Executive search
Reviewing AI search outputs: trace the claim before approving the draft
Check an AI-assisted search brief against its evidence. Use a filled claim ledger to catch wrong periods, unsupported attribution, double counting and unconfirmed instructions.
A preparation note says a candidate led a factory opening. The linked company report confirms the opening, but never names the candidate. The citation is real; the personal claim is unsupported.
For an executive-search consultant, that distinction matters more than how polished the note sounds. Review the statements that could change a question, a client account or the next action. Keep the original passage beside the proposed wording so another person can follow the decision.
This guide offers a claim-review method for working drafts. Candidate assessment and employment decisions remain with the appropriate people.
TL;DR
- Check what the source establishes about the exact person, organization and period.
- Separate facts, attributed accounts, inferences, calculations and proposed instructions.
- Record the correction beside the claim, then inspect the final wording.
- Evaluate the effort to capture, check and return the output as well as its prose quality.
A filled claim ledger
The following example, people, organizations and source records are entirely fictional. It demonstrates review decisions; it is not a customer case or a model-accuracy test.
A consultant is checking preparation brief v2 for internal use. The permitted evidence bundle contains:
- E1: Cedar Components' invented 2024 annual summary, stating that its Bristol facility opened that year. It does not identify the executive responsible.
- E2: Rowan Hale's invented career account, stating that Rowan was Cedar's finance director from 2022 to 2023. No current-role evidence is supplied.
- E3: An invented client-call note dated 28 October 2026: the client suggested consumer-sector experience; the partner still needs to confirm whether it is essential.
- E4: An invented assignment worksheet dated 2 November 2026. First interviews are scheduled for
P03andP05; prior relationship records exist forP01,P03andP05.
| Draft claim | Evidence and issue | Decision and replacement |
|---|---|---|
| Cedar opened its Bristol facility in 2024. | E1 supports the company event and year. | Retain, citing E1. |
| Rowan led the facility opening. | E1 does not name Rowan; E2 establishes an earlier role only. | Remove. Ask the consultant whether evidence of Rowan's involvement is needed. |
| Rowan is Cedar's finance director. | E2 covers 2022–2023, not the present. | Replace with “Rowan's supplied career account records a finance-director role at Cedar from 2022 to 2023; current role unverified.” |
| Five people have a prior relationship or a scheduled first interview. | Adding 3 and 2 counts P03 and P05 twice. | Replace with “Three distinct people appear in the union of the two selected sets.” |
| Consumer-sector experience is mandatory. | E3 records a suggestion awaiting confirmation. | Replace with “The client suggested consumer-sector experience; the partner must confirm whether it is essential.” |
| Two first interviews are scheduled. | E4 identifies P03 and P05 as of 2 November. | Retain with the worksheet date; check for changes before the meeting. |
The revised brief can state the supported company event, the dated career account and the current open question. It cannot connect Rowan to the facility opening simply because those facts appear near each other. The consultant also keeps the relationship count scoped to the selected worksheet rather than presenting it as the firm's entire network.
For your own ledger, add the reviewer, output version, intended audience and any unresolved decision. Copy the exact contested sentence rather than a label such as “check candidate background.” That makes the correction inspectable.
Open the source behind the citation
NIST's Generative AI Profile identifies confabulation risks, including misleading citations. Its suggested action MS-2.5-003 calls for checking sources and citations in generated outputs. This is voluntary cross-sector guidance, not a certification of a model or a Laisy workflow.
For each consequential claim, locate the supporting passage and compare its scope with the draft. Check the named entity, the relevant period and any qualification. If the source is inaccessible, record that limit and obtain suitable evidence or narrow the statement. A search snippet or a generated assurance does not complete the check.
OpenAI's explanation of ChatGPT's factual limitations warns that confident answers can include fabricated references and incorrect facts. That provider warning does not establish which model produced a particular draft. The evidence check is useful regardless of the model name.
