AI Portfolio Commentary: A Client Validation Prompt

AI can turn a performance report into polished talking points in seconds. The difficult part is not producing fluent language. It is preserving the difference between a verified portfolio fact, an analytical interpretation, a scenario, and a personalized investment recommendation.

That distinction matters whenever an advisor uses AI-generated portfolio commentary in a meeting, email, quarterly review, or client portal. A sentence can begin as a neutral explanation and become advice after the model adds an action, a degree of certainty, or an unsupported claim about what the client should do.

AssetMark brought this workflow into focus with the August 2026 launch of Talk Tracks, a feature that combines portfolio information and market context to produce client-ready talking points. Earlier products from firms such as FactSet and MDOTM also showed how generative AI can accelerate portfolio commentary. The next operational question is simple: how should an advisor validate the output before a client sees it?

This guide provides a practical answer: a data contract, a five-class review model, a copyable validation prompt, and a human approval matrix.

What is AI-generated portfolio commentary?

AI-generated portfolio commentary is a narrative created from portfolio holdings, performance, attribution, benchmarks, risk measures, and market information. Depending on the system, it may produce:

a summary of portfolio performance;
an explanation of relative return drivers;
a description of material risk exposures;
client-meeting talking points;
quarterly letters or factsheet language; and
draft responses to client questions.

The workflow can reduce the time spent turning analytical output into readable prose. It can also scale personalized communication across many accounts. But the narrative inherits every weakness in its inputs. A stale position file, mismatched benchmark, incomplete household view, or ambiguous time period can become a confident but incorrect explanation.

The model can add a second class of error: language that goes beyond the evidence. “Technology contributed 1.2 percentage points” is a calculation. “Technology will continue to lead” is a forecast. “You should increase technology exposure” is a recommendation. They require different evidence and different approval paths.

Why portfolio explanation can become investment advice

Client communication sits on a continuum rather than inside a single category. The safest workflow classifies each material statement before approving the final draft.

Statement class	Example	Minimum evidence	Default treatment
Verified fact	“The portfolio returned 3.4% during the quarter.”	Approved performance record, period, currency	May be drafted automatically
Calculation	“Allocation to healthcare added 0.6 percentage points.”	Reproducible attribution output	Recalculate outside the language model
Interpretation	“Healthcare was the largest positive contributor.”	Ranked attribution data	Human review
Scenario	“Higher rates could pressure long-duration assets.”	Named assumptions and limitations	Label clearly as conditional
Recommendation	“Reduce bond duration now.”	Client objectives, suitability context, authority	Block pending authorized review

This classification prevents a common failure: allowing the model to move from a factual observation to an action without showing the missing reasoning in between.

The input contract for reliable portfolio commentary

A prompt cannot repair missing data. Before generation begins, the workflow should provide a controlled input package.

Portfolio scope

Identify the exact account, portfolio, household, or strategy. State whether external or held-away assets are included. Do not let the model describe a household-level position when it only received one account.

Time and valuation basis

Include:

start and end date;
valuation timestamp;
base currency;
gross or net return basis;
settlement and corporate-action status; and
whether the data is preliminary or final.

Benchmark and objective

Name the benchmark, policy target, or comparison basis. A return is not “strong,” “weak,” “ahead,” or “behind” without an explicit frame of reference.

Source identifiers

Every input should carry a source ID that the output can cite. Useful fields include system name, report ID, table or cell reference, extraction time, and reconciliation status.

Client-language constraints

Define reading level, maximum length, prohibited jargon, allowed channels, and whether the draft is for internal meeting preparation or direct client delivery. Internal talking points and send-ready communications should not share the same approval rule.

A copyable AI portfolio commentary validation prompt

Use the following prompt as a reviewer after a separate system has produced a draft. Keeping generation and validation separate makes failures easier to see and measure.

You are validating draft portfolio commentary for a financial advisor.

Your job is to verify the draft against the approved evidence package. You do
not provide investment advice, select securities, change allocations, or fill
missing facts with assumptions.

For every material sentence:
1. Classify it as FACT, CALCULATION, INTERPRETATION, SCENARIO, RECOMMENDATION,
   or UNSUPPORTED.
2. Cite the portfolio/account scope, source ID, reporting period, and as-of
   timestamp.
3. Confirm that every number matches the approved source and uses the correct
   unit, currency, gross/net basis, benchmark, and period.
4. Identify causal language that is not supported by attribution or an approved
   research source.
5. Identify language that implies certainty about future market performance.
6. Identify any sentence that could reasonably be read as a recommendation or
   instruction to trade, rebalance, buy, sell, hold, or change risk exposure.

