Define the question, check which records are included and verify the answer against its sources. A practical method for portfolio teams using AI.
By Sopact · Updated September 12, 2026
Ask useful AI questions about portfolio data by specifying the organizations, period, definitions and evidence the answer should use. Require the answer to show its sources, calculation and missing coverage. Verify important findings against the underlying records before sharing them. A fluent answer is not enough: a reviewer should be able to explain how the answer was reached and what it leaves unresolved.
This lesson is for portfolio and impact teams working with investee or grantee reports, documents and check-ins. Bring the reviewed evidence and open monitoring questions from the previous chapter. You will build a question brief, a review prompt and a short answer-checking record. You can practice with authorized files before considering a connected workflow.
“How is the portfolio doing?” leaves too many choices to the assistant. Name the commitment, period, population and intended use. Ask for a reviewable output rather than an unexplained overall rating.
| Broad question | More useful version |
|---|---|
| Which partners are struggling? | For this quarter, which reviewed measures are below their agreed targets, and what unresolved evidence prevents assessment? |
| Why did outcomes improve? | For the same eligible cohort and observation window, what changed, and which partner explanations are supported or still hypothetical? |
| What should we tell the board? | Summarize confirmed results, material changes and unresolved questions for the reporting period, with source references and coverage. |
| Who has the best results? | Which organizations have comparable measures and periods, and what can be compared without hiding differences in population or context? |
Define the output format too. A table of finding, source, calculation, limitation and next action makes the answer easier to inspect than a polished narrative alone. Keep recommendations separate from observed facts.
Before asking about the whole portfolio, establish which organizations, periods and sources are available. A search may return relevant examples without retrieving every record needed for a total. Absence from the answer does not establish absence from the portfolio.
For an aggregation, require an explicit included and excluded record count. For a qualitative question, distinguish the number of comments reviewed from the number of people or organizations represented. Ten comments from one organization do not necessarily represent ten independent perspectives.
A reusable brief keeps the assumptions stable when a different colleague asks the question or the next reporting cycle arrives. Store the brief with the result and the source references.
| Brief field | What to specify |
|---|---|
| Decision | The review or action the answer will support |
| Scope | Organizations, population, reporting period and cutoff date |
| Definitions | Measure, unit, denominator and dictionary version |
| Evidence | Accepted sources, reviewed values and unresolved questions |
| Calculation | Aggregation or comparison rules and exclusions |
| Output | Findings, source locations, uncertainty and next actions |
| Review | Who checks the answer before it is used |
For a question about change over time, require the same observation window and comparable definitions. If a definition changed, ask the assistant to identify the break rather than produce a smooth trend from incompatible numbers.
Fictional exercise, continuing the course. Partner A has 35 confirmed starts against a target of 40. Partner B has 20 confirmed starts against a target of 20, using the same definition and quarter. Partner C’s expected submission is unresolved.
A defensible answer to “How many starts did the portfolio achieve?” is: “The two reporting partners have 55 confirmed starts. Partner C is missing, so 55 is not a complete portfolio total.” The comparable target for those two partners is 60, making their combined shortfall five starts. Do not compare their actuals with a target that includes Partner C without clearly explaining the mismatch.
Organization-level coverage is two of three, or about 66.7%. That is not necessarily coverage of people served, invested capital or the portfolio’s outcome commitments. Use the denominator relevant to the question.
If the assistant adds Partner A’s five future offers and reports 60 starts, it has changed the definition. Correct the answer and retain the reason: the five offers belong to a later observation, not current-quarter starts. If it says outreach caused the result, ask what evidence supports that causal statement.
Replace the bracketed fields and supply only records you are authorized to use. The prompt requests a review, not an automatic decision.
Answer [question] for [organizations and period], using [dictionary version] and [listed sources]. First list the scope you can actually access and the expected sources that are missing.
Return a table with finding, value or relevant quotation, source location, calculation, limitation and next action. For totals, show included and excluded records and the denominator. Keep observations, projections and partner explanations separate.