Keep identity, time and interpretation separate
A name match needs enough context to identify the intended person. A group announcement may concern a subsidiary; an employer name may refer to a former role. In E2 above, even the correct person and company do not establish present employment.
Distinguish the date of a record from the date of the event it describes. A newly edited CRM note can contain an old account. The candidate research-note structure gives the reviewer a place to retain those dates and distinguish source-backed findings from attributed accounts.
Then inspect the reasoning between evidence and conclusion. A company event does not establish an individual's responsibility. A recorded meeting does not establish interest in a role or permission for an introduction. If an interpretation is useful, label it and state the missing evidence or alternative explanation. Otherwise remove it.
For shortlist feedback, preserve who made each observation. Several comments should not become an invented consensus or an automatic suitability score. A missing answer is an unresolved question, not evidence of a person's motivation or integrity.
Check the captured material before polishing the summary
An output can reproduce an extraction error faithfully. Compare important names, dates, units and table values with the selected original, especially where text came from a scan or a complicated layout.
Microsoft's Azure AI Search transparency note gives a product-specific example: OCR errors can carry into later key-phrase extraction. That illustrates a possible evidence-path failure; it does not describe Laisy's implementation.
When the original says “2023 revenue” and the captured row is labelled “2024 revenue,” editing the final sentence cannot fix the underlying mismatch. Correct the captured value, regenerate or amend the affected output, and check it again. If the source cannot be read reliably, leave the field unresolved.
Recalculate the number for the question being asked
Write down the unit, population and period before calculating. In E4, the question asks for distinct people in either set. The union is P01, P03, P05: three people. Two people belong to both sets, so adding the set sizes gives the wrong answer.
Other reviews may need a different denominator. “Claims corrected” counts claims, not candidates. A selected-record sample does not describe the whole database unless the sampling design supports that inference. A correct percentage cannot repair an unsuitable denominator.
Retain the difference between a source's reported figure, an independently checked calculation and an invented teaching example. The six rows above were chosen to show different decisions. Their proportions say nothing about any model's error rate.
Review the final version and its next use
After corrections, inspect the actual brief that will be used. Check that a later edit has not removed the date, restored an unsupported inference or turned a suggestion into an instruction.
The AESC Professional Practice Standards treat AI as support for work while retaining professional judgment. These are member standards, not evidence that a particular tool complies with them.
Record what was reviewed and what remains open. For client-facing work, the client-update workflow helps separate the factual account from the decision request and intended recipients. An accepted draft does not itself authorize outreach, a CRM change or a candidate decision.
Compare the whole working task
If your existing Claude workflow already produces an adequate draft, use that as the baseline for an evaluation. A new tool needs to justify its role in the actual task. Record time spent selecting and capturing sources, checking statements, making corrections and returning the accepted output to the existing record. Keep setup and continuing administration visible alongside software cost and the privacy arrangements for the selected inputs.
The pilot-success framework can turn that into a bounded comparison: one output, the same permitted source set, clear acceptance criteria and a named reviewer. This article does not establish that Laisy is a better fit or produces better prose. A useful result may be to retain the existing workflow and fix one troublesome capture or handback step.
Frequently asked questions
Does a source link mean the statement is verified?
No. Locate the passage and check its entity, period and scope. E1 supports a facility opening, but not Rowan's responsibility for it.
Must every sentence get a ledger row?
Use rows for claims that could affect understanding or action, plus disputed statements that need a traceable correction. Review the whole artifact for consistency without turning routine copy edits into a separate administrative exercise.
Can another AI model replace source checking?
It can suggest issues to investigate. Agreement between generated accounts does not independently establish the source fact. The reviewer still needs evidence for the disposition.
What if one claim remains unresolved?
Remove it, qualify it or present it as an explicit question, according to the purpose of the output. If the intended decision depends on that claim, hold the affected decision until the evidence is sufficient.