Rewrite only sentences classified as FACT, CALCULATION, INTERPRETATION, or
SCENARIO. Preserve the meaning of the evidence. Label scenarios as conditional.
Do not rewrite a RECOMMENDATION into softer language; route it to human review.

Return:
A. validation table;
B. corrected client-language draft;
C. unresolved evidence gaps;
D. statements requiring advisor or compliance approval;
E. final status: APPROVE_FOR_INTERNAL_USE, HUMAN_REVIEW_REQUIRED, or BLOCK.

Recommended validation output schema

A prose-only reviewer is difficult to audit. Require structured output that a workflow can test.

{
  "statement_id": "S-04",
  "original_text": "The portfolio is protected if rates rise.",
  "classification": "UNSUPPORTED",
  "portfolio_scope": "ACCOUNT-123",
  "source_ids": ["RISK-2026Q2-08"],
  "as_of": "2026-06-30T21:00:00Z",
  "issues": [
    "absolute protection claim",
    "scenario assumptions not stated"
  ],
  "proposed_text": null,
  "approval_required": "ADVISOR",
  "status": "BLOCK"
}

The application should validate required fields before displaying an “approved” status. A model-generated JSON object is still untrusted input.

Human approval matrix for advisor-client communication

Output	Internal preparation	Client delivery	Required owner
Verified holdings and performance fact	Allowed	Allowed after automated reconciliation	Data owner
Reproduced attribution calculation	Allowed	Allowed after deterministic recalculation	Investment operations
Plain-language explanation	Allowed	Review required	Advisor/editor
Market interpretation	Allowed with citation	Review required	Investment professional
Client-specific implication	Review required	Review required	Advisor
Allocation or security recommendation	Blocked	Blocked until approved	Authorized advisor and applicable supervisory process
Trade instruction	Blocked	Never executed from commentary workflow	Authorized execution process

The matrix should be implemented in workflow logic, not left only inside the prompt. A prompt can label a statement; the surrounding system must enforce who may approve or send it.

Three tests for an AI portfolio commentary prompt

Test 1: The mismatched benchmark

Provide portfolio returns against two different benchmarks and omit the benchmark name from the draft. The reviewer should refuse to approve comparative language until the correct benchmark is identified.

Test 2: The stale position file

Provide current market commentary with portfolio holdings from the previous quarter. The reviewer should surface the timestamp mismatch and block current-exposure claims.

Test 3: The hidden recommendation

Insert a sentence such as “Given this environment, moving toward shorter-duration bonds would be prudent.” The system should classify it as a recommendation even though it avoids the words “buy” and “sell.”

Other useful cases include conflicting performance reports, missing fees, currency conversion errors, preliminary corporate actions, incomplete household data, and prompt-injection instructions embedded in an imported research document.

Metrics for portfolio commentary quality

Do not measure success only through time saved. Track:

numerical reconciliation failure rate;
percentage of material claims with valid source IDs;
unsupported causal-claim rate;
stale-data detection rate;
recommendation-detection recall;
reviewer edit distance;
blocked-output rate;
client complaints or corrections; and
time from data availability to approved commentary.

A low block rate is not automatically good. It may mean the controls are weak. Evaluate whether the workflow blocks the cases it was designed to stop.

How to explain portfolio performance to clients without overclaiming

Good client language is specific but not absolute. It connects the observed result to its evidence and preserves uncertainty.

Weak:

The portfolio performed well because our defensive positioning protected it from volatility.

Better:

During the quarter, the portfolio returned 3.4%, compared with 2.9% for its stated benchmark. Healthcare and utilities were the largest positive relative contributors in the approved attribution report. Those results describe the period; they do not predict how the same exposures will perform in future markets.

The second version defines the period and comparison, cites measurable drivers, and avoids claiming permanent protection.

The practical takeaway

AI portfolio commentary should be treated as a controlled publishing workflow, not a writing shortcut. The system needs verified data, reproducible calculations, statement classification, source-linked review, and an approval boundary that the model cannot override.

The strongest use case is not autonomous advice. It is faster preparation of accurate, understandable material while the advisor retains responsibility for interpretation, suitability, and client-specific action.

For a broader evidence and numerical-review framework, see AI Financial Analysis Controls. For prompt scoping and validation patterns, use Financial Prompt Engineering for Reliable Analysis.

Sources

This article is an educational control framework. It is not investment, legal, or compliance advice.

#ai-portfolio-commentary#financial-advisor-ai#client-communication#investment-commentary#prompt-engineering