A statement arrives with a familiar surname and the amount a preparer expects. The extraction is accurate. The worksheet agrees.
The statement belongs to a different household.
Arithmetic will not expose that mistake. The team must establish why the evidence supports this client, entity and period.
The next step in the series
Our Tuesday field note on consistency and authenticity separated those two checks. This question goes further: even an authentic, accurately read document can be used in the wrong engagement.
Our earlier reviewer-access test asked who may inspect the source. The API handoff test asked what survives a transfer. Now ask what makes the association justified.
Define the acceptance rule before the demonstration
Have the practice leader identify the evidence needed to associate a document with an engagement. Which client, entity, account and period details matter? Which conflicts require review? When should the workflow leave an item unresolved?
A filename, folder location or familiar total should not silently become the whole explanation. Equally, do not invent a universal matching rule: the practice determines what the engagement requires.
NIST's voluntary AI Risk Management Framework supports documented test conditions and evaluation relevant to intended use. It provides a reason to test confusing inputs, not only clean demonstrations; it does not prescribe or certify this exercise. NIST AI RMF Core, MEASURE 2.
Run three fictional cases in twenty minutes
Use an approved demonstration environment and entirely invented records. Keep genuine taxpayer files out of the exercise. All cases below are proposed and have not been run.
1. Minutes 0–5: establish the baseline. Create two fictional engagements with similar display names but different entity and account references. Add a current-period statement to one. Ask the preparer to show the association evidence using the firm's existing process.
2. Minutes 5–10: introduce an ambiguous source. Keep the surname and amount familiar but change the fictional entity or period. Observe whether the disagreement remains visible. An appropriate unresolved question is a useful outcome; forcing a match merely to avoid an exception is not.
3. Minutes 10–15: replace a source. Add a corrected statement after an earlier review. Ask which prior conclusion rests on the old version and who must reconsider it. If the system produces a summary, inspect its supporting engagement and version too.
4. Minutes 15–20: change reviewers. A second authorized person should identify the intended engagement, permitted source, unresolved question and next owner without a verbal walkthrough. Ask them to distinguish a proposed match from a practitioner-reviewed disposition.
The IRS Office of Professional Responsibility describes the risk of one client's context influencing an AI answer about another. Its June 2026 guidance also emphasizes review of AI-produced material. That makes derived summaries part of the review question, not just the uploaded file. IRS OPR Alert 2026-19.
Record the decision, not just the result
Use one observation row per case:
Expected engagement and period | observed association | source/version inspected | reviewer/date | unresolved question | next owner/action
Count incorrect associations separately from appropriate abstentions. Record missing explanations, stale dispositions and total verification effort. An unobserved outcome stays unknown. Compare another tool only against the same acceptance rule and comparable fictional cases; a reviewer who already knows the answer can distort the comparison.
A small demonstration can reveal a handoff worth improving. It cannot establish a production accuracy rate or prove that every client and document type is handled safely.
Where TaxFlo and SafeVault fit
I founded AmitaSoft and built TaxFlo on SafeVault. TaxFlo supports classification, extraction, discrepancies, missing-evidence visibility and authorized prior-year context. The evaluation should show how those capabilities help a practitioner inspect evidence for the intended engagement. We claim no perfect matching, automatic identity certification or demonstrated cross-client leakage prevention. Tax judgment and final approval remain with the practitioner.
SafeVault's broader life-and-business vision depends on earning permission and keeping evidence useful. Deeper CPA advisory relationships and relevant client introductions are opportunities we hope to develop, not outcomes established by this exercise.
Bring one recurring association problem to a 20-minute synthetic-workflow review. We can compare the current process, document the pain and agree a success measure before considering a controlled 10–25 engagement pilot. No taxpayer files are needed.
Try the fictional companion
Use five initial fictional source cards, then introduce a separate correction before the reviewer handoff. Open the printable worksheet and fictional inputs.
The observation sheet starts with no results. All cases remain NOT RUN until actual authorized execution. The facilitator key is separate from the first review. These are editorial examples, not product pass rates, taxpayer documents or tax advice.
Hero: synthetic editorial photograph with fictional documents; not an actual client or product screenshot.