Module 5 — The Workflow Prompt Library, Part Two

This module completes the library with five more workflows, including the cost segregation anchor worked at length and the citation-discipline workflow that is the profession's single most important defense against fabricated authority. The structure is unchanged: what it is and where AI helps, a copy-paste prompt, what good output looks like, and the review step. De-identify first; review always.

Workflow 6 — Trial-balance flux and variance analysis narratives

The workflow and where AI helps. Flux analysis — explaining period-over-period movement in accounts — has two parts: computing the changes, which a spreadsheet does exactly, and writing the narrative that explains them, which is slow to do well by hand. AI is strong at the narrative layer: given the changes and a little business context, it drafts a readable explanation of what moved and offers plausible drivers to investigate. Let the spreadsheet compute; let the model narrate; let you verify the drivers.

The prompt.

You are drafting a variance analysis narrative for internal review. Below is
a comparison of account balances between two periods, with dollar and percent
changes already calculated. Write a concise narrative that (1) highlights the
most significant changes by dollar amount and by percentage, and (2) for each,
offers a plausible business driver to investigate, clearly framed as a
hypothesis, not a conclusion.

Business context: [e.g., a growing HVAC contractor; added a second crew in Q2;
raw material prices rose]
Materiality guidance: focus on changes over [$X] or over [Y]%.

Data:
[paste the de-identified two-period comparison with changes computed]

Format the narrative by account category. Do not state a cause as fact —
use "may reflect," "possibly due to," and note where a change is unexplained
by the context I gave and should be investigated.

What good output looks like. A tidy, category-organized narrative that surfaces the material movements, ties each to a plausible driver consistent with the context you supplied, hedges appropriately ("may reflect"), and explicitly flags movements the context does not explain as items to investigate. It should read like a draft you would refine, not a final signed analysis.

Review and guardrails. The computed changes came from your spreadsheet, so the numbers are trustworthy; the explanations are the model's hypotheses and must be confirmed against reality before they appear in any deliverable. A model will happily propose a confident-sounding driver that is simply wrong for this client — that is the thing to catch. This is a good mid-tier-model task; the narrative quality is fine below frontier, and the human step is validating the drivers, not fixing the prose.

Workflow 7 — Fixed-asset and cost-segregation intake triage (the anchor, worked in full)

This is the course's central worked example, so we go deeper. It shows AI doing exactly what it is good at — reading messy source material and producing a structured first pass — while staying firmly inside the boundary that makes it safe.

The workflow and where AI helps. A cost segregation study reclassifies parts of a building's cost into shorter depreciable lives (five, seven, and fifteen years) rather than the long default recovery period, accelerating deductions. It begins with a fixed-asset list or construction cost detail that must be read, sorted into recovery-period buckets, and reconciled to a total. That reading and first-pass sorting is language-heavy transformation on material sitting right in front of the model — the model's strength. AI can propose a preliminary bucketing for a professional to verify, draft the client data-request for the missing pieces, and flag what a real study still needs. What it produces is a triage that speeds the front of the process.

The boundary, stated before the prompt because it governs the whole workflow. The AI's classifications are preliminary and are not filed. A cost segregation study that will support a tax position requires a full engineering-based study performed by qualified professionals, which involves site inspection, engineering cost estimation, and analysis the model cannot do from a spreadsheet. The AI triage is a preliminary aid, reviewed by a CPA, that helps scope the engagement and prepare the request — never the study itself. Keep that fixed as you read what follows.

The prompt.

You are assisting an accountant with a PRELIMINARY triage of a fixed-asset
list ahead of a possible cost segregation study. This is a preliminary
estimate for a professional to review, NOT a final classification and NOT a
substitute for a full engineering-based cost segregation study.

For each asset below, propose a likely recovery-period bucket from: 5-year,
7-year, 15-year (land improvements), 27.5-year (residential structure),
39-year (commercial structure), or "Needs review." Base the proposal only on
the description and any detail given. For each asset, give: the proposed
bucket, a one-line reason, and a confidence level (High/Medium/Low).

Rules:
- If a description is too vague to classify, use "Needs review" and say what
  detail would resolve it. Do NOT guess into a specific bucket on thin
  information.
- Do not invent asset costs or descriptions. Use only what is provided.
- Group land improvements (paving, landscaping, site utilities) as 15-year
  candidates; group tangible personal property serving the business (certain
  fixtures, specialty electrical/plumbing tied to equipment) as 5- or 7-year
  candidates; group structural components (roof, framing, standard building
  systems) as 27.5- or 39-year candidates per the property type given.

