John Parkinson: Secret Software Agents - 2
The process models are really important not only as a design and debugging aid, but as an "out-of-band" control mechanism to detect when the decision rules or smoothing filters need changing.
By comparing results from the real world against predicted results from the models, you can see what, where and when you need to adapt the rules.
Now that we have an idea how to build agents, which business processes are good candidates for transforming into agent-based systems?
In general, you should look at processes that are hard to do well if people are involved. Some symptoms:
Among the business processes that can benefit from software agents: credit scoring, fraud detection, product pricing, risk underwriting, inventory management and resource scheduling.
The potential from agent-based approaches is significant.
We have seen improved return on assets in the range of 7 percent to 15 percent, and it's possible to make the case for significantly more as the agents cover more of a business's operations. You also get a steady improvement in decision quality, because the system "remembers" its decisions, and you can analyze the outcomes as a way to improve both the rules and how they are applied. And you can get greater productivity from people at all levels in the organization, because they can work with better information and can focus their attention on critical process areas where we can't yet provide complete automation.
There are, of course, some downsides.
The closer you get to the theoretical optimum of a process's performance, the greater the risk of a catastrophic failure if you exceed it. When a process operates so close to its maximum that it can't react to a problem (because it has no spare capacity) as quickly as the problem impacts other processes (because they are so interconnected), things can spiral rapidly out of control.
You need a whole new set of avoidance and recovery strategies for agent-based systemsor you must build in adequate operational circuit breakers to isolate process failures before they become disasters.
Nevertheless, the potential for a big jump in productivity and asset yield through the use of agent-based approaches is realand we are going to see more and more software agents in automation designs. It's only a matter of time before this kind of "attention management" replaces most routine business decision making.
John Parkinson is chief technologist for the Americas at Capgemini.