AI for Wealth Managers: Drafting Proposals With a Compliance Review

A wealth management firm can use AI to draft the descriptive wording and layout of a client proposal, as long as an authorised adviser supplies the recommendation, every figure is checked against source, and the finished document goes through the firm's normal compliance approval before sending. Archive the approved version and the AI-assisted drafts together.

The line to hold is between drafting and advising. A model can turn an adviser's bullet points into clear paragraphs, tidy a fee table or rewrite jargon for a retiree. It should never choose a portfolio, estimate returns or judge suitability. If that line is written into your process, the rest is mostly housekeeping: how to prompt, what reviewers should look for, and what to keep.

Follow me on Instagram@sagnikteaches

What AI may and may not touch in a proposal

Take a six-adviser firm that sends about 15 proposals a month. Here is a division of labour (illustrative) that compliance teams can read at a glance:

Connect on LinkedInSagnik Bhattacharya

This table becomes part of your internal procedure. The rule for AI is simple: it may change how something is said, never what is claimed.

Subscribe on YouTube@sagnikteaches

Why review matters more with AI drafts

Two regimes show what a reviewer is protecting. FINRA Rule 2210, which governs communications by FINRA member firms, says communications must be fair and balanced, give a sound basis for evaluating the facts, and not contain false, exaggerated, unwarranted, promissory or misleading statements. A retail communication (one sent to more than 25 retail investors within 30 calendar days) needs approval by a qualified registered principal before use, with limited exceptions; correspondence (25 or fewer) is supervised and reviewed under the firm's procedures. By those definitions a proposal written for one client is usually correspondence, while a standard proposal pack sent to dozens of prospects in a month can become a retail communication, so ask compliance which category each document falls into. The rule also requires records, including the approving principal and date, the dates of first and last use, and the source of any statistical table, chart, graph or other illustration. For firms under the FCA's Consumer Duty, one of the four outcomes is consumer understanding (the others are products and services, price and value, and consumer support): information should be clear, timely and free of jargon so customers can make informed decisions, and the FCA expects firms to test that customers have understood key information. These are two examples, not a complete list: your own regulator and rulebook decide what applies, so ask compliance rather than relying on this summary.

Whatever the regime, AI drafts fail in recognisable ways. Reviewers should be told what to look for:

• Promissory language. "Your portfolio will grow steadily", "protect your family's future". Models reach for reassuring phrases that read as promises.
• Dropped qualifiers. A sentence that originally said "may fall as well as rise" becomes "historically performs well".
• Invented specifics. A model asked to make a proposal "more persuasive" may add a statistic, a ranking or a reference to "decades of experience".
• Changed numbers. A fee of 0.85% becomes 0.8% in a tidy-up, or a date shifts.
• Tone drift. Warm, informal wording is fine in a cover note but not in the section describing risks and charges.

An AI-assisted flag on the review form means the reviewer reads with these in mind. It also prompts the adviser to confirm that the facts section matches the fact-find.

A drafting workflow in seven steps

1. Prepare confirmed inputs. The adviser writes a short briefing: client situation in five or six bullets, goals, the recommendation in their own words, and the fee structure ID. No raw recordings, no account numbers.
2. Use placeholders for personal data. Draft with "[client first names]", "[amount]" and "[date]" and fill them inside your own system afterwards. This limits what ever reaches an outside tool.
3. Run the approved prompt. One prompt saved in a shared project or document, not retyped by each adviser.
4. Adviser edits and confirms. They check every statement against the briefing and delete anything the model added.
5. Insert fixed text and numbers. Risk warnings from the compliance library, tables from planning software, charges from the fee schedule.
6. Compliance review. The reviewer sees the draft with an AI-assisted flag, plus the briefing so they can compare.
7. Approve, send and archive. Covered in its own section below.

The saved prompt

You are drafting wording for a wealth management proposal. You are not giving advice.

Use ONLY the facts in the adviser briefing below. Do not add any figure, statistic, ranking, performance statement, tax statement or product claim that is not in the briefing.
Do not use promissory language such as "will", "guarantee", "safe" or "ensure" about outcomes.
Write in plain English at roughly a 12-year-old reading level. Keep each paragraph under 90 words.
Where the briefing is silent, write "[ADVISER TO COMPLETE]" instead of guessing.
Leave placeholders in square brackets unchanged.
Sections: 1 Your situation and goals. 2 What we recommend and why. 3 What it costs. 4 What could go wrong. 5 Next steps.
For sections 3 and 4, copy the briefing text and only fix spelling and grammar.

