From Data to Decisions: How Modern Analytics and AI Solve Real Business Problems
Every business makes decisions that affect its customers, operations and bottom line. An e-commerce company has to decide how much stock to keep. A logistics company has to determine how deliveries should be routed. A financial institution has to decide which transactions deserve closer attention.
These decisions may look very different, but they share something important: they increasingly depend on data.
Having large amounts of data, however, does not automatically make a business more data-driven. The difficult part is turning raw information into something useful enough to support a decision. That requires asking the right question, selecting relevant data, identifying meaningful patterns and understanding what those patterns mean in the real world.
This is where modern analytics and artificial intelligence are changing the workflow. Descriptive and diagnostic analytics can help organizations understand what has happened and why. Predictive methods can estimate what may happen next, while optimization techniques can help identify possible actions. More recently, generative AI has started to become an additional layer in this process, assisting with tasks such as data exploration, code generation, visualization and communicating analytical findings.
But the technology is only part of the story. An analytical model can produce a prediction without understanding the business context behind it. An AI assistant can generate an analysis without necessarily knowing whether the underlying assumptions are appropriate. The value comes from combining technology with good data, domain knowledge and human judgment.
This article explores that process through three practical business problems: demand forecasting in e-commerce, route optimization in logistics, and anomaly detection in financial services. Across all three, a common pattern emerges:
The goal is to understand not simply what modern AI can do, but where it fits into the larger journey from data to action.
From Business Problems to Decision Frameworks
A useful analytics project does not begin with a model. It begins with a business problem.
Consider a retailer trying to decide how much inventory to keep. The question is not simply “Can we predict sales?” The more useful question is “How can we estimate future demand well enough to make better inventory decisions?”
That distinction matters because the same dataset can be analyzed in many different ways. The purpose of the analysis determines what data matters, what methods should be considered and, ultimately, what constitutes a useful result.
A simple decision framework is:
The type of analytical question also changes as we move through this process.
In practice, these approaches are often connected rather than isolated. A company may first examine historical performance, investigate the reasons behind changes, predict future conditions and then evaluate possible actions.
AI can contribute at several points in this workflow. An analyst can use an AI assistant to explore a dataset, generate or explain code, suggest visualizations, summarize findings or investigate possible anomalies. Current tools such as ChatGPT and Copilot explicitly support several of these data-analysis tasks.
There is, however, an important line between assistance and decision-making. AI-generated output still needs to be checked against the underlying data and the business context. Microsoft, for example, notes that Copilot outputs can be nondeterministic and provides guidance for users to evaluate and validate them.
This distinction becomes clearer when we look at real business problems.
Amazon and Demand Forecasting
Imagine a retailer preparing for a major shopping period. Ordering too little can mean products are unavailable when customers want them. Ordering too much can leave the business holding excess inventory and tying up capital. The difficult part is that the company has to make these decisions before it knows exactly what customers will buy.
This is the problem that demand forecasting tries to address.
Demand forecasting uses historical demand and other relevant information to estimate what customers are likely to purchase in the future. In a simple setting, this might involve looking at previous sales and identifying trends or seasonal patterns. At a large e-commerce business, however, the problem becomes considerably more complex because forecasts may need to be produced across huge numbers of products, locations and time periods.
Amazon has publicly described the evolution of its forecasting systems. According to Amazon Science, its earlier forecasting systems relied on traditional time-series methods. As the forecasting problem became more complex, Amazon researchers incorporated machine-learning approaches and additional information such as demand, sales, product category and page views. Amazon has also described using models including Random Forest, XGBoost and Ridge Regression for demand prediction in its devices business.
From historical data to a forecast
The basic workflow can be thought of as:
The data is particularly important. Historical sales provide the starting point, but demand can be influenced by many other factors. Depending on the business problem, useful variables may include product characteristics, prices, promotions, inventory levels, seasonality, holidays and external market conditions. AWS guidance on demand forecasting also highlights the importance of combining historical demand with relevant internal and external variables.
The output does not necessarily have to be a single number. Amazon Science has described the importance of probabilistic forecasting, where the model estimates a range or distribution of possible future demand. This is particularly useful for inventory decisions because managers need to consider uncertainty rather than assume that one predicted number will definitely occur.
For example, suppose a model estimates that a product will have demand of around 1,000 units next month. A decision-maker may also want to know how likely demand is to be significantly higher or lower than that estimate. That additional information can influence how much inventory the business decides to hold.
