From AI Enthusiasm to Accountable Practice: A 90-Day Family Office Pilot

1. Define the decision before selecting the tool

Begin with a recurring internal task whose output can be checked and whose failure is reversible. One example is preparing a first draft of a briefing from a fixed set of approved documents. Do not bundle that pilot with authority to transfer funds, commit capital or send unchecked client communications. Name a process owner, a reviewer with relevant expertise and a person who can stop the pilot.

Write a one-page charter: the task, intended users, permitted input information, output recipient, approval boundary and reason for trying AI. Add the human purpose the office expects improved capacity to serve. “More efficiency” is too vague to evaluate. A useful objective might be reducing clerical preparation while preserving time for examining a decision’s unresolved assumptions.

Use the conceptual article’s question about stewardship as a design prompt, not as evidence that this protocol is effective. A family that values more reflective time should decide what that means in its own routines. No individual should be required to disclose health information to demonstrate that time has been well used.

2. Days 1–15: establish an inspectable baseline

Collect a representative set of eligible tasks from normal operations. Record complexity, source availability, the people involved and the complete time required for preparation, checking and correction. Define what makes an output acceptable before reviewing AI-assisted work. For document briefings, a material factual claim should point to supporting source material, important qualifications should be retained, and numerical statements should be checked independently.

Count both ordinary and difficult cases. If only easy work enters the pilot, its apparent performance will not describe the office’s real workload. Use de-identified or explicitly approved information. Record the permitted environment and access rules; confidentiality obligations do not disappear because a task is experimental.

The baseline need not be large enough for a causal research claim to be useful operationally. It does need to be honest about uncertainty. Document the number and kinds of cases available. If the office later seeks a publishable empirical evaluation, specify a justified sample design, consent where required and appropriate independent review before collecting additional participant data.

3. Days 16–45: compare complete workflows

Where operationally feasible, assign comparable cases between the existing process and the AI-assisted process using a pre-specified random allocation. Balance task difficulty and reviewer experience. Avoid judging an AI condition staffed by the strongest analyst against a baseline staffed by a novice. If randomization is impractical, use a clearly described matched comparison and retain the limits on causal interpretation.

Preserve the input packet, tool and model version, relevant settings, draft, corrections and approved final output. Ask reviewers to score outputs without knowing their production method where feasible. Separate factual correctness from style and confidence. A polished paragraph should not offset a material error.

Use net staff time per acceptable output as one operational measure: preparation plus verification plus rework, divided by the number of outputs meeting the pre-specified standard. Also track unsupported material claims, numerical errors, missed qualifications and correction severity. Record licensing, integration and training costs separately; saved minutes are not automatically cash savings.

The workplace findings in Brynjolfsson, Li and Raymond [1] and the task-dependent effects in Dell’Acqua and colleagues [2] motivate local testing. Their effect sizes are not the expected result of this pilot. Record results by task category so that a gain on routine summaries does not conceal a failure on ambiguous evidence.

4. Require an evidence handover

Before a work product reaches a decision maker, require an evidence note containing its important claims, source locations, checked calculations and unresolved questions. Name the reviewer and record the disposition: approved, corrected, returned for more information or rejected. Ask what evidence would change the recommendation. A citation that opens a document but does not support the nearby claim does not count as verification.

The reviewer should inspect original material rather than ask the same model to certify its own answer. For material decisions, provide a route to independent specialist advice. Record disagreement without treating it as a failure of team alignment. A person who raises an uncertainty should not have to overcome the social authority of a polished automated document.

NIST’s profile is a useful external reference for evaluating generative-AI risks [3]. This local evidence note is an editorial proposal inspired by those concerns, not a claim of NIST certification or an exhaustive compliance procedure.

5. Days 46–75: allocate capacity and test continuity

If acceptable output improves without unacceptable errors or costs, decide explicitly how to use any recovered capacity. Record the intended allocation, such as deeper review, staff learning or protected time. Check the actual allocation through ordinary work records and voluntary feedback. Do not infer wellbeing from an empty calendar or turn a family’s private relationships into a surveillance exercise.

Test whether the workflow functions when its usual champion is absent. Identify who can retrieve the source record, finish the task and respond to an error. A system that frees a founder’s schedule while making every exception depend on that founder has not solved the continuity problem.

For an internal family-purpose conversation, ask three practical questions: what do we want this office to make possible; which activities currently crowd that out; and who is missing from the discussion? Record disagreements and decisions at the appropriate level of confidentiality. This is an invitation to governance, not a clinical assessment or a requirement that every family agree on the same goals.

