Read the response against the question you asked
Start by checking whether the response solves the requested task and uses the information you supplied. A text can be broadly correct yet unsuitable for your situation. Identify the expected result, the conditions limiting the decision and the elements that were left out before you spend time editing its style.
For me, a good response must serve the situation in which it will be used. Well organised paragraphs and a confident tone help with reading, but they do not replace that fit. If I requested an analysis of a specific offer, a list of sales best practices leaves part of the work undone.
Imagine a hypothetical team trying to understand why contact requests increased while sales remained stable. The AI recommends increasing spending because contacts have grown. That recommendation skips the main question and assumes that more contacts mean better opportunities, without demonstrating the relationship.
Before correcting sentences, return to the request and mark where the response meets it. Check whether it distinguishes observations, possible explanations and proposed actions. When these are mixed together, a hypothesis can pass for a fact and a suggestion can seem like the only reasonable decision.
The work of giving AI context with clear criteria makes this check easier. Even with a good input, read the result looking for ignored constraints. A limited budget, an agreed deadline or an unavailable team can make a seemingly sensible recommendation impractical.
To prepare your next step, explore AI data analysis for businesses.
Check the claims that support the conclusion
Identify the claims that would change the decision if they were false, and look for support for each one. Open the relevant documents or records and check their content, date and scope. A reference supplied by AI only helps when it exists and actually says what the response attributes to it.
Not every sentence in a long response deserves the same effort. I would start with numbers, promises about capabilities and statements that justify taking action. If the recommendation depends on a commercial condition, that condition needs checking against material valid for the situation, rather than a generic recollection from the model.
You can ask AI to locate the passage supporting a claim in the documents supplied. Then open that passage and read what comes before and after it. A sentence may include a qualification that disappeared in the summary or refer to a specific group while the response presents it as a general rule.
Do not count two responses from different models as two independent sources. Agreement can help uncover points to investigate, but it does not by itself demonstrate that the information is correct. Verification needs to reach the document, record or person responsible for the information.
When you cannot find sufficient support, leave the claim unconfirmed and adjust the conclusion. “Could not verify” is a useful working status. Avoid filling the gap with a plausible sentence merely to make the report look finished; that choice hides precisely the point that needs attention.
Redo the calculations and check what the numbers mean
Check the original values, period, unit and operation before accepting a calculation. Then verify whether the indicator answers the question. A division can be correct while the conclusion remains wrong if it compares different populations, mixes periods or treats contacts as customers actually acquired by the company through the activity.
In a hypothetical example, a campaign produces 200 contacts and 20 confirmed sales. The proportion is 10%, provided those sales belong to the same group of contacts and the period under consideration. If some sales came from older contacts, the division is still simple, but it no longer measures that conversion.
Redo the calculation in a calculator or spreadsheet using the checked values. Keep the operation visible so another person can reproduce it. Asking AI to confirm its own response may uncover an error, but it does not replace checking the input and the meaning of the indicator.
This care begins with preparing advertising data for AI analysis. During the final review, check column definitions and possible duplicates again. A total of submitted forms may include several submissions by the same person; calling it the number of new customers changes the meaning of the data.
Consider the precision of the conclusion too. A change over a short period may justify investigation without demonstrating a lasting trend. If information about contact quality or sales follow-up is missing, keep that absence visible instead of attributing the entire result to the campaign.
Test the step from facts to recommendation
After checking the facts, ask whether they support the proposed action and what alternative explanations remain possible. Distinguish what the data shows from the interpretation placed on it. A useful recommendation should identify its conditions and the information still missing, without presenting a hypothesis as an inevitable conclusion for the team.
In the contacts and sales example, several questions may remain unresolved. The source of enquiries, the sales response time or the type of offer may have changed. Without investigating those possibilities, increasing spending could amplify a difficulty the analysis has not yet understood.
Ask for an alternative to the main recommendation and the reason for rejecting or testing it. The purpose is to expose assumptions, not to produce an endless list of scenarios. If every option depends on the same unconfirmed piece of information, that information deserves priority before debating which plan sounds more convincing.
Distinguish decision-making ability from writing ability too. When choosing a model for a business task, I assess the result I need to obtain. Good writing can help communicate an already validated decision, but it does not automatically give the model the context required to make it.
Close this part with an action proportionate to the available evidence. That may mean correcting a report, collecting missing information or testing a limited change. The review should make clear why that action is reasonable now and what would make the team reconsider it.
Keep the correction so you do not review the same error repeatedly
When you find an error, record what was wrong, the correction and the information supporting it. Update the instructions or data that contributed to the failure where appropriate. Keep the reviewed version and outstanding questions so the next person knows what they can use with confidence in their own work.
If AI confused contacts with customers because the file contained an ambiguous column, improve that field's name and definition. If it added an unsupported conclusion, strengthen the instruction to separate facts from hypotheses. The correction becomes more valuable when it helps reduce work on the next use.
It is also useful to distinguish errors in the supplied material from errors introduced in the response. An incorrect number at source requires a different correction from a correct number interpreted improperly. Without that distinction, you may change the request repeatedly while continuing to supply the same defective information.
The depth of review should match the consequences of use. An internal draft may tolerate later editing; a statement sent to a client creates a different commitment. Define who checks the relevant points and avoid confusing a response reviewed for language with one validated for that destination.
Finally, keep a version that shows what was confirmed, corrected and left open. The team does not need an appearance of certainty to proceed. It needs to know which parts of the response are solid enough to support the next step and which parts still require work.