The decision is more important than the prediction
This is where an important distinction appears.
Forecasting is not the final objective.
A forecast becomes valuable when it helps someone make a better operational decision.
Amazon has described using demand predictions to help determine questions such as how many units to manufacture, how much to ship, where to position products and when to move them.
This gives us a more complete picture:
The same principle applies beyond Amazon. AWS describes demand forecasting as part of a broader supply-chain process involving inventory management, procurement, production, warehousing, transportation and distribution.
Where AI changes the workflow
Modern AI can expand this process further. Machine-learning models can identify relationships in large datasets that would be difficult to capture through simple historical averages. Generative AI can also assist the people working with these systems by helping them explore data, write code, explain model outputs or work with less structured information.
However, better technology does not remove the uncertainty in forecasting. A model can learn from historical patterns, but the future can contain events that were not present in the trainingdata. Changes in prices, promotions, consumer behaviour, supply constraints or unexpected external events can all affect actual demand.
That is why forecasting should be treated as decision support rather than a crystal ball.
Interestingly, AWS’s more recent supply-chain guidance makes a similar point from a business perspective: improving forecast accuracy alone does not automatically guarantee better business outcomes. Demand planning also requires collaboration and business expertise to interpret the forecast and incorporate information that may not be captured by the model.
This brings us back to our framework:
Data provides the evidence. Analytics produces the forecast. Human judgment turns the forecast into a business decision.
The Amazon example therefore demonstrates more than the application of machine learning to sales data. It shows how analytics becomes valuable when a prediction is connected to a real operational decision.
And forecasting is only one side of the problem. Once products need to move through the supply chain, another question emerges:
What is the most efficient way to get them from one place to another?
That leads us to the next case study: route optimization in logistics.
Route Optimization in Logistics
Once a business has decided what needs to be delivered, another problem begins: how should those deliveries be made efficiently?
For a logistics company, this is not simply a question of finding the shortest path between two locations. A delivery driver may have dozens of stops in a single day, each with different delivery or pickup requirements. Traffic conditions can change, customers may have specific service windows, and the vehicle itself has practical constraints.
This makes route planning a decision problem involving many variables at the same time.
A useful real-world example is UPS’s On-Road Integrated Optimization and Navigation system, known as ORION. UPS developed ORION to optimize the daily routes of its delivery drivers. Research published by INFORMS describes ORION as a metaheuristic optimization system that considers the packages to be picked up and delivered and generates optimized routes for drivers.
From locations to an optimized route
The basic idea can be represented as:
The input is more complicated than simply a list of addresses. UPS has described ORION as considering the stops a driver needs to make and factors such as service requirements. Its later UPSNav system also incorporated detailed location information, including delivery points such as loading docks and receiving areas.
The analytical challenge is therefore to find a route that works well across several competing objectives.
For example, a route may need to balance:
- Distance travelled
- Time
- Number and order of stops
- Pickup and delivery requirements
- Vehicle and operational constraints
- Changing road conditions
This is where optimization differs from a simple prediction.
A forecasting model might answer:
How much demand should we expect tomorrow?
An optimization system asks:
Given everything we know, what action should we take?
That makes route optimization a useful example of prescriptive analytics.
Why optimization matters
Even a small improvement in each delivery route can become significant when repeated across thousands of vehicles and millions of deliveries.
UPS has reported that ORION and related routing technologies were designed to reduce miles driven, fuel consumption and emissions. The company has also continued developing the system so that routes can respond to changing conditions during the day.
This illustrates an important principle in analytics: the value of a model often comes from its scale and repeated use.
A recommendation that saves a few minutes on one route might seem insignificant. Applied across a large delivery network every day, however, small operational improvements can become meaningful business outcomes.
From static optimization to dynamic decisions
Traditional route planning can be thought of as creating a plan at the beginning of the day. Modern logistics systems increasingly need to respond as circumstances change.
A traffic problem may appear. A new pickup request may arrive. A delivery may take longer than expected. A route that looked efficient in the morning may no longer be the best option a few hours later.
UPS has described work on dynamically adjusting routes using changing traffic conditions and remaining deliveries and pickup requests.
This represents a broader shift in analytics:
The system does not simply produce an answer and stop. It can become part of an ongoing decision loop.
