Extract outcomes from a grantee report by identifying the result claimed, the people and period it describes, and the source evidence behind it. Keep activities, outputs, observed outcomes and causal claims separate. Preserve the original passage and record missing definitions or conflicting values. The output is a review table, not a set of numbers that AI has filled in to make the report complete.
This lesson is for grant and portfolio teams reviewing narrative reports, tables and supporting files. You will prepare an extraction table and a focused clarification request. It develops the field-level extraction skill used inside a broader report review.
- Confirm the grant, reporting period and available source versions.
- Identify each relevant claim without changing its meaning.
- Separate inputs, activities, outputs and outcomes.
- Record the measure, population, timing and evidence status.
- Resolve contradictions or state the missing information.
- Approve only the findings the evidence supports.
What are you looking for in the report?
Start with the agreed reporting question. A report can contain important evidence that is not part of a mandatory metric, but do not require every possible indicator or monetary proxy from every grantee. Establish what applies to this grant and period before labeling a field missing.
Keep a source register with document name, version, date, relevant page or table and access scope. If a file cannot be opened or read reliably, record an extraction limitation. An inaccessible attachment is not proof that the grantee failed to collect the evidence.
How do you separate outputs from outcomes?
An output describes delivery or reach. An outcome describes a change or condition for the relevant people or setting. The report may state an outcome without enough evidence to assess it. Preserve the statement and its status rather than promoting it to an established finding.
| Fictional report passage | Category | What to record |
|---|---|---|
| “We delivered 12 workshops.” | Activity or delivery output | Workshop count and period; check the source definition. |
| “There were 240 attendances.” | Participation output | Attendance occasions, not automatically 240 unique people. |
| “36 participants were employed at follow-up.” | Reported outcome status | Count; ask which checkpoint, eligible population and known-status base. |
| “The program transformed employment prospects.” | Broad interpretation | Retain the claim; identify what evidence would support its scope. |
Do not treat the categories as a grading scale where every output is inadequate. Delivery evidence can answer a legitimate question. It simply cannot substitute for an outcome the report has not measured.
What fields belong in the extraction table?
For each relevant finding, record the source passage, claim type, measure, value, unit, population, period, denominator when needed, source location and review status. Add the applicable agreement or dictionary definition where available.
Keep the value separate from the reviewer’s interpretation. “36 employed” is a count. “60% employed” requires a denominator of 60 from the same eligible group and period. Do not derive that rate from an unrelated table or assume that every nonrespondent was unemployed.
Record when a source describes a checkpoint rather than continuous duration. Employment on a date does not establish uninterrupted employment through the preceding months. A clearer table should preserve that distinction, not smooth it away.
How should missing evidence be handled?
Use specific states: not reported, definition unclear, source inaccessible, not applicable, not yet due or conflicting. “Missing” alone may lead to the wrong follow-up. A future outcome not yet due should not be treated as a late submission.
In the fictional example, suppose the report gives 36 employed but no follow-up base. Ask for the eligible and known-status counts and the checkpoint. If the answer is 60 known statuses among 80 eligible people due for follow-up, the observed employment rate is 36/60 = 60%, with 60/80 = 75% coverage. Twenty statuses remain unknown.
If the denominator cannot be recovered, retain the count with its limitations. Adding a new collection field can improve the next cycle, but it does not repair the missing historical observation.
What if two sources disagree?
Compare scope and versions before treating the difference as an error. A narrative may report a cumulative total while a table reports the quarter. Two values may both be correct for different periods. Preserve each and request a specific clarification if they should describe the same thing.
Keep the reviewer’s resolution, source and date. Do not automatically choose the larger number, the latest uploaded file or the most detailed paragraph. Use the agreed source and review policy, and retain uncertainty where it cannot be resolved.
Should in-kind contributions be converted to money?
Not for every report. First record the contribution as described, with its unit and source. A volunteered hour count and an assigned monetary value are separate items. Where valuation is actually required, use a separately reviewed method and its assumptions rather than inventing a price during extraction.
Similarly, do not attach a financial proxy to an outcome merely because a report mentions it. Outcome extraction establishes what evidence exists; valuation is a later specialist task with additional requirements.
How does Sopact support extraction?
You can build the table manually from a report and its attachments. Sopact’s connected approach keeps recurring submissions and files with the relevant grant or partner history, and supplies definitions as context for analysis. AI can propose extracted passages and identify possible gaps as material arrives.
Test the workflow with scanned tables, footnotes, changed headings and contradictory statements. Verify that the source passage actually supports the extracted value. Review access before sharing findings; a report review does not imply that its underlying documents should be available through an unrestricted link.
Practice: approve one finding and hold another
Use the fictional workshop and employment passages. Build two extraction rows. Approve the output only to the extent its definition and source support it, and write a focused request for the missing employment base. Then record what can still be reported if that base remains unavailable.
Frequently asked questions
Can AI fill a missing outcome from the narrative?
It can extract a supported statement but should not invent an unreported value. Mark what is missing and identify a source or future collection step. A plausible estimate is not an observed outcome.
Does a source-linked result prove impact?
No. Traceability helps a reviewer inspect the evidence. The strength of the claim still depends on the measure, coverage and evaluation design. Keep observed status or change separate from causal attribution.
Should every grantee report the same fields?
Use the agreed common core where comparison is needed, with relevant grant-specific fields. Check applicability before following up. Map compatible local measures explicitly and retain differences that cannot be resolved.
When is extraction complete?
When relevant claims have been recorded and reviewed, with gaps and conflicts visible. That does not mean every requested finding is available. A complete review can conclude that a particular rate or claim cannot yet be reported.