Property type: [e.g., commercial retail / residential multifamily]
Asset list:
[paste the de-identified asset list with descriptions and amounts]

After the table, produce two things:
1. A subtotal of proposed cost by bucket, and the grand total, so I can check
   it foots to the asset list total.
2. A "Data request" section: a short list of documents and clarifications a
   real study would need that are missing here (e.g., construction cost
   detail, invoices, blueprints, a breakdown of lump-sum line items).

What good output looks like. A per-asset table with a proposed bucket, a short reason, and a confidence flag; liberal use of "Needs review" for vague lines rather than false precision; bucket subtotals plus a grand total you can foot against the source; and a data-request section naming the construction detail, invoices, and lump-sum breakdowns a real study would require. A good output is visibly humble about the thin lines — it does not pretend a one-word asset description supports a confident five-year call.

Review and guardrails. This workflow concentrates the whole course. Every proposed bucket is a hypothesis a qualified professional verifies against the actual asset detail and the applicable rules — the model does not know your client's specifics or the current authority, and a confident "15-year" on a vague line is exactly the hallucination pattern to distrust. Check that the subtotals foot to the source total; a triage that does not reconcile is not usable. Treat the "Needs review" and data-request items as the genuinely valuable output — they scope the real work. And carry the boundary all the way to the client: nothing here is filed, and the engagement, if it proceeds, is a full engineering-based cost segregation study by qualified professionals. A frontier model does noticeably better at the reasoning and the "what's missing" analysis and is worth using; the value, though, is the professional's verification of every line.

A closer look at reviewing the cost-seg triage, because the review is the workflow. It is worth walking through what a professional actually does with the model's output, because this is where the value is created and where the boundary is held. Suppose the model returns a table classifying "decorative lighting package — $48,000" as 5-year with High confidence and "site electrical distribution — $210,000" as 39-year with Medium confidence. The reviewer does not accept either on the model's confidence label; confidence from an LLM reflects how typical the answer looks, not how correct it is for this property. The reviewer instead asks what a real study would examine: what does the decorative lighting actually consist of, is it tied to the building's operation or to specific business function, what does the underlying invoice detail show, and does the site electrical serve the building generally or specific equipment. The model's proposed bucket is a starting hypothesis that tells the reviewer where to look first; the invoice detail, the construction documents, and ultimately the engineering analysis tell the reviewer what is true. A lump-sum line the model flagged "Needs review" is often the most useful output of all, because it points precisely at the cost that has to be broken down before anything can be classified — which is exactly what the data-request section exists to gather.

Notice that the model has done something genuinely valuable here without doing anything a professional could rely on unverified. It read a long, messy asset list in seconds, organized it into a reviewable structure, surfaced the lines that need decomposition, and drafted the client request — turning hours of tedious first-pass sorting into a reviewed starting point. What it did not do, and cannot do, is the engineering cost estimation, the site inspection, and the application of current authority to this specific building that a study supporting a filing requires. Holding both of those facts at once — real acceleration, strict boundary — is the whole skill, and cost segregation teaches it more clearly than any other workflow because the temptation to over-trust a clean, confident table is strongest exactly where the stakes are highest.

Workflow 8 — Tax and accounting research summarization with citation discipline

The workflow and where AI helps. AI can accelerate research by summarizing material you provide — a ruling, a standard, a guidance document you paste in — into plain language and by helping you organize what you are reading. Where it is dangerous, and where this workflow's discipline matters most, is the moment you let it supply authority from its own memory. As Module 1 warned, fabricated citations are a signature failure: real-looking section numbers, plausible case names, Revenue Rulings that do not exist, all delivered with total confidence. The workflow is built to get the summarization benefit while making fabrication impossible to miss.

The prompt (summarizing a source you provide — the safe mode).

Below is the full text of a tax/accounting source document I am providing.
Summarize it in plain language for a professional. For every substantive
statement in your summary, cite the specific section, paragraph, or page of
THIS document it comes from. Do not add any rule, exception, or authority that
is not present in the text I gave you. If the document does not address
something, say "not addressed in this source." Do not reference outside
authority.

[paste the full source text]

The prompt (when you want the model to point you toward authority — the high-risk mode, quarantined).

I am researching [topic]. Suggest, as leads only, the types of authority I
should look for (e.g., "a Code section governing X," "IRS guidance on Y").
Do NOT provide specific citation numbers, case names, or ruling numbers from
memory — I will find the real authority myself in a primary source. If you do
mention any specific citation, explicitly label it "UNVERIFIED - confirm
against primary source before any use."

What good output looks like. In the safe mode: a grounded summary where every claim points to a spot in the document you can check in seconds, and honest "not addressed" notes where the source is silent. In the high-risk mode: search leads and topic framing, with any specific citation loudly marked unverified. What you should never accept as good output is a confident citation to external authority presented as fact.