That idea is particularly relevant as AI becomes more capable. Modern AI systems can help process complex information and support decisions at a scale that would be difficult for a person to manage manually. But the underlying optimization problem, data quality and operational constraints still matter.
There is also an important human element. A mathematically efficient route may not always be the most practical route for a driver. Local knowledge, unusual delivery conditions and information missing from the dataset can affect the final outcome.
The lesson from logistics is therefore similar to the lesson from demand forecasting:
The purpose of analytics is not to produce an impressive algorithm. It is to produce a useful decision under real-world constraints.
Demand forecasting helps answer what might happen. Route optimization moves one step further by asking what action should be taken given the available information.
Get Ameera Kallil’s stories in your inbox
Join Medium for free to get updates from this writer.
But there is another kind of business decision where the problem is less about optimizing a known system and more about identifying something unusual.
A bank may process millions of transactions that look completely normal. Somewhere among them could be a small number of transactions that require investigation.
Finding those unusual patterns brings us to the third case study: fraud and financial-crime detection.
Fraud Detection in Financial Services
A different challenge appears in financial services. Instead of predicting how many products customers will buy or finding an efficient delivery route, the objective may be to identify transactions that do not fit expected patterns.
Consider a payment system processing millions of transactions. Most transactions are legitimate, which means a fraud-detection system cannot simply treat every unusual transaction as fraud. A transaction made from a new location, for example, may be perfectly legitimate. The challenge is to identify patterns that warrant further investigation while minimizing unnecessary disruption to genuine customers.
This is fundamentally an anomaly-detection and classification problem.
From transactions to risk signals
A simplified workflow looks like this:
The available data can include information such as transaction amount, time, location, merchant information and historical customer behaviour. The precise variables and models vary between organizations and use cases.
Machine learning can help identify relationships within these large datasets. Some approaches learn from previously labelled examples of fraudulent and legitimate transactions. Others focus on identifying unusual patterns without requiring every suspicious pattern to have been labelled in advance.
Research from the Bank for International Settlements (BIS) illustrates this approach in payment systems. Its work on machine learning for payment anomaly detection describescombining supervised and unsupervised machine-learning techniques to identify potentially suspicious transactions.
The distinction is useful.
A supervised model might learn from historical examples:
“These transactions were previously identified as suspicious.”
An unsupervised or anomaly-detection approach can instead ask:
“Does this transaction or group of transactions look significantly different from the patterns we normally observe?”
In a real financial system, these approaches can complement existing rules and other controls rather than necessarily replacing them.
Why fraud detection is difficult
Fraud detection presents a particularly challenging data problem because fraudulent transactions are usually a small minority of all transactions.
Imagine a system processing one million transactions where only a very small fraction are fraudulent. A model that simply predicts “legitimate” for almost everything could achieve a high overall accuracy while still failing to identify the transactions that matter.
This is why accuracy alone can be a misleading measure for highly imbalanced problems.
The cost of different errors also matters.
A false positive occurs when a legitimate transaction is incorrectly flagged. Too many false positives can inconvenience customers and create unnecessary investigation work.
A false negative occurs when fraudulent activity is missed. Depending on the situation, this can result in financial losses and other consequences.
The analytical challenge is therefore,
“Can the system identify meaningful risk while keeping the number of unnecessary alerts manageable?”
That is a much more useful business question.
From rules to machine learning
Traditional financial monitoring often relies heavily on rules. For example, a system might flag a transaction because it exceeds a certain amount or violates a predefined condition.
Rules remain useful because they are understandable and can capture known risks. Machine learning adds another capability: it can identify more complex relationships and patterns within historical data.
The Bank of England’s work on machine learning in financial services describes applications including fraud detection and anti-money-laundering activities, while also discussing the importance of governance, explainability and model risk.
This suggests that the future of fraud detection is not necessarily about choosing between rules or AI. In many applications, the two can work together.
A simplified system might therefore look like:
Where human judgment becomes particularly important
Fraud detection also demonstrates why human validation matters.
A model can identify a suspicious pattern, but suspicious does not necessarily mean fraudulent. There may be legitimate reasons for unusual behaviour that are not visible in the available data.
For example, a customer could suddenly make a large purchase because they are travelling, moving house or making an unusual one-time payment. A model may recognize the behaviour as unusual without knowing the reason behind it.
This is why the output of an analytical system is often better understood as a risk signal rather than an unquestionable verdict.
The system helps prioritize attention. People, business rules and additional evidence can then be used to determine what should happen next.