Briefing:
[paste]

The final instruction on sections 3 and 4 is deliberate: those are the sections where rewording does the most harm. The output reads less polished, and that is acceptable. Compliance would rather review a clunky sentence they recognise than a smooth one they do not.

A worked pass through one proposal

Say an adviser prepares a proposal for a couple in their late fifties who want to consolidate three pensions. The briefing includes the goal ("retire at 65 with a target income of $4,500 a month before tax"), the recommended structure (one consolidated account in a moderate-risk model portfolio) and the fee summary (advice fee 0.75%, platform 0.25%, fund costs 0.18%, all illustrative).

The model returns, for section 2 (sample, illustrative):

We recommend moving your three existing pensions into one account. This would give you a single statement and one set of charges to follow. The account would be invested in a moderate-risk portfolio, which matches the level of risk you told us you are comfortable with. Moving pensions is not right for everyone, and we have checked your existing plans for guarantees or exit charges first: [ADVISER TO COMPLETE].

What the adviser fixes: "we have checked your existing plans" is a claim the model could not know, so the placeholder is right and the adviser must replace it with the factual result or remove it. What compliance looks for: there is no promise, the reason links to a stated client objective, and the sentence about suitability sits next to a caveat. The reviewer also spots that the phrase "single statement" implies simplicity but does not mention that consolidating can lose benefits, so they add the library paragraph on transfers.

Checking the numbers

Numbers do not go through the model at all. For any table, the process is: export from the planning system, paste into the template, and have a second person check against the source. If you want an AI check, use it as a second pair of eyes after the human has done the work, giving it both the briefing and the final document and asking for inconsistencies only:

Compare the fee table in the final document with the fee schedule below. List every difference in figure, label or date. If there are none, say "no differences found". Do not correct anything.

Treat a "no differences found" answer as weak evidence, not proof, because models can miss things in tables. For the longer discussion of drafting investment narrative while keeping the numbers under human control, see Can AI write an investment proposal?

Archiving drafts so you can show your working

Create a folder or record per proposal containing:

• The adviser briefing and fact-find version used.
• The prompt version (a date or number is enough) and the raw AI output.
• The edited draft with the adviser's changes visible, if your tool tracks them.
• The compliance review record: reviewer, date, comments, approved or sent back.
• The final document as sent, with its first-use date.
• The source of each chart, projection or illustration.

FINRA's rule lists approval details, dates of use and statistical sources among its records, and your own retention schedule sets how long. Do not rely on a chat history: ChatGPT and Claude histories are personal to an account and can be deleted, and OpenAI's help centre says self-service data export is not available in ChatGPT Business workspaces. Copy what you need into your document system at the time. If a client complains two years later, you want one place to answer the question "who wrote this and who approved it?"

A reviewer's checklist, filled in for the example

Give reviewers a one-page form so that a review is a record rather than a nod. Here is the example proposal after review (illustrative):

The "read aloud by a colleague" check is cheap and effective. A reviewer who knows the product reads past jargon; someone who does not will stop at "drawdown" or "platform fee" and tell you. If your firm works under a consumer-understanding standard, this also produces a small piece of evidence that you tested the communication.

Before and after: rewriting a dense paragraph for a retiree

Another good use for AI is making existing approved text clearer, as long as the meaning does not change. Here is a dense paragraph an adviser might inherit (illustrative):

Your pension will be subject to a phased drawdown arrangement wherein capital is crystallised in tranches with the residual fund remaining invested, thereby exposing the residual to market risk during the decumulation phase.

Prompt: "Rewrite this for a reader with no financial background. Keep every fact and every risk. Do not soften the risk. Do not add anything." A realistic output:

You would take money out of your pension in stages. The part you have not yet taken stays invested, so its value can go down as well as up while you are taking income.

What you check: the rewrite kept "can go down as well as up", which is the point. If a version had said "stays invested for growth", that would be a change of meaning and the reviewer would reject it. Send both versions to compliance, since the second is a new document even if it started from approved text. If it passes, add it to the approved phrase library so the next adviser does not need to repeat the exercise.

