Scan every client return for planning opportunities

Traditional vs AI-native

The traditional way

1. A preparer notices something during prep ('this Schedule C is big. Has anyone talked about an S corp?') and leaves a note in the file.
2. The note may or may not reach a partner; if it does, it gets raised at the next client meeting, often months later.
3. Planning ideas depend on which staff member worked the return and what they happen to know.
4. Year-end planning letters go only to the handful of clients who ask or who the partner remembers.
5. Nobody can answer 'how many of our clients should be looking at a PTET election?' without pulling every file.

Where it hurts

• Opportunities are missed systematically for quiet clients who never ask.
• Advice is inconsistent: two similar clients get different recommendations depending on who prepped them.
• The firm's advisory revenue is reactive and hard to forecast.
• Clients learn about a strategy from a neighbor or a competitor and ask why their CPA never mentioned it.

The AI-native way

1. 1You

   Confirm consent and scope before touching data

   A partner confirms, with counsel, which screens fall within the Treas. Reg. §301.7216-2(h) exception for the firm's own other accounting services and which need the client's signed §7216 consent (for example, anything that involves a third party). Clients without a needed consent are excluded from those screens. Consent status is a field in the practice management system and the scan filters on it.

   Tools: Practice management custom field · Consent tracking

2. 2Automation

   Export structured return data for every client

   Pull a per-return data set from tax software (report writer, CSV export, or API where available): filing status, AGI, taxable income, Schedule C net profit, Schedule E and K-1 items by entity, itemized vs. standard deduction, SALT paid, W-2 wages and withholding, retirement contributions, HSA fields, state of residence, and entity links. Where exports are thin, parse the filed return PDFs into the same schema and flag those rows as lower confidence.

   Tools: Tax software report/export · Python + pandas · Document AI for PDF fallback

3. 3AI

   Normalize and link households and entities

   The model matches 1040s to the 1120-S, 1065, and Schedule C businesses they own, using EINs, entity names, and K-1 recipient data, so a screen can see an owner's full picture. Ambiguous links (two entities with similar names, a K-1 from an entity the firm doesn't prepare) go to a review list rather than being guessed. Output is a household-level JSON record per client.

   Tools: LLM with structured output · Entity/EIN lookup table

4. 4Automation

   Run deterministic planning screens

   A versioned rules library runs against every household. Examples: Schedule C net profit above a firm-set threshold and no S election (S-corp candidate); pass-through income in a state with an elective PTET where the owner itemizes or is phased down on SALT (PTET candidate); self-employment income with no retirement plan contributions (Solo 401(k)/SEP), or stable owner income high enough to consider a cash balance plan; taxable income well below the client's recent average (Roth conversion window); C-corp stock held by the client or an original-issue founder (possible §1202 QSBS review); HDHP coverage indicated without HSA contributions; balance due plus an estimated-tax penalty (underwithholding). Every threshold lives in a table with a source and effective year, never in model memory.

   Tools: Python rules engine · Versioned threshold table (YAML/CSV)

5. 5Automation

   Compute savings ranges in code

   For each triggered screen, a deterministic calculator produces a low/high savings range with its assumptions. The S-corp calculator, for example, runs reasonable compensation at two levels, computes employer and employee payroll tax, the §199A effect of W-2 wages, added payroll/processing and return-prep costs, and any state entity-level tax or minimum fee from the state table. Retirement calculators use the current-year contribution limits from the maintained table. The model never produces a dollar figure.

   Tools: Python calculators · Annual limits table

6. 6AI

   Apply judgment and rank per client

   The model reviews each client's triggered screens with the underlying data and prior-year notes to catch what rules can't: income that looks one-time (a large capital gain driving a false 'high earner' flag), a business winding down, a client already known to have a plan at another provider. It assigns a confidence level, writes a two-sentence rationale citing the specific data fields, lists the questions a human must answer, and ranks opportunities by the code-computed range and confidence.

