# Drafting deliverables: memos, decks, and client reports
A second-year associate at a mid-size advisory firm feeds twelve data-room documents into an AI tool and asks for a due-diligence memo. Ninety seconds later she has eight pages of clean prose. It looks finished. It is not. Two of the risk flags are hallucinated (invented by the model), and the working capitalworking capitalWorking capital is the difference between a company's current assets and current liabilities, measuring short-term liquidity and the funds available to run daily operations.View full definition → summary cites a figure that appears nowhere in the source files.
This is the core tension of AI-assisted drafting in professional services: the machine gets you to a first draft fast, but the last mile, the part clients actually pay for, still runs on human judgment.
This lesson shows you how to build a prompt-and-review workflow that captures the speed without inheriting the risk.
Most professional services output is structured, repetitive, and template-driven. An engagement letter (the contract defining scope, fees, and terms of a client project), a due-diligence memo, a board deck: all follow predictable shapes. That structure is exactly what large language models handle well.
But these documents also carry professional liability. A misstated indemnity clause or a fabricated financial metric is not a typo. It is a client relationship, and sometimes a lawsuit.
So the goal is not "let AI write the deck." The goal is a workflow where AI handles assembly and structure, and humans own facts, judgment, and accountability.
Think of it as four stages. AI does heavy lifting in stages one and two. You dominate three and four.
Do not start by asking for a finished memo. Start by asking the model to organize your inputs.
Suppose you have deal notes, three financial statements, and a management interview transcript for a due-diligence review. Your first prompt:
> "You are helping structure a due-diligence memo. Below are my raw notes and documents. Extract and list, without commentary: (1) all quantitative figures with their source document, (2) stated risks, (3) open questions. Do not add analysis. Flag anything where the source is unclear."
This does two things. It forces the model to tie every number to a source, and it surfaces gaps before they hide inside polished prose.
Now give the model the skeleton. Professional firms already have house templates. Paste yours in.
> "Using only the extracted facts above, draft Section 3 (Financial Overview) of our standard DD memo. Follow this structure: [paste headings]. Where a fact is missing, write [GAP: describe what is needed]. Do not infer or estimate."
The [GAP] instruction is the workhorse. Instead of the model smoothing over missing data with confident-sounding filler, it leaves visible holes for you to fill.
This is where partner judgment lives. Run every AI draft through a fixed checklist:
A useful mental rule: AI can describe. Only you can recommend.
Once facts and judgment are locked, send it back to the model for polish.
> "Tighten this section for a partner audience. Keep every number and caveat exactly as written. Reduce hedging language. Target one page."
Polish is safe to automate because you have already validated substance.
Here is how the four stages play out on a real deliverable type.
Input: Buyer's counsel sends you a target company's financials and asks for a risk memo before a Tuesday committee meeting.
Stage 1 turns the data room into a sourced fact list. You immediately see the model flagged that revenue in the pitch deck (say, an estimate cited by management) does not match the audited statements. Good. That mismatch is now visible instead of averaged into a single fake number.
Stage 2 produces a structured draft with [GAP: no explanation for revenue discrepancy] sitting in the text.
Stage 3 is where you earn your fee. You call the client, learn the discrepancy is a revenue recognition timing difference, and you write the judgment sentence: "The variance reflects timing under the new revenue standard and is not a going-concern indicator." No model can make that call. It requires knowing the client, the standard, and the stakes.
Stage 4 formats it into the committee's preferred one-page brief.
Total time: perhaps a third of a fully manual draft. But notice the time you saved went into extraction and formatting, not into thinking. The thinking stayed with you.
🎬 [VIDEO: "How to Write Effective Prompts for Business Writing" — youtube.com — a practical walkthrough of prompt structuring for professional documents]
| Task | AI leads | Human leads |
|---|---|---|
| Organizing raw notes | Yes | Review |
| Drafting boilerplate sections | Yes | Review |
| Summarizing long documents | Yes | Verify |
| Formatting a deck | Yes | Sign off |
| Assessing materiality of a risk | No | Yes |
| Recommending a course of action | No | Yes |
| Legal or fee terms in an engagement letter | No | Yes |
| Final client-facing sign-off | No | Yes |
A grounding principle from research on AI reliability: models are fluent, not truthful, by default. Fluency is not evidence. For a plain-language primer on why models fabricate, see this overview from the Alan Turing Institute.
Never paste confidential client data into a public consumer tool. Use your firm's enterprise or private deployment, where inputs are not used for training and data stays governed. Check your engagement terms and confidentiality obligations first.
Keep a source binder. Every figure in a deliverable should trace to a document you can produce. The AI workflow makes this easier if you enforce citation in stage one.
Disclose per firm policy. Some clients and some regulators now expect disclosure of AI assistance in deliverables. Know your firm's stance before you send.
The accountability rule: whoever signs the deliverable owns every word, including the words the model wrote. "The AI said so" is not a defense to a client or a court.
Knowledge check
1. According to the lesson, what is the core tension of AI-assisted drafting in professional services?
2. Why does the lesson describe professional services deliverables as both the 'perfect AI use case' and the 'perfect trap'?
3. In the prompt-and-review workflow, why should you NOT begin by asking the model for a finished memo?
4. Select ALL correct answers. According to the lesson, which responsibilities should humans own rather than delegate to the AI?
Select all the correct answers.
5. Select ALL correct answers. What kinds of errors in the associate's AI-generated memo illustrate the danger of trusting AI output uncritically?
Select all the correct answers.
The same four stages transfer, with tweaks.
Consulting decks: AI is strong at converting a validated findings memo into slide-level bullets and a logical flow. Prompt it to produce a "storyline" first (one sentence per slide), review the argument, then generate content. Never let it invent data for a chart. Feed it your numbers.
Engagement letters: Use AI to assemble standard clauses from your approved library and to check the draft against a scope checklist. Do not let it draft fee terms, liability caps, or indemnities from scratch. Those are legal and commercial decisions. A partner and, where relevant, general counsel own them.
The pattern holds across all three: automate assembly, own judgment.
[GAP] markers, run a human review pass, then automate polish.