on the one section that matters most, not all of them.
5 AI Prompts That Make Your Tax Research Defensible
Same mechanic, applied to one specific job: turning a quick tax question into something you could defend later if the IRS asked.
Read → https://nexairi.com/article/Accounting/5-ai-prompts-tax-research-client-questions-cpa-2026/
Prompt engineering best practices for 2026
Anthropic’s own guidance on what actually separates a sharp AI answer from a generic one, straight from the people building the model.
Source → https://claude.com/blog/best-practices-for-prompt-engineering
How are finance teams really using AI and automation?
The Journal of Accountancy’s study of 277 accountants found structured AI workflows closed books 7.5 days faster and lifted client support 55%, the real-world version of the prompting discipline above.
The 10-minute advisory prep workflow accountants are using with AI
Karbon’s look at how firms are actually structuring AI requests before a client meeting, worth comparing against your own habits.
Source → https://karbonhq.com/resources/ai-accounting-advisory-prep/
The 2026 AI Index Report
Stanford HAI’s read on where AI stands this year: less evangelism, more evaluation. The exact shift a stopping-condition prompt forces onto a single AI answer.
Source → https://hai.stanford.edu/ai-index/2026-ai-index-report
The five prompts above are a start. The free checklist below covers where AI actually fits into finance workflows without breaking your controls.
Get the checklist → https://nexairi.com/downloads/finance-ai-workflows-controls
P.S. If your team starts feeding real client financials into prompts like these, the next question is what data policy governs that. The CPA AI Policy Kit is $29 and covers exactly that gap.
No posts
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.