The conceptual architecture common to the business intelligence solution is depicted in figure below. While all deployments don’t function exactly like the system seen below, this architecture is becoming more and more common.

Source data
Source data comes from one or more operational databases and sometimes from third-party data, like census or industry data. Source databases are most often relational; however, delimited text files and spreadsheets are common as well.

Extraction, transformation, and loading (ETL)
ETL describes the processes used to copy data from source databases to the data warehouse, but it’s more complicated than simply moving bits. Quite often, data is transformed and validated on the fly. The end-to-end ETL tasks are combined into “packages,” which are scheduled to run automatically at preset times.

The data warehouse
Transformed source data is consolidated into a single relational database called a “data warehouse.” The warehouse physically resides on an industrial-strength relational system, such as Oracle, Microsoft’s SQL Server, or IBM’s DB2. A data warehouse contains read-only data depicting the state of an organization’s information at regular points in time—weekly, daily, or even hourly. Keep in mind that data quality is especially important. You’ll be wasting time and money if end users have issues with the credibility of your data. Another important consideration is query speed. Frustration results if many seconds pass between submitting a query and displaying results. Relational databases don’t always respond quickly to complex queries, so multidimensional “cubes” are increasingly used to bridge the gap.

The cube
A cube is a complex, efficient, and proprietary data structure that includes data and data aggregations (precalculated summary information), as well as security information that controls who can access what. Cubes are lightning-fast when responding to complex queries—at least compared with relational databases. Also, the data within cubes is almost always compressed to reduce physical storage requirements. A multidimensional cube may contain tens of millions of records (an individual scan at the supermarket is a record, for example) and may reach several gigabytes in size. Finally, cube updates are often performed nightly as part of the ETL package so current information is available first thing in the morning.

End-user tools A variety of applications is available to meet the requirements of end users, or you may write your own if you so choose. Client applications fall into one of seven areas:

1. OLAP. This common term stands for “On-Line Analytical Processing,” meaning the end user has direct—or “on-line”—access to the cube by way of PC-based analysis software. Browsing the information freely permits spontaneity that makes spotting trends and relationships easier. Better client applications combine an intuitive interface with rich visuals designed for rapid comprehension.
   
   
2. Static and live reporting. Static reports give you a view of information arranged in a predetermined way—sales by month by region, for example. Different views, say sales by product category, require an expensive special report. BI systems make it fantastically simple to generate special reports.
   
   Live reports allow end users to interactively manipulate information and drill down to more granular levels, in more or less predefined ways.
   
   
3. Balanced scorecards. Scorecards represent a remarkable technology from the minds of Robert Kaplan and David Norton of the Harvard Business School. Scorecards emphasize frequent and timely relevant measuring of individual and team performance against key financial and nonfinancial objectives. Scorecards directly reinforce strategy because performance measures map directly to your organization’s strategic initiatives.
   
   
4. Budgeting and forecasting. In most organizations, budgeting is traditionally painful, tedious, and distracting. Your accounting team painstakingly assembles and then distributes budget packets to departments and branches throughout the organization. Each department then spends many additional and painful hours completing the packets. Later, the packets are collected, consolidated, and reviewed. Often, this cycle is repeated another time or two before the budget is finalized.
   
   
5. Data mining. Here the objective is to recognize patterns and relationships not apparent through simpler analysis methods. Data mining models generally describe buyer characteristics (which group is best targeted given a specific product) or predict a dependent value (which product is best targeted given a specific group). Data mining is the core of high-efficiency database marketing.
   
   
6. Exceptions and notifications. When key performance measures are out of line, software agents instantly take notice and take immediate action. Traditional systems require that humans first observe and comprehend out-of-bounds facts before action is taken. Advanced systems allow end users or managers to link events with appropriate notifications. For example, if non-billable overtime hours were greater than 2 percent of total hours, Joe and his boss will receive an e-mail, a fax will go out to Sue, and Fred will be paged
   
   
7. Business process input. Human beings traditionally carry information from process to process. Now we’re designing information systems that will access information without human intervention. An automated purchase-order system may receive information directly from the BI system to establish a reorder quantity based on an analysis of what’s in the sales pipeline and possibly other factors. Because humans aren’t involved, value-chain processes move faster, better, and at a lower cost.