A third example: cleaning up a messy fee section layout

Layout is where AI saves time with the least risk. A proposal assembled from three past documents often has three styles of fee table, inconsistent headings and percentages written as "0.75%", "0.75 per cent" and "75bps". Give the model the text and ask: "Make heading levels consistent, use one format for percentages (0.75%), put each charge on its own line in the order advice, platform, fund costs, and list every change you made. Do not alter any number." Then compare the change list with the document. Models sometimes sneak in a rounding; the change list makes it findable. This is a ten-minute job done in two, and the reviewer sees a cleaner document.

Rolling it out across the firm

For a firm of six advisers, a sensible sequence over a month looks like this:

1. Week 1: compliance and a principal agree the division of labour table, the saved prompt and the review form. One adviser volunteers as pilot.
2. Week 2: the pilot adviser drafts three real proposals this way, in parallel with the old method for one of them to compare time and quality.
3. Week 3: the review comments are collected, the prompt is revised, and a short list of the model's most frequent mistakes is added to the review form.
4. Week 4: the remaining advisers are briefed in a 30-minute session using the pilot examples, and the AI policy is updated to name the approved tool and the prohibited uses.

Do not open the process up to every adviser's personal account. A shared workspace under a business plan means the prompt, the library and the settings are managed once. Check the vendor's pricing page for current seat prices before you commit; the figures in the next section were checked on ChatGPT's and Claude's pricing pages in November 2026.

What to tell clients

Your terms of business or a footer line can say that your firm uses software tools, including AI, to help draft documents, and that every document is reviewed and approved by a named person at the firm before it reaches the client. Avoid saying AI "writes" your proposals: it overstates the machine's role and understates the adviser's. Your compliance team decides whether and how this is disclosed, so draft the wording with them.

Which tools to use for the drafting step

For a firm of this size a business-plan chat assistant is usually enough: ChatGPT Business lists Standard seats at $25 a month or $20 billed annually, from two users, and Claude Team lists Standard seats at the same $25 or $20 for teams of 2 to 150. Both vendors say business plans do not train on your content by default. Both let members share Projects inside the workspace, where you can store the saved prompt, tone guidance and a library of approved phrases so every adviser starts from the same instructions. On Claude Team an admin can switch project sharing off, so confirm the setting when you configure the workspace. If your adviser CRM or proposal tool already has drafting built in, compare it before adding another subscription; what is built in versus bolted on covers that decision. Whatever you pick, put it through your vendor due diligence and update your AI policy; the governance checklist for small advice firms lists the questions.

Prospect packs, review letters and other edge cases

A prospect proposal sent before the fact-find is complete. Use the same workflow but with a "no advice" label and no recommendation section; the AI drafts the firm's introduction and services description from approved copy. Compliance decides whether that document is a marketing communication and what approval it needs.

An annual review letter. Reuse the prompt with the review briefing. The extra check is that the letter's performance wording follows approved templates, because past-performance statements are the usual trouble spot.

An adviser pastes a competitor's proposal as a style guide. Stop this: copying another firm's material raises copyright and accuracy issues. Use your own approved past proposals as the style reference.

The model produces a persuasive paragraph the adviser likes. Make the adviser trace each claim in it to the briefing. Usually two of five claims have no source. Delete them, and the paragraph is as good as it needs to be.

A realistic mistake and how it surfaced

Say an adviser, short of time, pastes the whole fact-find into the model including the clients' full names, dates of birth and account numbers, and asks it to "write the proposal". Three things go wrong at once. Personal data has gone to a tool the firm may not have assessed. The output contains a section on tax efficiency that nobody on the briefing supplied. And no prompt version or briefing was kept, so the reviewer cannot tell what was asked. It surfaced when the compliance reviewer asked where the tax paragraph came from and the adviser could not say.

The fixes were procedural, not technical: a one-page briefing template that is the only permitted input, a saved prompt that forbids unsourced claims, a rule that personal identifiers stay in the firm's own systems, and a review form with a mandatory prompt-version field. Notice that none of these needs better AI. They need a process that makes the safe route the easy one.

Keeping the prompt and phrase library healthy

Assign one person, usually the compliance officer or a senior adviser, to own the saved prompt and approved phrase library. Review it every quarter and whenever a vendor changes its models or settings, because behaviour shifts when the underlying model does. Keep a change log with the date, the change and the reason ("added rule against 'secure' after reviewer comments in March"). Test the prompt on two old, approved proposals after each change: the output should not introduce claims that were not in the briefing. If it does, fix the prompt before anyone uses it on a live client.