That distinction becomes increasingly important as AI systems become more sophisticated. A more powerful model can process more information, but it does not automatically possess all the context needed to make a consequential business decision.
The bigger lesson
The three examples now reveal three different roles for analytics:
Although the business problems are different, the underlying principle is similar. Data is transformed into a signal that can support a decision.
And this is where generative AI adds a new dimension. Traditional analytics models have generally been designed for specific analytical tasks. Generative AI can interact with analysts in natural language and assist across several stages of the workflow.
The question is no longer simply whether a model can make a prediction.
It is also:
How can AI become a useful partner in the analytical process without turning the analyst into a passive consumer of machine-generated answers?
That is the next part of the story.
The New Layer: Generative AI in Analytics
The three examples above use different forms of analytics, but they share a common characteristic: the analytical system is designed around a specific business problem.
Generative AI introduces something different. Rather than being limited to one particular prediction or optimization task, it can interact with an analyst using natural language and assist with several parts of the analytical workflow.
That makes it less useful to think of generative AI as simply another type of model. In many business settings, its more immediate role is as an interface between people and data tools.
From writing code to asking questions
A traditional data-analysis workflow might require an analyst to write SQL or Python, clean a dataset, create visualizations and then interpret the results.
Generative AI can assist with several of these steps.
For example, an analyst could provide a dataset and ask:
“Which products experienced the largest decline in sales, and what patterns should I investigate?”
An AI assistant could help explore the data, generate code, suggest visualizations and summarize potential findings. OpenAI’s current data-analysis guidance describes workflows that include uploading CSV or Excel files, exploring and cleaning data, generating visualizations and extracting key findings.
This does not mean that the analyst no longer needs Python, SQL or statistics. In fact, understanding those fundamentals becomes more important when AI is generating part of the work. The analyst needs to recognize whether the generated code makes sense, whether the calculation answers the original question and whether the result is consistent with the data.
The role begins to shift from writing every instruction manually toward defining the problem, directing the analysis and evaluating the output.
AI inside familiar analytics tools
This change is also happening inside tools that analysts already use.
Microsoft’s Copilot in Power BI, for example, allows users to interact with data using natural language and can assist with tasks such as creating reports, generating DAX expressions and exploring insights.
This is significant because the AI does not have to exist as a separate chatbot. It can become part of the environment where the analytical work already happens.
A simplified workflow might look like:
The same principle can apply across different tools and platforms. The technology may change, but the underlying objective remains the same: reduce the friction between a business question and useful evidence.
But faster does not automatically mean better
There is a temptation to assume that if AI can produce an analysis in seconds, the analytical process has become more reliable.
That would be a mistake.
Generative AI can produce incorrect calculations, misunderstand a question, overlook important context or generate an explanation that sounds convincing without being supported by the data. Even AI features embedded inside established analytics platforms have limitations.
Microsoft explicitly notes that Copilot outputs in Power BI are nondeterministic and that users need to evaluate and validate the outputs they receive. It also emphasizes the importance of preparing the underlying data and semantic model properly; poor preparation can lead to inaccurate or misleading results.
This brings us to an important principle:
Generative AI can reduce the effort required to perform analysis, but it does not remove the responsibility for the analysis.
That responsibility becomes especially important when the result affects customers, finances, safety or other high-impact decisions.
AI as an analytical assistant, not an automatic decision-maker
Consider our three examples again.
In demand forecasting, AI could help an analyst investigate trends or compare forecasting approaches. But someone still needs to determine whether the forecast is reasonable and how much uncertainty the business can tolerate.
In logistics, AI could help analyze operational data or explore possible routes. But the final plan has to work within actual vehicle, driver and delivery constraints.
In fraud detection, an AI system may identify an unusual transaction pattern. That signal still needs to be interpreted within the organization’s fraud controls and investigation process.
The pattern is consistent:
This is not simply a philosophical distinction. AI risk-management guidance from NIST emphasizes the need to consider trustworthiness, evaluation and human oversight throughout the AI lifecycle. Its Generative AI Profile specifically addresses risks associated with generative AI systems and ways organizations can manage them.
The strongest use of generative AI in analytics, therefore, may not be replacing the analyst. It may be making the analyst more capable and more efficient while keeping accountability with people.
And that raises a final question: what happens when the AI generated answer looks convincing but is wrong?
That is why human validation is not an optional final step. It is part of the analytical process itself.
