Wrapping up my series on decision agents, here’s the third post.

1. How AI Agents and Decision Agents Combine Rules & ML in Automation
2. Building Decision Agents with LLMs & Machine Learning Models
3. Designing AI Decision Agents with DMN, Machine Learning & Analytics [this post]

The most effective way to define decision agents is using decision modeling. Just as you build a data model for a database or a process model for workflow, a decision model lets you create a visual blueprint for your decision agents. We use the industry standard notation for decision models – the Decision Model and Notation or DMN. This is partly because it’s a standard and partly because the basics of DMN are incredibly simple – three shapes and two lines. Yet this simplicity supports enormously complex decision agent designs.

I won’t get into the details here – there’s a lot of material out there including my book with Jan Purchase. Suffice it to say you can model out enormously complex decision agents, breaking down their decision-making into its component pieces and then specifying the logic for any piece that must be consistent and prescriptive while identifying the right kind of machine learning model for probabilistic decisions such as determining the sentiment of a text field for instance.

The prescriptive decisions in such a model can easily be implemented using a decision platform while the others can be executed on AI/ML models and even specified using standards such as PMML – Predictive Model Markup Language, an XML standard for interchanging predictive models or ONNX, Open Neural Network Exchange, which exchanges graph models between different ML platforms.

Each decision agent can leverage one or more decision services defined this way using MCP to communicate with the stateless services.

Besides integrating LLMs into the model for execution, you can also use them to help you build the models. While it is quick to build these by hand–10x faster than writing requirements documents, LLMs can accelerate this even further by taking your policy documents, Standard Operating Procedures and regulations and extracting initial partial decision models from them. These won’t be 100% right because most organizations don’t have everything documented but they will accelerate your process. You can also use LLMs trained on programming languages to extract models from code.

The DMN model represents a precise, visual definition of your decision-making that matches the behavior of your decision agents. This allows you to easily track changes, mix and match the right technology for each agent, engage business owners in the definition of your agents and produce regulatory documentation of how you decided. All things that are REALLY hard to do any other way.

To learn more about decision modeling with DMN, there are several options:

• For those who prefer video, check out Designing AI Decision Agents with DMN, Machine Learning & Analytics on the IBM Technology channel
• For those that prefer to read, download our white paper on Agility and Efficiency with Decision Modeling or check out Real-World Decision Modeling with DMN, the book I wrote with Jan Purchase.

Connect with me here or on LinkedIn if you want to talk about doing this in your own environment.

Continuing my series on decision agents, here’s the second post.

1. How AI Agents and Decision Agents Combine Rules & ML in Automation.
2. Building Decision Agents with LLMs & Machine Learning Models [this post]
3. Designing AI Decision Agents with DMN, Machine Learning & Analytics

If you are building a sophisticated agentic AI system that is intended to solve complex, real-world problems, you’re going to need decision agents. These systems must make autonomous decisions that directly impact customers, operations, and business outcomes. However, the technology that powers modern agentic AI—large language models—creates a fundamental paradox: while LLMs excel at many tasks, they are poorly suited for the precise, consistent, and transparent decision-making that business-critical applications require.

So why are Generative AI models, LLMs, unsuitable for advanced decision-making? Several reasons:

• They are inconsistent by design, not something we look for in decision-making
• They are opaque and black-boxy, making it hard to explain why a decision was made
• They are poor at mathematical analysis and much worse than other machine learning techniques
• It’s hard to make small, focused changes to their behavior, limiting continuous business-driven improvement

A good decision agent should be ruthlessly consistent, completely transparent, easy to change, accessible to domain experts and able to embed advanced analytics and machine learning. Which means they should be built using a platform that meets these criteria – a Decision Platform or a Business Rules Management System. These are widely used and are ideal for building decision agents. For instance, Decision Agents can leverage any of IBM’s Decisions technology – IBM Operational Decision Manager (ODM), Automation Decision Services (ADS), Decision Manager Open Edition (DMOE) or the new Decision Intelligence.

Plus, these platforms can be enhanced with generative AI by using it to ingest unstructured information, explain decisions made in natural language and suggest improvements.

To learn more about decision agents, there are two options:

• For those who prefer video, check out Building Decision Agents with LLMs & Machine Learning Models on the IBM Technology channel
• For those that prefer to read, download our white paper on Building Effective Decision Agents.

