Can AI Write an Investment Proposal? Draft Narrative, Check Numbers

Yes. An AI investment proposal generator can extract statement data and draft a wealth proposal's explanation or investment policy statement, but the adviser must verify every holding, valuation date, fee, allocation, assumption and disclosure against the underlying records. The adviser, not the generator, owns the recommendation and the final compliance review before anything goes to the client.

The useful boundary is between drafting and deciding. A polished chart can conceal a duplicated holding, a cash balance from a different date or a fee stated on the wrong base. Build a source-to-output trail first, then ask AI to explain figures that have already passed arithmetic and suitability checks.

Follow me on Instagram@sagnikteaches

Decide which part of the proposal AI may draft

A wealth proposal usually blends four jobs: assembling current holdings, comparing current and proposed positions, explaining the recommendation in plain language, and producing an approved document. These jobs need different controls. An assistant can put notes into coherent prose. It cannot make an incomplete client fact-find complete, or turn an unapproved allocation into advice merely by describing it persuasively.

Connect on LinkedInSagnik Bhattacharya

Specialist products differ. Feathery says its proposal workflow imports statements, notes and client records into a firm template and routes the document through existing review; it explicitly says it does not generate investment recommendations. Investipal describes statement extraction, portfolio comparison, commentary and investment policy statement drafts that an adviser reviews. Tecla describes a tailored proposal and IPS agent with a review gate. Their pages establish possible workflows, not guaranteed accuracy on your statements. Ask each supplier to demonstrate your messiest document format and your firm's actual approval path.

Subscribe on YouTube@sagnikteaches

A general assistant is an alternative for a small, controlled drafting task: give it a verified table and approved template, then ask for a narrative. It is a weaker choice for raw multi-account extraction, live valuations or final disclosures unless your firm has approved that handling. The broader wealth-manager proposal drafting tutorial covers the overall production flow; the question here is the numbers gate that prevents a plausible draft becoming a misleading proposal.

Make a statement ledger before you write a sentence

Start with a one-row-per-position ledger. Record the source document, account identifier in an approved internal form, page, statement end date, security identifier, units, price or quoted value, currency, cost basis if needed, and whether the line is a holding, pending transaction or cash. Mark anything the statement omits as unknown. Never turn an unknown cost basis into zero; zero is a real number with a different meaning.

For a household with several accounts, keep account totals separate. A transfer in flight can appear as cash in one statement and a position in another. An assistant that combines both may double count it. Likewise, a statement can show a total portfolio value and a separate informational section for assets held elsewhere. If you add the latter again, the headline assets under review become inflated. Note the inclusion rule next to the ledger, not only in a footnote after the proposal is designed.

Illustrative check: a statement shows 120 fund units at $75, a bond holding of $18,000 and $3,000 cash. The three lines total $30,000. If the extracted table says 120 units at $7.50, its total becomes $21,900. The error is visually small on the unit-price line but changes the proposed portfolio weight and the client's starting point. The reviewer should return to the source page before any chart is drawn.

Have the extractor output a page reference and a confidence or uncertainty note for each field. Then inspect the page itself. A page citation is a pointer, not proof: some PDFs number a cover page and a printed page differently. If a value cannot be traced, remove it from the proposal until a person resolves it. The same source-tracing discipline appears in the financial statement numbers-check tutorial, but a proposal needs an additional link from verified figures to the recommendation.

A $600,000 proposal with an independent tie-out

Consider an illustrative advice firm preparing a proposal for a household whose approved fact-find records a medium-term income need and a separate long-term growth goal. The most recent verified statements show $240,000 in broad equity funds, $210,000 in bonds, $90,000 in cash and $60,000 in a concentrated single holding. The ledger totals $600,000: equity funds 40%, bonds 35%, cash 15% and the single holding 10%. Those are observations, not yet a recommendation.

The adviser considers a proposed allocation of 45% diversified equity, 40% fixed income and 15% cash, with the concentrated holding reduced. At $600,000, the draft target values are $270,000, $240,000 and $90,000. The current equity fund value is not simply subtracted from $270,000 because the concentrated holding also belongs to the wider equity exposure. The adviser must decide how to classify it, what will be sold, which account bears the change, and whether the transfer or tax consequences alter the plan. AI should not silently convert an asset-class comparison into trade instructions.

Suppose the tool drafts: “Your current portfolio is 50% equities and the proposed mix increases equities by five percentage points.” That can be arithmetically consistent only if the single holding is included in equity: $240,000 plus $60,000 equals $300,000, or 50%. But the phrase “increases equities” is then wrong if the proposed target is 45%. The reviewer would amend it to “reduces total equity exposure by five percentage points while diversifying the remaining equity allocation”, subject to the adviser's actual decision. This is why a correct sum does not guarantee a correct sentence.

Now take an illustrative fee schedule: 0.60% advisory charge on $600,000, or $3,600 a year, plus a weighted fund cost of 0.18%, or $1,080 if the same base applies. The combined illustration is $4,680, equivalent to 0.78% of $600,000. That number is only usable if the adviser has checked the chargeable base, fund selections, whether any fee is tiered and whether transaction, platform or other costs are excluded. If only $510,000 is billed for advice, applying 0.60% to the whole household would overstate that component by $540. The proposal should name its basis instead of presenting one neat percentage without conditions.

