Generic AI drafts fail because they pull from broad, impersonal sources resulting in textbook-like output with wrong tone, structure, and terminology. The solution is not better AI, but better training; teach it your firm’s own voice using your past work.
If you have experimented with AI for report writing, you have probably hit the same wall most valuation professionals hit: the output is competent but generic. It reads like a textbook wrote it. The section order is not yours, the terminology is slightly off, the tone is wrong for your clients, and by the time you have rewritten it to sound like your firm, you have saved maybe 20 minutes.
Here is the thing: that is not an AI limitation. That is a training gap, and you are the one who has not done the training!
Your best 10 valuation reports contain everything AI needs to draft like your firm: your section structure; your standard limiting conditions; the way you phrase a company-specific risk premium discussion; and the transitions you use between the industry analysis and the financial review. That intellectual capital is sitting in your archive, and most analysts never put it to work.
This article shows you how, step by step, in an afternoon.
Why Generic AI Drafts Fail
When you ask a general-purpose AI to “write an income approach section,” it draws on everything it has ever seen (academic papers, blog posts, other firms’ styles, and plain guesses). The result averages out to nobody’s voice.
What you actually want is an assistant that has internalized your patterns:
- Your report skeleton: the exact sections, in the exact order, with your heading conventions
- Your standard language: the paragraphs that appear in nearly every report with minor tailoring
- Your voice: sentence length, formality level, how you hedge, how you cite
- Your analytical framing: how you walk a reader from data to conclusion
Get those four things into the AI’s context, and the quality jump is dramatic. Not because the AI got smarter but because you finally told it what “good” looks like at your firm.
Step 1: Pull Your Best 10 Reports (Not Your Last 10)
Choose deliberately. You want reports that represent the writing you would be proud to replicate; ideally spanning your common engagement types: a couple of gift and estate valuations, a shareholder dispute, a marital dissolution, an SBA or transaction-related engagement. Variety teaches the AI how your voice flexes across contexts while the core style stays constant.
If a report contains sections you always rewrite because a former partner drafted them badly, leave it out. You are about to clone this writing. Clone the good stuff.
Step 2: Sanitize Before Anything Touches an AI Tool
This step is non-negotiable, and it is where most analysts get sloppy. Before any report goes into an AI system:
- Replace names: client company becomes “the Company,” individuals become roles (“the Owner,” “the non-titled spouse”)
- Remove identifiers entirely: EINs, SSNs, account numbers, case captions, docket numbers
- Generalize locations: “a metro area in the Mountain West” instead of the actual city
- Scale the financials: multiply every figure by a consistent factor; ratios and patterns survive, real numbers do not
And match the data to the tool tier: sanitized excerpts on an enterprise-grade AI agreement you have actually read, or a local model if the engagement demands it. Consumer chatbot accounts are not the place for client work, sanitized or not, if the terms allow training on your inputs. (Next time, I will create a decision tree, so you can be confident when to redact and when it is OK to proceed.)
Step 3: Extract the Three Assets
Now the interesting part. Feed your sanitized reports to the AI one or two (depending on the length) at a time and have it do the analysis of your writing not of the companies. Three extraction passes:
Pass 1 Structure. Prompt: “Analyze the section structure of these valuation reports. Produce a master outline showing every section and subsection in order, noting which appear in every report versus only certain engagement types.” The output is your report skeleton, documented, possibly for the first time ever.
Pass 2 Standard language. Prompt: “Identify paragraphs and passages that appear in substantially similar form across multiple reports. Extract each as a reusable template, marking the variable elements in brackets.” This surfaces your boilerplate library (limiting conditions, standard of value definitions, methodology descriptions, certification language).
Pass 3 Voice. Prompt: “Describe this firm’s writing style in specific, replicable terms (typical sentence length, level of formality, use of first person, how conclusions are hedged or asserted, transition patterns, and citation style). Write it as instructions another writer could follow.” This is your style guide; the thing big firms pay consultants to produce.
Step 4: Assemble the Assistant
Combine the three assets into a single instruction document—your firm’s “writing constitution.” It should read something like: “You are a report writer for a business valuation firm. Follow this section structure: [skeleton]. Use these standard passages where applicable: [templates]. Write in this style: [voice guide]. Always flag any factual claims you cannot support from the provided engagement data.”
Where you put this document depends on your tools. In Claude, create a Project and load it into the project instructions along with two or three sanitized sample reports as reference material. In an API-driven workflow, it becomes your system prompt. Either way, every future drafting request now starts from your firm’s DNA instead of from zero.
That last instruction, flagging unsupported claims, matters more than it looks. Your assistant should draft your language, not invent your facts.
Step 5: Test It on a Real Section, Then Iterate
Take a current engagement, sanitize the key inputs, and ask for a single section such as the economic outlook or the company overview. Compare it against what you would have written. The first attempt will be noticeably better than generic AI but not perfect. When you spot a miss (“we never say ‘in conclusion,'” “the guideline company discussion always precedes the transaction method”), add that correction to the instruction document.
Three or four iterations in, you will hit the threshold this article promised: drafts that arrive 80% complete, where your job shifts from writing to reviewing and refining, which is where your professional judgment belongs anyway.
The Compounding Payoff
Here is what most firms miss: this asset appreciates. Every correction of feedback makes the next draft better. New staff can draft in the firm’s voice on day one. Partner review time drops because drafts arrive consistently. And the exercise of extracting your structure and standard language often reveals inconsistencies across your own reports that were worth fixing regardless of AI.
Ten reports. One afternoon of sanitizing and extracting. A writing assistant that actually sounds like your firm.
Your archive has been holding this the whole time.
Colin Brown, CTO of Syncnet, helps consultants integrate AI into their practices.
Mr. Brown may be contacted by e-mail to cto@syncnet.com.