   Tools: Enterprise LLM · Prior-year notes (RAG)

7. 7You

   Partner review of the ranked list

   The responsible partner or manager works a queue sorted by estimated value: accept, reject with a reason, or mark 'needs data'. Rejection reasons are captured as structured codes so screens that produce noise can be tightened. Nothing reaches a client without this approval.

8. 8AI

   Draft the one-page client summary

   For approved opportunities only, the model drafts a plain-English page: what we noticed (with the client's actual figures), what it could be worth (the code-computed range, labeled as an estimate), what it would take, and what we need to know. It avoids guarantees and includes a line that the estimate depends on facts to be confirmed. The partner edits and approves.

   Tools: LLM drafting · Firm letter template

9. 9Automation

   Create engagements and track the pipeline

   Each approved opportunity becomes a proposed engagement in practice management with the estimated value range, owner, and a follow-up date. A dashboard shows the firm-wide pipeline by opportunity type, conversion rate, and realized fees, and feeds next year's threshold tuning.

   Tools: TaxDome / Karbon / Canopy · BI dashboard or spreadsheet

What you need

• Structured return data for every client (report writer, CSV export, or API)
• Counsel's read on which screens fit the §301.7216-2(h) exception and which need signed §7216 consent
• A versioned table of thresholds and limits, each with a source and effective year
• Deterministic calculators for the savings ranges
• A partner who works the review queue

The stack

Named tools are examples, not endorsements. Check any tool against your own data rules first.

Copy the prompt

You are a senior tax planning reviewer at a US CPA firm. You will receive one client household as JSON:
- HOUSEHOLD: return data fields (filing status, AGI, taxable income, Schedule C/E, K-1s, SALT, withholding, retirement and HSA fields, state), with a data_source and confidence per field.
- SCREENS: planning screens that fired, each with screen_id, the fields that triggered it, and a savings_range computed by the firm's calculator (low, high, assumptions).
- NOTES: prior-year preparer and partner notes, if any.

Tasks:
A. For each screen, decide: "likely", "possible", or "unlikely" for this client, using only the data provided. Explain in at most two sentences, citing the specific field names and values.
B. Flag any reason the screen may be a false positive: one-time income, a business that appears to be winding down, missing data, a note saying the client already has a plan or has declined before.
C. List the specific questions a human must answer before recommending it (e.g., "Does the client have employees who would need to be covered by a retirement plan?").
D. Rank the screens you rated "likely" or "possible" by the high end of savings_range, then by your confidence.
E. Note any obvious opportunity the screens did not catch, as a question only, with no dollar amount.

Rules:
- Never compute, change, or invent a dollar amount. Use savings_range exactly as provided or say "not computed".
- Never state tax law thresholds, limits, or phase-outs from memory. If a threshold matters, name it as a field the reviewer should check.
- Do not recommend any action as certain; this is a screening aid for a credentialed reviewer.
- If a field has confidence below 0.8, say so and lower your rating by one level.
- Output JSON only:
{ "client_id": "", "opportunities": [ { "screen_id": "", "rating": "", "rationale": "", "false_positive_risks": [], "questions": [], "savings_range": { "low": 0, "high": 0 } } ], "ranked_screen_ids": [], "unscreened_ideas": [], "data_gaps": [] }

Example

Fictional data.

client: Marisol Vega (MFJ), state: GA, itemizes: yes
Schedule C (landscape design): net profit $212,400 (2024), $188,900 (2023)
No S election on file; no retirement contributions; SE tax $28,950
W-2 (spouse): $64,000; withholding covered 71% of total tax; §6654 penalty $410
SCREENS FIRED:
 SCORP_01 savings_range $6,800–$13,900 (comp $85k–$110k, payroll + prep cost $3,400)
 SOLO401K_01 savings_range $9,200–$16,700 (at 24%/32% marginal)
 WITHHOLD_02 penalty avoidance $400–$600
NOTES: "Client mentioned hiring a part-time crew lead in 2025."

