Define the decision before uploading a file
Useful analysis begins with a question that can change a concrete decision, such as reviewing a campaign, organising capacity or investigating a sales decline. Explain the context and available alternatives. That makes it possible to select necessary data and distinguish an interesting observation from information that helps choose the next step for the business.
Asking AI to analyse everything can produce a long report with few consequences. I prefer questions such as: where are unanswered requests accumulating? Which offers require more delivery work? How do contacts from two sources differ after they have had enough time to move through the sales process?
The question also reveals missing information. If you want to assess sales but only have click data, the model cannot reconstruct an unrecorded journey. It can help organise hypotheses and identify gaps. That contribution is useful when it remains clear that there is not yet enough evidence for a conclusion.
In Sextas Ímpares #142, about advertising analysis with Claude, I work with campaign data and discuss ways of obtaining information for analysis. The Portuguese-language class provides practical context. Here, I extend the reasoning to preparing decisions that connect marketing, sales and operations across different sources of information.
Confirm the meaning and limitations of the data
Before analysing, identify the source, period, unit and meaning of every field. Check for total rows, duplicates and missing information. AI should receive those definitions alongside the file, because similar names can represent different events and lead to comparisons that do not correspond to how work is actually performed or recorded.
In the class #142 passage on data exports, I showed that a file may contain only visible columns. That detail is easy to forget. If the view omitted an important metric, the analysis starts with a limitation that must be identified before interpreting the result or comparing campaigns.
In a sales file, ask whether a date means enquiry arrival, proposal submission or closing. In a revenue report, confirm how cancellations and refunds are treated. In a work table, distinguish planned time from actual time. These definitions change comparisons even when the values themselves have been added correctly.
Keep a small field dictionary and the extraction date. Anyone repeating the analysis should know whether they are using the same definition. The sales automation with AI page shows why consistent states help explain an enquiry’s journey and avoid reports that combine different stages under the same label.
Choose comparisons that respect the business journey
A comparison should bring together equivalent periods, groups and definitions while considering how long results take to appear. If conditions changed, identify the change before assigning a cause. AI can organise those differences, but needs context about campaigns, offers, the team and events that are not described anywhere in the file itself.
Imagine comparing two months of enquiries. One included a new offer and a period when part of the team was absent. A difference in sales can have several explanations. Before concluding that the campaign deteriorated, check enquiry volume, response time, proposals sent and the usual length of the decision process.
Avoid comparing newly arrived contacts with groups that have already had weeks of follow-up. The first group has not completed its journey. You can monitor early indicators, but state what is being compared and what remains unobserved. Analysis then follows reality instead of anticipating outcomes that have not yet had time to develop.
The article on growing a business with attention to margins, teams and AI connects numbers to operations. A change in capacity can explain delays and influence future sales. That context belongs in the discussion even when the spreadsheet only contains dates and values without explaining what happened around them.
Check calculations and separate observations from hypotheses
Important calculations should be checked with a suitable tool and explicit definitions, without relying solely on the model’s written response. Distinguish what the data shows from the explanation proposed for that result. A hypothesis can guide investigation, but needs further evidence before it justifies a significant change in the way the business operates.
For example, if one hundred enquiries produce ten sales, the rate for that group is ten per cent. You still need to know whether every enquiry had time to develop and whether sale has the same definition throughout. The calculation can be correct while interpretation remains incomplete because of the group selected.
An average can also hide different situations. A reasonable average response time may combine many promptly handled enquiries with a few forgotten for days. Ask for the distribution, look at examples and check what happens in extreme cases. That gives the team a concrete issue to investigate, alongside a number to monitor.
I want reports to identify when a statement is a possibility. If sales fell while response times increased, there is a relationship to investigate. We do not automatically know how much of the decline came from that delay. Conversations, history and other changes help test the explanation before treating it as a definitive conclusion.
Turn the report into work someone follows through
A report should state the period analysed, relevant observations, uncertainties and actions someone will check. Assign an owner and a time to review each decision. Information becomes useful when it reaches the team’s work and allows comparison between what was expected and what actually happened after an agreed change was put into practice.
In class #142, during the discussion of automation and reports, I explained a sequence using spreadsheet data and information sent through automation. The delivery channel is an operational choice. The content needs enough clarity to support decisions without forcing the recipient to reconstruct the entire context first.
An internal alert might identify unassigned enquiries and link to the authorised place containing details. A campaign summary can highlight a change and request review of a hypothesis. Avoid putting personal contacts or complete files into a message when aggregated information and a link can serve the purpose adequately.
The business process automation page develops the transition between analysis and execution. The article on turning leads into customers adds the sales perspective. Good analysis helps the team choose where to look and record what it learned from the investigation, including findings that challenge the initial explanation.
Build an analysis routine the team can maintain
An analysis routine needs known sources, stable definitions and review when the business changes. Define who prepares data, who verifies results and who follows actions. The process should remain understandable to another person, with examples and documentation that allow repetition without depending on the usual owner’s memory or availability for every question.
In my 2022 book, I advocated documenting processes and lessons to support continuity. Applied to AI analysis, that includes questions used, filters, known limitations and decisions taken. If a hypothesis failed, keep the result. Avoid having the team spend time reaching the same conclusion again without considering what has already been observed.
Choose information appropriate to the environment being used. Prepared or aggregated data may be sufficient for an initial assessment. The AI application security page explains access and handling considerations. The model should receive what the task needs, with authorisation and without credentials or unnecessary personal information that adds nothing to the analysis.
To discuss implementation, present the need to SpartAds. To learn how to prepare analysis and review decisions in your own project, explore AI and business mentoring. Bring a concrete question and a description of available data: they are the best starting point for understanding what could be improved.