Record the review in a small tie-out sheet: statement value $600,000; extracted value $600,000; proposed targets $600,000; fee basis checked against the agreed schedule; concentration and liquidity considerations reviewed by the adviser; final proposal version and reviewer identified. This example is deliberately simplified. A real household may have different valuation dates, restricted assets, commitments, liabilities and costs that require separate treatment.

The adviser's numbers-check list

Use this list on the rendered proposal, not only the spreadsheet behind it. Copying a correct number into the wrong chart or paragraph is still a proposal error. Have one person mark each line complete and a second reviewer verify material exceptions according to firm policy. The made-up figures checklist is useful for general drafting errors; the controls below are specific to investment comparisons.

• Scope: List included accounts and excluded assets. Check the account count against the fact-find and source package.
• Date: Print the valuation date beside every current value. Do not combine a month-end balance with a later cash figure without labelling the difference.
• Identity: Verify security names, identifiers and units. Similar fund share classes can have different charges.
• Total: Sum positions and cash independently. Reconcile to each statement total and investigate material differences.
• Currency: Show the currency of each account and any exchange rate used. A converted headline needs an identified rate and date.
• Weights: Recalculate each current and target percentage from the same verified denominator. Rounded rows need not sum visually to exactly 100%, but the unrounded model should.
• Costs: Recompute advice, fund, platform and transaction costs on their actual bases. Distinguish current from proposed cost.
• Tax and restrictions: Flag gain, loss, lock-in, withdrawal or account constraints for qualified review. Do not use an AI estimate as a filed or payable amount.
• Performance: Confirm dates, benchmark, gross or net basis and any required method before including a return or projection.
• Consistency: Compare every chart, caption, table and prose claim with the approved calculation sheet.

Give each failed check a disposition: correct the extraction, change the wording, obtain a missing source, or pause the proposal. “Reviewed” without a recorded resolution is weak evidence when the error later appears in a client copy. A simple issue log with version, page, disputed figure, source, owner and resolution is enough for a modest workflow.

Draft the narrative from approved figures only

Once the ledger and calculations are signed off, supply the assistant with a deliberately narrow brief. Include client objectives in the firm's approved wording, the verified table, the proposed allocation that an adviser has chosen, factual reasons, limitations and the approved template. Ask it to distinguish a verified fact from an assumption. Require an “unresolved” section; otherwise the model may fill blanks with confident prose.

Draft a client-facing explanation for the approved proposal.
Use only the verified figures and adviser decisions below.
Do not choose investments, calculate returns or invent a fee.
For every numerical claim, show its ledger row in brackets.
Separate facts, adviser rationale and open questions.
If a statement or fact-find field is missing, write "needs adviser review".
Keep the tone clear and avoid promises of performance.

An illustrative output might read, “The proposal moves the portfolio from 50% equity exposure to a diversified 45% target, retaining 15% cash for planned withdrawals [allocation rows 4–7].” The reviewer should still inspect whether the cash target is genuinely linked to withdrawals, since the fact-find may say only “reserve for foreseeable needs”. Edit the rationale to match the client's recorded objective, not a generic story invented to make the allocation attractive.

For another illustrative case, a prospect has $80,000 of cash earmarked for a purchase within months. A draft that calls this “uninvested money ready for growth” changes its purpose. The correct edit is to label it as a near-term commitment and exclude it from the investable allocation if that is the adviser's documented decision. This is a narrative error even when the $80,000 is transcribed perfectly.

What the investment policy statement must agree with

An investment policy statement, or IPS, records the client's objectives and the rules for managing the portfolio. AI can help turn approved notes into a first draft, but the IPS and proposal must agree on the accounts covered, target ranges, rebalancing trigger, permitted assets, income needs, liquidity and review frequency. A proposal that recommends 45% equity while its IPS allows 20–35% is not ready for sign-off.

Take an illustrative IPS with a 40–50% equity range and a 15% cash target. A generated sentence says, “We will rebalance whenever any asset class moves by 5%.” That could mean five percentage points from target or a relative five-per-cent move. The difference matters. If equity rises from 45% to 49%, one interpretation does nothing; another may trigger action. The adviser must replace the vague phrase with the firm's approved rule and ensure the same rule appears in the operational record.

Another example is a proposed annual income withdrawal of $24,000 from a $600,000 portfolio. The naive percentage is 4% of the current value, but the proposal must not treat that as a sustainable or guaranteed return. Ask the adviser to check the withdrawal timing, cash reserve, fees, tax position and what happens if markets fall. AI can explain a reviewed plan; it cannot validate the plan merely by calculating $24,000 divided by $600,000.

Five failure patterns that survive a neat layout

Stale values. An illustrative statement package contains one account at month-end and another updated a week later. The generator labels the combined total “as at month-end”. Fix the caption or obtain matching-date values. Otherwise the comparison appears more precise than its inputs.