1. SOLO401K_01 (likely). Sched C profit $212,400 and $0 retirement contributions.
   Risk: note says hiring a crew lead; a Solo 401(k) may not fit if an
   eligible employee is hired. Question: hire date and expected hours?
2. SCORP_01 (likely). Profit above threshold two years running ($188,900, then $212,400).
   Question: willing to run payroll? If elected, run the state PTET screen next.
3. WITHHOLD_02 (likely). Penalty $410; set 2025 estimates to prior-year safe harbor.
unscreened_ideas: "Would an SE plan decision change if the business becomes an
   S corp mid-year? Coordinate SCORP_01 and SOLO401K_01 calculations."
data_gaps: none

Guardrails

• IRC §7216 is the first gate. Using return information to identify clients for and offer services other than preparing that return (planning engagements, retirement plan setup, entity conversions) generally requires the taxpayer's prior written consent under Treas. Reg. §301.7216-3, in the format Rev. Proc. 2013-14 describes for 1040 clients. Treas. Reg. §301.7216-2(h) lets a CPA firm use return information for its own other accounting services to that client, so have counsel decide which screens fall inside that exception and which (for example, referrals to a third-party plan provider) need consent. Build consent into the engagement letter and have counsel review before running the scan.
• Run everything under an enterprise AI agreement with zero data retention or a no-training commitment, and add the scan, its data exports, and where they are stored to your Written Information Security Plan under the FTC Safeguards Rule.
• Savings ranges come only from deterministic calculators fed by a maintained limits and thresholds table with a source and effective date for every value. The model ranks and explains; it does not calculate or recall limits.
• Treat law-sensitive screens as dated rules: several individual energy credits were cut off early by Pub. L. 119-21 (July 4, 2025), for example §25C and §25D after 2025, and §1202 changed for stock acquired after July 4, 2025. Every rule carries an effective-date range and an owner who confirms it against current guidance each year.
• Client summaries are estimates, labeled as such, and approved by a credentialed reviewer. If you give written advice on a recommendation, Circular 230 §10.37 standards apply: reasonable factual assumptions, relevant facts considered, no reliance on unreasonable representations.

What can go wrong

How to prove it worked

Where this goes

The rules library becomes the firm's institutional planning knowledge: every strategy a partner knows is encoded once, dated, and applied to every client every year. Run it continuously as documents arrive during the year rather than once after filing, and the firm starts calling clients before year-end with specific, quantified moves instead of explaining after April what they could have done.

Demo

References

• IRC §7216 and Treas. Reg. §§301.7216-2 and 301.7216-3 (use and disclosure of tax return information; consent requirements)
• Rev. Proc. 2013-14 (form and content of §7216 consents)
• IRS Notice 2020-75 (deductibility of state and local income taxes paid by partnerships and S corporations)
• IRC §1202 (qualified small business stock) and IRC §6654 (failure to pay estimated tax)
• Treasury Department Circular 230, §10.37 (requirements for written advice)

Summaries for orientation. Verify against the current standard before relying on any of them.

Keep browsing

• Large impactTax

  Turn a pile of client uploads into a complete, pre-filled tax file

  Get from whatever the client uploaded to an indexed, extracted, complete file, with a specific list of what is still missing.

  Saves 30 to 60 minIntermediateWorkflow
• Large impactCAS and bookkeeping

  Run month-end close with an agent doing the first pass

  Close a client's books each month with an agent that reconciles, categorizes, proposes adjusting entries, and prepares the review package, while the accountant approves every change.

  Changes the workflowAdvancedAgent
• Large impactClient service

  Build a cash dashboard around what one owner actually cares about

  Give a business owner a live view of their cash built around their own questions (can I make payroll, when can I buy the truck, how much can I take out) instead of a standard report package.

  Changes the workflowAdvancedClient experience