Updated
Unstructured data is information that doesn't arrive organized in a fixed field structure for analysis. Free comments, recordings, images, and documents can contain details that a sales column doesn't explain. The challenge is to transform them into useful evidence, preserving context and recognizing what the sample doesn't allow us to conclude.
A spreadsheet can show how many deals were lost. The salesperson's notes can suggest why. Reading these notes helps formulate hypotheses about the offer, communication, and service; it doesn't, on its own, prove the cause of a loss.
Structured, semi-structured and unstructured data
The distinction presented by IBM stems from the organization of the data: a predefined schema facilitates consistent queries, while free-form content requires other interpretation processes. A file can contain different types of information. The file extension format, in isolation, does not resolve the classification.
| Type | Example | Care |
|---|---|---|
| Structured | Date, order code, and value in defined fields. | Organized fields can also contain errors or gaps. |
| Semi-structured | JSON record with variable keys and structures. | The presence of keys does not guarantee standardization across origins. |
| unstructured | Free text from an interview or audio recording of a customer service interaction. | It is necessary to interpret the language, context, and quality of the record. |
A CRM can store value and stage in structured fields and, alongside them, a free-form description of the conversation. A sensor measurement with time and temperature doesn't become unstructured just because it comes from a machine. Sort by content and the organization used in the analysis.
What separates a piece of data from an insight?
In this script, we call an insight a useful interpretation to guide a decision, supported by observations and open to verification. “We received messages about the deadline” is a record. “The promised deadline is not clear to replacement buyers” is an interpretation that needs to be confronted with the material and other explanations.
One possible course of action could be to display the timeframe by region before the quote request. The test should verify if this reduces doubts or improves the suitability of contacts. It's not enough to ask an AI to produce "insights" and treat the response as a proven discovery.
Example: from sales observations to a decision
Imagine a fictional sample of 40 lost deals over a defined period. The team read the main reason recorded for each one and created the ranking below. The numbers are for illustrative purposes only and do not represent a survey or a client of Dois Z.
| Classified record | Quantity | Next check |
|---|---|---|
| Incompatible deadline | 12 | Compare the requested, promised, and available timeframes. |
| Inadequate specification | 9 | Check what the advertisement and technical material explained. |
| Price quoted | 7 | Understanding alternatives, conditions, and perceived value. |
| Another reason identified | 5 | Review the cases before creating new categories. |
| Insufficient record | 7 | Improve filling without inventing the cause. |
The 12 instances of payment deadlines represent 30% of these 40 losses. This does not mean that 30% of all clients in the market reject payment deadlines. Only losses were observed, and seven records do not even allow for classification. Interviewing some buyers and comparing won deals could change the interpretation.
An analysis workflow that allows you to check the result.
- Define the decision. Choosing a message change is a more verifiable goal than "understanding everything about the customer".
- Define the sample size. Record the origin, period, inclusion, exclusion, and quantity. Do not present a partial export as the complete universe.
- Prepare the materials. Remove duplicate analysis, identify unreadable sections, and preserve a reference to the original record with controlled access.
- Write the categories. Explain what goes into each class and allow for "other" and "insufficient" options. Define whether a record can have multiple labels.
- Review a sample. Compare classifications and discuss discrepancies before automating the rest.
- Test an action. Document the hypothesis, the person responsible, the indicator, and the review deadline.
Where AI, transcription, and text analysis help.
IBM describes natural language processing and machine learning among the techniques used to analyze unstructured content. In practice, a tool can transcribe audio, suggest categories, or locate passages related to a question. The choice depends on the material, language, quality, and volume.
Ask each person to point out the passage that supports their classification. Separate extraction from interpretation: first what was said, then the hypothesis about its meaning. Review negations, irony, product and unit names, as a plausible summary may miss crucial details. A confidence index for the model does not replace verification.
Before sending data to a service, evaluate access, purpose, retention, and the need for identification. Start with minimized material authorized for that use. Do not automatically include complete conversations or personal data in a tool trial.
How to bring analytics to marketing and CRM.
Connect revised categories to the fields needed for the decision: product, stage, period, and origin when known. This allows the team to compare objections by offer and discuss what needs to change in the ad, the page, the technical material, or the customer service.
A recurring question about application can guide content; a recurring loss due to deadlines may require an operational decision. Not every problem encountered in conversations will be solved through communication. The acquisition, service, and CRM guide shows how to organize this passage through the funnel.
The useful result is not just a word cloud or an elegant summary. It is a decision with locatable evidence, explicit limits, and a follow-up test that allows one to learn whether the interpretation was correct.