Using AI in Financial Planning: Tax Return Data Into a Client Plan

A financial planner can use AI to extract candidate figures from a client's tax return into a source-linked worksheet, then move only checked figures into the plan. Confirm the client, tax year, schedules, units and signs against the original pages; reconcile totals; and ask which past amounts are likely to recur. The adviser signs off on every planning assumption.

The key separation is between a historical fact and a forward-looking assumption. A return can show that a client received $18,400 of interest last year; it cannot establish that the same income will recur. Keep the extracted value, its page reference, the client's explanation and the plan assumption in different columns. That small discipline prevents a polished AI summary from silently becoming advice.

Follow me on Instagram@sagnikteaches

Decide what the return is allowed to tell the plan

Start with a narrow question: which parts of the existing plan may need updating because the return contains new evidence? Typical candidates are earned income, interest, dividends, realised gains, business or rental activity, retirement distributions, charitable giving and carry-forward items. The return can also expose accounts or entities absent from the discovery questionnaire. It does not provide a current account balance, a future rate of return or a complete picture of spending.

Connect on LinkedInSagnik Bhattacharya

Write a field list before uploading anything. For each field, define the document location, what counts as a match, how material it is to the plan and who may approve it. A household with straightforward wages may need a dozen fields. A household with several schedules may require a longer review. Do not ask AI to "find all planning opportunities" first: that tends to produce conclusions before the team has checked whether the inputs are complete. The broader adviser software evaluation tutorial covers vendor selection; this process starts after a tool has passed the firm's checks.

Subscribe on YouTube@sagnikteaches

For an illustrative client, the previous plan records $110,000 in salary, $3,000 in bank interest and no rental income. The newly received return appears to show $116,200 in wages, $5,640 in interest and a $9,800 rental loss. Those differences are prompts to investigate, not plan updates. The interest may reflect a temporary deposit after a property sale; the loss may include a one-off repair. The right next output is a question list, not an automatic change to the cash-flow forecast.

Set a permission boundary around client documents

Before processing a real return, check what the client authorised, what the engagement covers and which system the firm has approved for sensitive records. Limit access to the people handling that household. Confirm the vendor's current retention, deletion, subcontractor and model-use terms through the firm's procurement process. A product's marketing claim that it serves advisers is not itself permission to upload a client's complete return. If the terms are unclear, test the workflow with a synthetic return and wait for a decision.

Give every document a stable name such as household ID, document type and tax year, but keep names and identifiers out of prompts unless essential. Store the original in the normal document repository and retain a link or reference to it in the review worksheet. Do not create a second uncontrolled copy in a team chat. If a client sent a password-protected file, follow the firm's existing handling method; a failed AI upload is not a reason to remove protection casually.

An illustrative intake failure: a planner receives a corrected return three days after the first upload. The two PDFs share the same filename. If the extraction record says only "latest tax return", nobody can tell which version produced the figures. Give the corrected file a version identifier, mark the original extraction superseded and rerun only the affected checks. Preserve the reason for the change, because a later reviewer may need to understand why a plan assumption moved.

A specialist tax-return product may fit this task. Holistiplan describes tax reports and scenario analysis for financial advisers. FP Alpha says its platform reads uploaded tax returns and summarises key data. Those are verified product descriptions, not accuracy guarantees. Ask each vendor to demonstrate how the actual output exposes source pages, corrections and export fields before buying. The specialist comparison handles the product choice in more detail.

Make a source-linked extraction sheet

The extraction sheet is the hand-off between document processing and advice. Its purpose is to make every proposed plan input inspectable without forcing a second reviewer to search the entire PDF. Include the field name, extracted value, period, source page, status, reviewer initials, correction and planning relevance. Keep a separate column for the eventual plan assumption. A blank assumption is a valid result when the client has not answered a question.

Every figure in that table is illustrative. A page reference should point to the actual client's PDF page, which may differ from the printed form page number. If the tool supplies only a summary without page references, add them during human review. If it merges multiple schedules into a single figure, keep the components visible until the reviewer agrees that the total is correct. A sum that happens to match is not enough if a component is assigned to the wrong person or year.

You can use the following prompt in an approved environment after stripping unnecessary personal identifiers. It asks for evidence, not conclusions. A general assistant may still miss an attachment, so the output remains a draft.

Extract only the fields in my approved list from this one tax-return file.
For each field give: value, unit, tax year, PDF page, printed label,
and any ambiguity. Preserve minus signs and distinguish blank from zero.
Do not recommend a strategy or fill missing values from other years.
If a figure is only on an attachment, name that attachment.
Return a table plus a list of fields you could not locate.

An illustrative output might say, "Interest income: $5,640, page 2, high confidence." The reviewer opens page 2 and sees $5,640 on the return, but the supporting statement shows $5,460. That is not a harmless formatting difference. The team should check whether the return includes a second account, whether the statement belongs to another period and whether the AI cited the correct line. Mark the field unresolved until the discrepancy is explained. Confidence labels are useful for sorting work, not for waiving it.