Measuring whether it helps

Track three numbers for a month, using your own illustrative baseline: time from briefing to compliance submission (for example down from 90 minutes to 45), the number of review comments per proposal, and the number of proposals sent back for wording problems. If time falls but sent-back rates rise, the drafting step is saving time by moving work onto compliance, which is a net loss. Adjust the prompt or the review checklist, not the reviewer's standards.

If you want to work through the process with someone, that is what my AI implementation consultation covers. For the related question of how advice records are kept, see AI note takers for financial advisers.

Questions compliance officers and principals ask

Can AI decide which investments go in a proposal?

No. In this workflow AI only drafts wording and layout around recommendations a qualified adviser has already made. Suitability, asset allocation and product selection stay with named, authorised people. Treat any recommendation text a model volunteers as an error to delete.

Do AI-drafted proposals need separate compliance approval?

They need the same approval as any proposal. The reviewer should know AI helped, because the common errors differ: invented performance claims, softened risk wording and wrong figures. Many firms add an AI-assisted tick box to the review form.

What should we archive?

Keep the final version sent, the approval record with name and date, the source of every figure or illustration, and the prompts and intermediate drafts for AI-assisted documents. Your retention schedule and regulator rules decide how long, so check them with compliance.

Can we put client data into ChatGPT or Claude?

Only under a business plan your firm has approved, with a data-processing agreement, and only the minimum needed. Many firms draft from placeholders and add personal details inside their own systems. Check your firm's AI policy and the vendor's terms first.

Further reads

• Can AI Draft a Suitability Report That Passes Compliance Review? — The stricter cousin of a proposal: reports that must pass compliance.
• Can Financial Advisers Use ChatGPT? What Compliance Allows — What compliance typically allows before any tool is switched on.
• AI Compliance Checklist for Small Businesses: What Applies to You — A wider checklist of what applies when you adopt AI.
• How Advisers Use AI to Prepare for Annual Client Reviews — Reuse the drafting routine for annual reviews.
• AI Financial Planning Software for Advisers: What to Test First — What to test in planning software before AI touches proposals.
• Best AI App for Wealth Management: Jump, Zocks and Wealthbox — Jump, Zocks and the Wealthbox notetaker compared on price, recording and archiving, with a break-even sum and a pilot a small firm can run.
• AI for Wealth Management Firms: A 90-Day Rollout Plan for Small Firms — A week-by-week 90-day plan for a small wealth firm: policy first, a meeting-notes pilot, vendor checks, training, and measured hours and errors.
• AI Marketing for Financial Advisers: Content That Clears Compliance — Prompt rules built from the SEC Marketing Rule and NASAA's AI guidance, a review workflow, and a before-and-after post a compliance officer can sign.
• AI Tools for Investment Managers: Client Reports and Fact Sheets — How a small investment manager can draft quarterly reports and fact sheets with AI while the numbers stay locked and compliance gets a reviewable packet.
• AI Tools for UHNW Wealth Managers: Seven Jobs With Privacy Rules — A job-by-job guide to AI in a small wealth team, including access limits, sample outputs and review gates.
• AI for Financial Advisers UK: Suitability Letters and Meeting Notes — A workflow for drafting suitability letters and meeting notes with AI while the adviser stays accountable and Consumer Duty evidence stays intact.
• AI in KYC for a Small Wealth or Advice Firm: Checks That Stay Human — Which parts of client KYC AI can handle for a small advice firm, which must stay with a person, and what records to keep.
• SEC Recordkeeping for AI Notes and Chats: What an RIA Should Archive — Rule 204-2 and AI: which meeting notes, transcripts and chatbot chats an RIA should keep, where to store them and for how long, tool by tool.
• AI in Wealth Management Examples: Five Admin Jobs Worth Trying — Meeting notes, follow-ups, review packs, onboarding checklists and newsletters: five admin jobs a small wealth firm can give AI, plus the records caveat.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: FINRA Rule 2210 text on finra.org; FCA Consumer Duty pages on fca.org.uk; ChatGPT Business and Claude pricing pages; OpenAI help articles on data export and Projects; Claude help article on Projects; all checked 4 November 2026.