Share-class substitution. A fund title is shortened during extraction and the narrative cites a lower charge for another share class. Match the exact identifier and current charge source before calculating the proposed cost. A product name alone may be insufficient.

Cash mistaken for income. A statement shows a $12,000 distribution credited to cash. The model treats it as a new recurring annual income stream. Check whether it was an exceptional distribution, reinvestment or account movement before using it in a spending plan.

Unapproved assumption. A draft says the client can tolerate a 20% fall, although the fact-find records no agreed loss capacity. Remove the number and return to the adviser. Risk vocabulary must be grounded in the client's actual record.

Chart and prose drift. After a reviewer edits the target table, the AI-written opening still says bonds rise to 35% rather than 40%. Re-run a final comparison on the exported PDF. A version number and approval date on both calculation sheet and document make this easier to catch.

These failures are reasons to specify a reviewer, not reasons to abandon the tool. The cost of checking should be measured in the pilot. If the team spends more time repairing source errors than it previously spent assembling proposals, narrow the task to narrative drafting from verified inputs. Do not call a document “compliant” because a vendor page says its workflow supports compliance; the firm still needs its own disclosures, permissions and review record.

Choose a tool by testing the control points

Ask a specialist supplier for a live demonstration using consented, representative documents: a clean statement, a scanned page, a position with an unfamiliar name, and a multi-account household with a cash transfer. Look for per-field source references, an exception queue, change history, template locking, access controls and an approval gate that prevents an unreviewed send. Confirm where uploaded data goes and whether third-party connections fit firm policy. A glossy sample proposal is not a useful test of extraction quality.

A general assistant can be a sensible first pilot when the firm already has a checked spreadsheet and simply needs clearer explanation. In that case, paste only the approved, minimised material into an approved workspace and store the final document in the firm's normal system. OpenAI says business workspace data is not used for training by default, but that is only one part of a data decision; access, retention, contracts and internal permissions still matter. A broader wealth-management software evaluation helps when the team is ready to compare full platforms.

Use a trial scorecard rather than a single “time saved” claim. For ten representative proposals, record extraction corrections, arithmetic discrepancies, narrative edits, missing disclosures, reviewer time and any document returned for rework. Keep the denominator visible: eight clean statements and two complex ones are not the same test as ten clean examples. The scorecard tells you whether to expand from drafting to extraction, keep the task narrow or stop.

The final release gate belongs to the firm

Before sending, the adviser should open the exact file the client will receive and read it alongside the source ledger and approved recommendation. Confirm the client's identity, accounts and objectives; trace every number; inspect tables and footnotes; check the IPS and disclosure versions; and confirm that the proposed action is the one the adviser actually endorses. Record who approved the file and when. If any calculation changes, regenerate the document and repeat the final pass.

AI earns its place when it reduces re-keying and makes a sound recommendation easier to understand. The defensible workflow keeps source evidence, arithmetic and professional judgement visible at every step. If the proposal cannot explain where a figure came from or who authorised the recommendation, its polished wording is a reason to slow down, not to press send.

Questions advisers ask before sending a draft

Can AI choose the proposed portfolio allocation?

It can describe an allocation supplied by the adviser and flag apparent differences from the current portfolio. The adviser must decide whether that allocation fits the client's objectives, capacity for loss, constraints and the firm's approved process. A plausible paragraph is not evidence that a recommendation is suitable.

Should I upload a prospect's full statement to a general assistant?

First check the firm's approved tools, data-processing terms, retention rules and the prospect's permissions. If the assistant is not approved for confidential statements, use a redacted or synthetic example to draft the structure and complete the real figures inside the approved environment.

What is the fastest useful arithmetic check?

Reconcile every extracted holding to the statement, add positions plus cash to the stated account total, then independently calculate weights and annual fees from verified values. An apparent one-cent rounding difference is different from a missing account, stale valuation date or duplicated holding.

Further reads

• AI Tools for UHNW Wealth Managers: Seven Jobs With Privacy Rules — Consider data permissions in sensitive adviser workflows.
• AI for Wealth Management Firms: A 90-Day Rollout Plan for Small Firms — Pilot proposal drafting within a controlled firm rollout.
• Best AI App for Wealth Management: Jump, Zocks and Wealthbox — Compare adjacent adviser tools for notes and client workflows.
• AI in KYC for a Small Wealth or Advice Firm: Checks That Stay Human — Keep identity and client checks separate from drafting.
• AI Proposal Software vs a General AI Assistant: Which to Pay For — Choose between a workflow platform and a general writing assistant.
• AI Tools for Asset Management: Factsheet Commentary With Sign-Off — A month-end workflow for factsheet commentary: source-only drafting, a check on every figure, and a named approver who signs a log.
• 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 and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Feathery Wealth Management Proposal Automation (feathery.io/wealth-management/proposal-generation); Investipal Information for AI Assistants and proposal feature pages (investipal.co/ai/); Tecla AI Investment Proposal & IPS Agent (tecla.io/ai/finance/wealth-management/ai-investment-proposal-ips-agent); OpenAI Business data privacy (openai.com/business-data/). Accessed 4 November 2026.