Reconcile the return before interpreting it

Use two passes. In the first, compare each extracted number and sign with the visible source. In the second, test relationships among figures. Do the return's income components reconcile to its reported income total under that return's own structure? Does the plan import show the same period and currency? If a value is derived, retain the calculation as well as the original inputs. Never have the model invent a missing figure to force a total to match.

Rank fields by their effect on the plan. An omitted $40 bank-interest entry may still need correction, but it might not delay a meeting. A mistaken $40,000 distribution or a negative sign lost from a rental result can materially change a cash-flow discussion. Set an internal rule that high-impact or ambiguous fields require a second pair of eyes. The threshold should reflect the firm's work, not a universal dollar amount. Record the threshold and the exceptions so reviewers can apply it consistently.

Here is a small illustrative reconciliation. An extraction shows $116,200 wages, $5,640 interest, $4,100 dividends and $42,000 realised gain, giving $167,940 when simply added. The return's reported total is $168,740. The $800 gap is not permission to label the output "approximately correct". Search the source for another income line, check whether the gain figure is net or gross and identify any amount counted elsewhere. If the relationship cannot be explained from the document, put the total on hold and ask the client's tax preparer.

Watch for four recurring document errors. Optical recognition can read 8 as 3; a minus sign can disappear at a page margin; a two-column schedule can join the wrong row and value; and a return packet can contain both an original and amended version. A reviewer should inspect the pixels, not merely compare two AI-generated tables. The distinction between return review and software diagnostics matters here: a correct extraction can still rest on a return that needs professional tax review.

Carry the figures into a plan without copying the past forward

Once the source values are checked, create a separate assumption review. Each entry should answer: is this a historical fact, a likely recurring cash flow, a one-off event, or an unresolved item? Then ask what current evidence would change that classification. A planner might confirm salary from a current pay statement, interest from account statements, expected distributions from a client conversation and a rental forecast from records. The return is one input among those sources.

A second illustrative case shows why the separation matters. The return records a $42,000 gain from a sold asset. AI proposes using $42,000 as an annual "investment income" assumption because it appears in an income section. The adviser rejects that import: realised gains depend on transactions, and the asset may no longer exist. The plan retains the gain as a historical note and asks about proceeds, current holdings and any future sale already planned. The old figure may inform a discussion, but it does not become a forecast by default.

Another case: a self-employed client has $68,000 of business profit on the return. They report a new contract and expect $92,000 this year. Do not replace the return value with the client's estimate inside the extraction sheet. Preserve $68,000 as the verified historical source, put $92,000 in the scenario column with the date and source of the client's statement, and test a lower case if the work is uncertain. The arithmetic is easy; the judgement about recurrence belongs to the adviser and client.

Keep the plan's assumptions and tax calculations distinct. Moving checked evidence into a financial plan does not settle whether a particular transaction is tax-efficient. If the plan raises a tax interpretation, relief, carry-forward or filing question, involve a qualified tax adviser before presenting it as a recommendation. The boundary between adviser administration and advice is especially useful if the firm wants to automate more of the workflow.

Work one household through the whole hand-off

Consider an illustrative planning firm reviewing a household's completed return. The existing plan assumes $110,000 wages, $3,000 interest and $12,000 annual retirement distributions. The new return contains $116,200 wages, $5,640 interest, $4,100 dividends, a $42,000 realised gain and a $9,800 rental loss. The AI extraction also lists a $21,000 retirement distribution. The reviewer records page references for all six items and notices that the distribution appears on two pages with different labels.

First, the reviewer checks the client and tax year on the return and confirms that the packet includes the relevant schedules. They compare the extraction with each printed field. The $21,000 distribution is actually a total that includes a $9,000 rollover recorded elsewhere. The planning worksheet therefore marks $21,000 as the return's reported total, but it does not label the entire amount as spendable income. The reviewer notes the ambiguity and asks the client and tax adviser to explain the transaction. AI's initial label is corrected in the log, with the source pages retained.

Second, the adviser asks four questions. Is the $116,200 pay level still current? Did the interest arise from cash that remains in the account? Was the $42,000 gain linked to a one-off sale? Did the rental loss reflect a continuing operating shortfall or an unusual repair? The client's answers are illustrative too: current pay is $120,000, half the interest came from a temporary deposit, the sold asset is gone, and $7,000 of the rental loss came from a one-off repair. None of those answers alters what the return said; each informs a separate planning assumption.

Third, the adviser updates the plan with the agreed current pay assumption of $120,000, leaves the gain out of recurring income, and models rental cash flow using current evidence rather than simply repeating -$9,800. The interest assumption remains pending until an account statement confirms the current balance and rate. The retirement distribution remains pending until the transaction is clarified. A client-facing note says which items were updated, which were held, and which require a tax adviser. The result is deliberately less complete than an AI-generated "finished plan" because unresolved inputs should remain visible.