6. Days 76–90: continue, revise or stop

Set decision thresholds before reviewing aggregate results. The office should define a tolerable error profile appropriate to the task, minimum acceptable evidence quality, a cost boundary and conditions that trigger suspension. A material confidentiality incident or an unchecked consequential action requires immediate attention rather than waiting for a favorable average at the end of the pilot.

At the closing review, show the distribution of outcomes, not just a mean speed improvement. Identify failures, exceptions, reviewer effort and the cases excluded from analysis. Choose among continuing the bounded workflow, revising and retesting it, or stopping it. Retain the rationale and the next review date. Expansion to a new task requires new evidence, even if the interface and model stay the same.

The ultimate record should answer two separate questions: did the workflow perform better on its defined measures, and did the institution use the change for a purpose it had chosen? A positive answer to the first does not establish the second. The distinction makes Swire’s question about the purpose of wealth operational without reducing that purpose to a financial metric.

7. Limitations and reporting

This is a proposed protocol, not a completed primary study. Ninety days is an organizing horizon rather than a validated duration. A small family-office pilot will usually be underpowered for broad statistical claims, and its results may not generalize across models, people, tasks or market conditions. Changes in supervision or attention can also affect performance.

Report the exact workflow, model version, evaluation dates, case selection and deviations from the plan. Publish aggregate operational findings only when authorized and appropriately de-identified. Avoid disclosing family assets, private relationships or commercially sensitive source documents. If reporting errors reveals a weakness, preserve that finding rather than rewrite the evaluation as a success story.

References and source record

1. Brynjolfsson, Erik; Li, Danielle; Raymond, Lindsey (2025). Generative AI at Work. The Quarterly Journal of Economics, 140(2), 889–942. https://doi.org/10.1093/qje/qjae044

2. Dell’Acqua, Fabrizio; McFowland III, Edward; Mollick, Ethan; Lifshitz, Hila; Kellogg, Katherine C.; Rajendran, Saran; Krayer, Lisa; Candelon, François; Lakhani, Karim R. (2026). Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality. Organization Science. https://doi.org/10.1287/orsc.2025.21838

3. Autio, Chloe; Schwartz, Reva; Dunietz, Jesse; Jain, Shomik; Stanley, Martin; Tabassi, Elham; Hall, Patrick; Roberts, Kamie (2024). Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile. NIST AI 600-1. https://doi.org/10.6028/NIST.AI.600-1

4. Opal Group (2026). Family Office & Private Wealth Management Forum West. Event context: Napa, 21–23 October 2026. https://opalgroup.net/finance/family-office-private-wealth-management-forum-west-2026/

Editorial contribution and provenance

Paired conceptual article: Beyond the AI Margin: Human Judgment and the Purpose of Family Wealth. https://journal.research.university/journals/family-office-private-wealth/articles/napa-2026-ai-human-judgment-purpose-of-wealth. The proposed pilot has not been tested.

Forum source: an attendee-supplied excerpt, received 21 October 2026, attributed by the attendee to Andrew Roettger (speaker) and Sophia Swire (moderator). It is not a complete or timestamped transcript. The event page verifies forum context, not individual statements. The supplied excerpt is paraphrased selectively; no complete recording or transcript is reproduced.

Prepared by the journal editorial research desk at Dr. Freedom Cheteni’s direction with AI assistance for synthesis, source checking and drafting. The speaker and moderator are credited as participants, not manuscript authors or endorsers. This is an editorial analysis, not an independently peer-reviewed research article. No new participant data, study results or DOI are claimed.

Implementation record

Open the cited sources

1. [1] Brynjolfsson, Erik; Li, Danielle; Raymond, Lindsey (2025). Generative AI at Work. The Quarterly Journal of Economics, 140(2), 889–942. ↗
2. [2] Dell’Acqua, Fabrizio; McFowland III, Edward; Mollick, Ethan; Lifshitz, Hila; Kellogg, Katherine C.; Rajendran, Saran; Krayer, Lisa; Candelon, François; Lakhani, Karim R. (2026). Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality. Organization Science. ↗
3. [3] Autio, Chloe; Schwartz, Reva; Dunietz, Jesse; Jain, Shomik; Stanley, Martin; Tabassi, Elham; Hall, Patrick; Roberts, Kamie (2024). Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile. NIST AI 600-1. ↗
4. [4] Opal Group (2026). Family Office & Private Wealth Management Forum West. Event context: Napa, 21–23 October 2026. ↗