Review and guardrails — the catch-fabrication routine. For any citation to real-world authority, the rule is absolute: verify it against the primary source before it is used or relied upon, every time, no exceptions. Pull the actual Code section, ruling, or case and confirm it exists and says what the model claimed — checking existence is not enough, because a real citation can be attached to a claim it does not support. Never let a citation move from an AI response into a memo, a client communication, or a filing without that independent confirmation. The safe mode above is safe because the "authority" is a document you already have; the high-risk mode is quarantined precisely because it is the fabrication zone. Treat the AI as a research accelerator over sources you control, never as a source of authority itself. This is not a place to save time on review; it is the place the profession's credibility is won or lost.

Workflow 9 — Meeting notes to action items and follow-ups

The workflow and where AI helps. Converting a page of raw meeting notes into a structured set of decisions, action items, owners, and deadlines is pure transformation on material you supply — a task LLMs do reliably and quickly. AI turns messy notes into an organized follow-up list, extracts commitments, and drafts the recap, saving the tedious sorting while you confirm accuracy.

The prompt.

Below are raw notes from a client meeting. Transform them into:
1. A short summary of decisions made (bullet points).
2. An action-items table: Action, Owner (as stated in the notes), Due date
   (if mentioned), Status (New).
3. A list of open questions or items needing follow-up.

Use only what is in the notes. Do not invent owners, dates, or commitments.
If an action has no clear owner or date in the notes, leave that cell blank
and add it to the open-questions list. Then draft a brief, professional
recap email to the client summarizing the decisions and their action items.

[paste the de-identified meeting notes]

What good output looks like. A clean decisions summary, an action-items table that assigns only the owners and dates actually present in the notes (blanks where they are absent, surfaced as open questions), and a short recap email. The discipline to leave a cell blank rather than invent an owner is exactly what you want to see.

Review and guardrails. Confirm the extracted items match what was actually agreed — the model can misattribute an action or infer a deadline that was never set, and a recap that puts words in a client's mouth is worse than no recap. Check the draft email before it goes out. This is a strong cheaper-model or local-model task: the transformation is well-defined and low-ambiguity, so a small model handles it well, keeping cost near zero and, with a local model, keeping the notes on your own hardware.

Workflow 10 — Spreadsheet formulas, Excel and Power Query, and simple scripts

The workflow and where AI helps. Accountants constantly need a formula, a Power Query transformation, or a small script and do not want to hand-write the syntax. AI is excellent at generating this from a plain-language description of the goal, because code and formulas are highly patterned text. It also explains what an inherited formula does and helps debug one that is misbehaving. You describe the intent; the model writes the syntax; you test it on real data before trusting it.

The prompt.

I need an Excel formula (or Power Query step / short script — specify which).
Describe your solution and explain how it works so I can verify it.

Goal: [plain-language description, e.g., "In column F, flag any row where the
invoice date in column C is more than 30 days before the payment date in
column D, but only if column E is 'Paid'."]
My columns and their meaning: [describe columns and data types]
Constraints: [e.g., must work in Excel 2019; no VBA]

Provide: (1) the formula or code, (2) a plain-language explanation of each
part, (3) one worked example with sample values showing the expected result,
and (4) any edge case I should test (blanks, text-vs-number, errors).

What good output looks like. A working formula or query step, a part-by-part explanation, a worked example with sample inputs and the expected result, and a short list of edge cases to test — blanks, type mismatches, error values. The explanation and the worked example are what let you verify the logic rather than pasting on faith.

Review and guardrails. Always test generated formulas and code on real data, including edge cases, before relying on them in a workpaper. A formula can look correct and mis-handle blanks, text-versus-number, or an off-by-one boundary — and a wrong formula that runs silently is worse than one that errors visibly. Never let AI-generated logic touch a client deliverable without testing it against known-correct results. For simple formulas a cheaper model is fine; for gnarlier multi-step transformations or scripts, a frontier model produces cleaner, more correct code and better edge-case coverage. Either way, the test on real data is the guardrail, not the model's confidence.

The pattern across all ten

Step back and the ten workflows are one workflow in ten costumes. The model reads or transforms material you place in front of it and returns a structured first pass; a qualified professional reviews that pass against the source and their own judgment and turns it into a deliverable they are responsible for. Cheap or local models handle the well-structured, low-ambiguity, low-stakes transformations; frontier models earn their cost where ambiguity, reasoning, or client-facing stakes are high. The two workflows that concentrate the profession's real risk — cost-segregation triage and research citation — are exactly the two where the boundary and the verification matter most: preliminary triage that is never filed, and authority that is never trusted without checking the primary source. Get those two right and the discipline generalizes to everything else. Module 6 turns from individual workflows to the firm: choosing tools on a budget, managing cost, writing the AI-use policy, and building the whole thing safely and cheaply.