A Service Level Agreement is a contract between the network and the user. The user agrees to restrict their traffic so that its bandwidth is below a specified traffic profile; in exchange, the network agrees to guaranteed transmission characteristics like throughput, delay and number of dropped packets.
This is a standard feature of all network services, both commercial and residential.
When someone refers to network service as a "pipe", they are wrong. They are confusing access line speed with a service level agreement based on a traffic profile.
The capacity internal to a network is designed based on statistical multiplexing, meaning that the internal capacity is less than the total of the access line speeds. How much less can be calculated knowing the historical demand statistics, i.e. how often people transmit packets at their access line speed, hence the term statistical multiplexing.
Since the internal network capacity is purposely designed to be less than the total of the access line speeds (lowering the cost to the users!), then there have to be mechanisms to ensure that the traffic experienced by the network is actually within the design parameters, the usage statistics.
The mechanism for that is to define a traffic profile and sell a service level agreement. The traffic profile specifies the "statistics": the maximum average number of bits per second, maximum number of bits per second, maximum burst size. The service level agreement is between the network and the user: if the user agrees to transmit at or below the traffic profile, the network agrees to guaranteed transmission characteristics.
This 8-minute tutorial explains how this works, and how a network can tag packets so that differentiated services ("Diff-Serv" in Cisco-speak) can be provided for different customers.
