I am attending the IBPMS Expo in Chicago this week, and first keynote was from Jim Sinur, who presented some of the ideas behind his new book (with Jim Odell and Peter Fingar) on distributed intelligent agent oriented BPM. I included some notes from his talk.
Whatever we do, we need to remember CEOs just want Revenue. Our job in BPM is to show CEO how to get higher revenue. At this show, the standard case study is account opening, and that fits well with this idea. Business is in a battle of high revenue per hour worked. In China, they have traditionally felt they have lots of people to work and don’t need to worry about optimum, but even in China that is where the next revolution will come from: optimizing that resource.
It used to be that a business model would be in effect for many years. But now executives need to be on top of it, because there is no such thing as a static business model. Who likes change? Most people hate change, and it is a struggle for everyone. But, how else will you get the edge on the competition?
Processes evaporate in an agent world. Full cycle is observe, orient, decide, and act. (Reference to John Boyd) Processes are only the “act” part. It is going to work like a fighter jet. Every two years someone says “BPM is dead”. Jim doesn’t buy that. There are plenty of healthy BPM vendors. However, BPM is not dead, but they need to be more responsive in real time.
There are many terms used for BPM, and they all mean slightly different things.
- forms driven workflow – for eliminating manual tasks – want to follow existing forms flow, and a focus on data and content capture
- straight thru processing – rule guided machine processing – want automation of repetitive tasks with no human interaction
- content collaboration – shared tasks and content – want to share work and leverage knowledge for speed.
- guided navigation – self-service with guard rails – want work displacement to client and ease of use.
- case management – multi-role/emergent/knowledge intensive – want completion of complex work and emerging situations.
- intelligent processes – context sensitive/agile/poly-analytic – want situational awareness using human skills with technology assists, and near real time responses.
(These map reasonably well to my 7 domains of predictability, maybe I can work with Jim and come up with more detail to map them.)
We need some real world example, that get into messy reality. People and agents will be more interchangeable in the future. If someone is too busy, you might need to give it to an algorithm. When it gets a lot more important, it may be allocated back to a human. These are interchangeable.
People fight about “should I model or should I not model“. In the future you will have models that show what happened. There is nothing wrong with planning by starting with a model. What is wrong is turning that into a theology. It is a starting point, and that is all. Do it, try it, fix it. Ok to start with a model, but not OK to try to make it perfect.
Strong recommendation to get all of these parts from an integrated system. Jim does not recommend composing this by yourself.
If we want to talk about intelligent agents, we need to find a scientific way of measuring intelligence. Jim introduced the Cumulative Process Intelligence Quotient. There are five dimensions: (1) raw intelligence, (20 social intelligence, (3) agility, (4) autonomy, and (5) visualization. Autonomy is the thing that made him think that distributed agents is the right thing.
He presented a couple of use cases:
- a dynamic surgery center where all of the participants, including the people waiting in the lobby, were tagged with sensors that tracked their situation, and allow everyone to keep informed. Given the current situation (how many supplies, how much to do) the system does a forward simulation of what might happen going forward, and assess potential outcomes.
- airport operations in the UK. Start tracking flights when they are 35 minutes away, and do forward simulation of how the terminal will be affected. This is intelligent process. Increased predictability of handling passenger flow.
- a farming operation that uses sensor to determine how much moisture there are in the different parts of the field, and how much fertilizer. Usign this information, they can fine turn the watering that goes, and time the application of fertilizer to when there is enough moisture. Claim to track “by plant” the moisture and are working on figuring what the optimal moisture for a given plant.
- goal modeling: this use case from Compsim which does goal modeling which tells you how goals are related in the predator operation program. (He compared this to Whitestein.) Some goals conflict with other goals. Showed a dashboard taht visually displayed the relationships between goals, adn these relationships can be tweaked.
IBO (Intelligent Business Operations) – collaborative distributed intelligence with machine learning. Example is a security trading desk with many different specialized agents.
Jim calls this a tidal wave of intelligence. Dive in!
Thinking Matters 
Keith, while I know that Jim is right about where the technology will be going, I also know from experience that there is a huge gap between what can be done and what the people in business will accept. Agents sound cool and modern, but they are FAR AWAY from BPM in the real world today. They are even further away from BPM practice. As a first step BPM technology would have to switch to an ACM approach to allow flow, rule, agent and performer interaction in the same domain.
The key problem being that the agent first needs to be able to ‘recognize’ the knowledge in the process system (and that is not just the database, but the process as it played out so far and how it might play out) and then it has to have some knowledge engine how to tell it what to do when. Current BPM systems are utterly incapable of linking such an agent to their systems as they just run an SQL database underneath. You can’t run a real-time knowledge query against such databases. Agents that just execute rules are utterly useless in the real-world as the AI community can attest to. Yes, we can write software modules that control the physical real-world using fuzzy control mechanisms like the automated trains at the airport. But these are closed systems with very limited range of possible situations.
Yes, goals are an essential element of using agents and they are a way to direct the agent. Whitestein goals have nothing to do with agent goals, they just have the same name. Agents will not be able to simply balance conflicting goals, particularly as goal definitions can be simple data rules or they can be larger conceptual situations.
And then there is the human side of using agents. We have introduced the User-Trained Agent in 2007 for the Papyrus Platform and have used it in various projects over the years. We found that businesses find it unsettling that they do not know why the agent will recommend something. The pattern that the agent discovers is known so there is a basic of if-then-else logic, but pattern recognition does not write rules that a business user can understand. We showed in 2008 a demo where we connected a NAO Robot to the Papyrus system and we trained him to recognize color command cards. He would then perform an action connected with the recognized pattern. People loved it, but they find it hard to apply that to the current process and business thinking.
For practical use we stayed far away from allowing agents to perform actions because that was a definitite NO-NO from business analysts and managers.
As for distributed agents that cooperate, we get even deeper into how to analyze the real world and how to apply agent knowledge. We can run any number of agents on any number of distributed nodes, being trained by a variety of performers and creating a library of knowledge related to a particular role. All the agents would do is to recommend an action to a performer. ‘I have seen this pattern of events and data before and the performer with the role clerk added goal to clarify the family status of the insured.’ But IT people will not trust such a ‘thing’ that they have not programmed. The main problem: HOW DO WE TEST IT? IT people think in frozen knowledge that they encode. New knowledge is risky and dangerous and might contain errors or something unexpected. Learning becomes nearly impossible in such environments.
What is discussed as IBO is a mix-and-match of technologies that does not really go together and creates a completely unmanageable construct. It requries a huge software stack and huge manpower to try and put knowledge into the ‘system’. I don’t see it as an advance. It is a step backward. Nevertheless, it is sold and covered by analysts so it must be good.
But do not expect that people will jump at BPM agents any faster than they are with understanding the benefits of ACM …
If you use agents in their ultimate sense, it is early for most folks. However agents can help in a passive mode in dealing with the internet of everything. We talk about two major phases of agent use in BPM in the new book, so we believe in a phased introduction of agents 🙂