You've inherited that workbook. Some prior staff member built it — assumption inputs at the top, projections rolling forward, valuation tying out cleanly, even a documentation tab attached. The kind of file nobody wants to touch.
Then halfway down a row, you spot it. A formula that copies the same two cells across every future year instead of rolling forward. No #REF! error. Nothing flagged by Excel. The number still looks clean — but it's been wrong every month, every quarter, every file someone copied from it.
It looks polished. Sometimes makeup just hides something ugly underneath.
AI can be like that, too. It ships polish in seconds, and polish is what we're trained to read as trust. The file looks done long before it's true. That's why AI prompting for accountants matters more than the model release of the week: in CAS work, the file lands harder on you than anywhere else.
The polish proxy is yours to defend
Those files — decks, workbooks, financial models — are the bread and butter of a CAS practice. Management reports. Forecasts. Board packs, year-end financials, and compilations. Advisory deliverables.
These are exactly the artifacts that travel beyond the room they were made in. A spreadsheet gets forwarded. A chart gets reused. A number becomes part of a decision your client makes about hiring, pricing, financing, or selling.
In our profession, the polish itself is what clients are paying for as a proxy for trust. The clean tie-out. The professional formatting. The absence of obvious errors.
When AI delivers all of that in seconds without the discipline underneath it, the proxy breaks. That's not a hypothetical future risk. That's the file you sent on Friday.
Last week's piece on quality control in AI bookkeeping named the role the human still holds when the work commoditises. This piece is about the habit that makes that role real, one prompt at a time.
Ask AI for the spec, not the deliverable
Here's the single most useful prompting habit you can build this year. Stop asking AI for the deliverable on the first prompt. Ask for the specification first.
Wrong prompt: "Build me a Q1 management report from these files."
Right prompt: "I need a Q1 management report built from these files. Before you build it, tell me what sources you'd use, the status of each — current, prior period, draft, estimate — what the report's narrative spine should be, and where you'd need my judgment. Wait for my approval before generating."
Wrong prompt: "Build me a three-year forecast for this client."
Right prompt: "I need a three-year forecast for this client. Before you build it, tell me what assumptions you'd need, where each would come from, what's a fact and what's a judgment call, and the workbook structure you'd use. Wait for my approval before generating."
Same AI. Same files. Dramatically different result. The wrong prompt jumps straight to generation and hopes you'll catch the mistakes in review. The right prompt forces source preparation and specification into the open, where you can correct them before the file exists.
Most practitioners haven't built this AI specification prompt habit. The ones who do find the time they "lose" to the spec round is paid back ten times over in rework they never have to do.
Four decisions hide in every AI workflow
Every AI build in your practice — a one-shot prompt, a saved template, a skill (a packaged, reusable AI workflow with instructions the AI follows every time it runs), an agent (AI that takes a sequence of actions on its own, not just answers one question) — is really four decisions. Source prep. Specification. Generation. Verification.
Most fail because three of the four happen by accident. The habit above forces stages one and two into the open. Once you see all four, the better question becomes obvious: for each stage, what's the right vehicle?
Source prep usually stays human. You know what's current. An LLM doesn't.
Specification is where prompting earns its keep. A good spec prompt is reusable — that's why it eventually becomes a saved template or part of a skill.
Generation is where AI is genuinely strong. Automate it aggressively.
Verification belongs in a separate context. Never the same chat that built the thing — a fresh conversation, sub-agent, or model, architecturally separated from the build.
I run this every year-end with our QBO finalizer skill. The skill reads the QBO trial balance, applies a client-specific account mapping, and generates the financial statements PDF plus the GIFI schedule for T2 import (both GIFI and T2 are Canadian-specific filings).
Source prep is enforced — the skill won't run without the P&L export, the balance sheet export, and the mapping CSV. Specification lives in the mapping itself; every account either has an approved presentation line or stops the workflow until I approve a new one. Generation is deterministic. Verification — my pass over the generated statements — is the only stage I still do by hand, and that's where I want to be.
Four stages. Four decisions. Made deliberately, not by accident.
Champions design AI workflows. Copilot users hope the polish holds.
The habit produces better outputs starting tomorrow. The principle produces better workflows over the next quarter. Together, they produce a practice where AI is used by design — where the question stops being "should I use AI here?" and becomes "how do I decompose this across the four stages?"
That shift separates a Champion from a Copilot user — one designs the work, the other hopes the polish holds.
Pick the next AI build sitting on your desk this week. Before you ask it for the file, ask it for the spec. Are you willing to ship what comes back?
If you want help working out what the four stages look like inside your own CAS practice — which workflows to encode, which to leave human, and where the verification context belongs — book a free consultation.