Do not treat missing values as zero, silently combine incompatible measures or infer causation. If sources conflict, show both and identify the clarification needed. If the evidence cannot answer the question, say what is missing. End with the items a person should verify before using the answer.
Test the answer with a small known example first. A longer prompt cannot compensate for inaccessible evidence, inconsistent definitions or an incorrect relationship between records.
Open the cited source for important findings. Confirm that it supports the actual claim, not merely the general topic. Recalculate consequential totals or rates, and check that exclusions have not disappeared from the narrative.
| Check | Question for the reviewer |
|---|---|
| Source | Can I reopen the exact record, page or row? |
| Scope | Does this evidence cover the named organization, population and time? |
| Arithmetic | Do the included values and denominator reproduce the result? |
| Interpretation | Is the conclusion observed, inferred or proposed? |
| Coverage | Which expected records are absent or excluded? |
| Change history | Did a correction or definition revision affect the answer? |
A valid-looking citation is not a guarantee. An assistant can cite a real passage and still overstate what it establishes. For a board or funder report, retain the checked calculation and review record alongside the final wording.
The Model Context Protocol documentation describes MCP as a standard for connecting AI applications with external systems, data and tools. A connection can make information available to an assistant; it does not, by itself, make that information complete, comparable or correctly interpreted.
Before relying on a connection, confirm the supported assistant, available operations, access permissions, source freshness and retrieval limits. A connector that can search documents may not provide an exhaustive dataset for portfolio totals. Test the exact question and operation you need.
Keep reading and analysis separate from actions that change records or notify people. Establish who authorizes those actions. Do not assume that connecting an assistant grants permission to expose every partner document to every user.
A useful longitudinal question names the observation and comparison periods and retains the versions used. “What changed since the last review?” should show new evidence, corrected values and unresolved questions that were already known.
For Partner A, distinguish a new job start from a correction to an earlier count. If the five future offers later become starts, record them in the appropriate period. If a prior record was wrong, preserve the restatement. Otherwise, the next trend can confuse operational progress with data cleanup.
Save the question brief and reviewed answer at each checkpoint. That gives the team a way to understand why two answers differ rather than treating the newest response as a replacement for the entire history.
Watch the portfolio reporting demonstration · 2 minutes 16 seconds
A Sopact demonstration of connected portfolio records and reporting. This demonstration introduces connected portfolio reporting. The demonstration introduces connected portfolio reporting. Use the three-partner exercise to test source coverage and the answer in your own setup.
Start with the three-partner exercise. Test whether responses, documents and reviewed values stay linked to the correct partner and period. Apply the shared dictionary, ask for the comparison and inspect whether the answer exposes the missing partner and excludes future offers.
Then test a follow-up question using a corrected record. Confirm which assistant and connection arrangements are supported for your setup; do not assume a particular integration or permission model. The practical benefit to assess is whether your team can revisit a question with its evidence and definitions intact, while still reviewing the answer.
Only within the records it can access and the operations it supports. Verify coverage before using phrases such as “whole portfolio,” especially for totals.
No. Check that the source supports the specific claim, the scope is correct and the arithmetic can be reproduced.
Yes, using authorized data and clear identifiers, periods and definitions. Record the snapshot date and refresh it deliberately; an upload does not stay current automatically.
No. It provides a connection standard. Data quality, access, retrieval coverage, calculation and interpretation still require checks.
Identify the gap and limit the answer. It should not insert zeroes, invent explanations or present an incomplete total as complete.
Reuse the question brief, check definition and population changes, and preserve source versions and corrections. Explain differences in evidence as well as differences in results.
Next, use the portfolio roll-up lesson to combine only the results that are genuinely comparable. Bring the included and excluded record list from this exercise.
Bring one portfolio question, its definitions and sample evidence. Work through the collection and analysis workflow with Sopact.
Explore Impact & ESG Portfolio →