Time and cost vary by document complexity and tool contract. For a pilot, measure the team's own minutes for upload, extraction review, client questions and plan updates on comparable households. Do not count only the seconds to generate a report. If a specialist subscription is being considered, ask for a quote tied to the expected household volume and the firm's review needs; current product pages are not a substitute for a current commercial proposal. A cheap extraction that creates a long correction queue may cost more in staff time than a slower but clearer process.

Write a client question log that can be answered

A raw list of differences is hard for a client to use. Convert each unresolved item into a question that names the observed figure, asks for a current fact and avoids suggesting an answer. Keep tax-return details in the secure client channel. The question log should have an owner and a deadline. If the answer arrives by telephone, record who spoke, when, and the exact assumption approved. Do not ask an AI model to infer the client's intention from a prior meeting note.

For example, instead of "Why did your income increase?", an illustrative message might read: "The completed return shows $5,640 of interest, compared with $3,000 in your previous plan. Is the account that generated this interest still held, and can you share a recent balance statement?" That question identifies the evidence needed for the next step without implying that the client did anything wrong. If the client replies only "yes", the planner still needs to clarify whether "yes" means the account exists or the figure is recurring.

A filled-in log entry could read: "Field: rental result; source: return PDF page 11; extracted: -$9,800; checked: yes; current status: client says $7,000 was a one-off repair; plan action: hold recurring rental assumption pending current rent and expense figures; reviewer: adviser; decision date: pending." This small record is more useful than a paragraph of AI-generated prose because it identifies the exact gap. The broader tax-return preparation review workflow handles the preparer's separate sign-off; the planner still owns the plan assumption.

Test the process on mistakes, not just clean documents

Before scaling to a whole client book, run a small pilot with permission and controlled access. Include an ordinary return, one with attachments, one with an amended page and one with a poor scan. Keep a manually reviewed reference sheet for each. Ask the tool to extract the same approved fields, then count incorrect values, missing fields, wrong years, wrong signs and absent page references. Record time to correct each error. A test that reports only "the summary looked good" will miss the risks that matter.

For an illustrative pilot of ten returns, suppose the tool outputs 120 candidate values. Reviewers find 112 correct, five wrong and three missing. The useful question is not merely whether 112 out of 120 is a high percentage. Which five were wrong? If one was a large distribution with a lost minus sign, the process needs stronger controls before live use. If the three missing values all sat in the same attachment type, alter the intake and checklist. Track the denominator and the consequences, not an unexplained accuracy score.

A final check should compare the plan against the reviewed worksheet, not against the AI output. For every changed assumption, can a second adviser identify the return page, the client confirmation and the decision-maker? Can they see that an unresolved field stayed unresolved? Can the firm reproduce the plan if the AI vendor changes? If the answer is no, improve the record before increasing volume. The document-to-plan hand-off is successful when the adviser can explain every figure without trusting a black-box summary.

Questions before using a return in a plan

Can a tax-return summary replace reading the return?

No. Treat the summary as an index to figures and possible questions. A reviewer should compare each material value with the original page, confirm the tax year and household, and inspect attached schedules before any figure enters the planning record. A plausible summary can still omit a schedule, misread a minus sign or mix years.

Should the adviser put the whole return in a general chatbot?

Only if the firm's client agreement, confidentiality obligations and approved tool policy allow that specific upload. A safer starting point is a firm-approved document workflow with controlled access. If the tool has not been approved for client records, use redacted synthetic data to design the process and keep the actual return in the existing secure document system.

Does a figure on last year's return become this year's plan assumption?

Not automatically. A return describes a completed period. Ask whether pay, interest, distributions, business income and realised gains are expected to recur. Record the source figure separately from the forward assumption, its date and the client's confirmation. A qualified tax adviser should review tax interpretations or proposed tax actions.

Further reads

• Tax Workpaper Prep With AI: Tie the Return Back to Source Documents — Trace return figures back to supporting documents.
• AI Software for Tax Professionals: Run a Pilot and Measure Cost per Return — Plan a measured document-processing pilot.
• ChatGPT for Tax Preparers: How to Redact a Return Before You Paste — Design a redaction process before testing general AI.
• Free AI for Tax Professionals: Where the Free Tier Stops at Client Data — Examine free-tier boundaries for sensitive client material.
• Holistiplan Pricing: Basic vs Premium Tax and Who Needs Which — Holistiplan Basic Tax vs Premium Tax priced by household tier from 30 to 750, with a Premium break-even, the billing rules and a 7-day trial checklist.
• Year-End Tax Planning Letters With AI: Personalise Without Errors — A tax practice's method for personalised year-end letters: data merged from your records, AI used for wording only, every figure checked before sending.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Holistiplan home page (holistiplan.com) and FP Alpha home page (fpalpha.com), checked 4 November 2026.