Connect with me here or on LinkedIn if you want to talk about doing this in your own environment.

Business automation is being transformed by Agentic AI. Generative AI (Large Language Models – LLMs) and the Agentic AI pattern together create new ways to solve business problems. As you would expect, I am particularly interested in what you might call “decision agents” – agents in an agentic framework specifically focused on automating business decisions. How do you identify, integrate, specify, design and build such decision agents and what technology and design approaches do you need?

As part of thinking about this topic, I recorded three lightboard videos for the IBM Technology YouTube channel and blogged about them here. Today I’m starting some longer pieces on the three topics.

1. How AI Agents and Decision Agents Combine Rules & ML in Automation [this post]
2. Building Decision Agents with LLMs & Machine Learning Models
3. Designing AI Decision Agents with DMN, Machine Learning & Analytics

Agentic AI combines a new architectural approach with LLMs to significantly improve business automation. It’s particularly well suited for creating systems with greater autonomy and such systems rely on agents making critical business decisions without human intervention. That said, agents based on LLMs have real limitations of transparency, consistency, state management and regulatory compliance of complex logic. Plus relying on large language models for autonomous decision-making fails to leverage existing investments and best practices in business automation.

Multi-method Agentic AI – an architectural approach that combines agents built with large language models with others based on proven automation technologies including workflow and decision platforms – resolves these limitations and leverages existing investments.

A multi-method approach divides agents into different types or classifications, applying the right technology (or mix of technologies for each). For instance:

• Chat agents for natural language interactions use generative AI to handle questions and requests from users.
• Orchestration agents use generative AI to route interpreted requests to the right specialty agent.
• Policy agents use Retrieval Augmented Generation (RAG) and generative AI to answer a wide range of general questions.
• Workflow agents use process management technology to handle complex sequences of steps and manage state.
• Decision agents deliver consistent and explainable decisions using decision platforms and business rules.
• Document ingestion agents use generative AI to extract needed information from free from documents.
• Explainer agents translate decisions made by decision agents into natural language for customers and staff.
• Companion agents use generative AI to support staff as they handle manual steps and reviews.

While many agents in such a framework do use generative AI, not all do. Such a multi-method approach improves confidence and transparency, leverages existing technology investments AND puts generative AI to work effectively. It combines the best tools for each task, maintains transparency where it matters, and builds systems that serve both customers and stakeholders with excellence.

To learn more, here are two options:

For those who prefer video, check out How AI Agents and Decision Agents Combine Rules & ML in Automation on the IBM Technology channel

For those that prefer to read, download our white paper on Multi-method Agentic AI.

Connect with me here or on LinkedIn if you want to talk about doing this in your own environment.

Along with a bunch of friends and colleagues, I have spent the last few years working on DecisionAutomation.org – a multi-vendor organization dedicated to documenting decision automation best practices. A key focus of this organization has been on how best to use decision modeling with the Decision Model and Notation standard. This matters because OMG, the organization that runs the standard, explicitly forbids “methodology” in its standards – you can’t standardize how to do things, only the result. This means there are elements of DMN that are rarely a good idea, but the standard can’t tell you when they are – as long as there is SOME reason to use them, the standard needs to define them.

To help organizations adopt DMN effectively and maximize their value, DAO identified multiple ways they might want to use DMN. Each of these uses different pieces of the standard and needs different kinds of software support. To make it easy to identify what will work for you, we built a great tool we call the Interactive Usage Scenario tool. This lets you identify what you are trying to achieve and see which elements of DMN and which software features in a DMN tool will help you be successful.

The tool has you pick a business path such as transformation through technology, people development or focusing on a specific target use case – and then pick one of the core scenarios for decision modeling – anything from a simple decision inventory to detailed decision definitions to deployable decision services. Based on these, it identifies DMN and tool capabilities you will need across decision requirements diagrams, model support, decision logic and KPIs, interpretability, implementation and terminology management. It will show you what’s needed, what’s helpful and what’s optional.

You can watch an introductory video here

Check out the Interactive Usage Scenario tool, see who’s participating and sign up at Decision Automation.Org

Business automation is being transformed by AI, specifically Generative AI or Large Language Models (LLMs). More recently, the Agentic AI pattern has gained real traction among those trying to apply LLMs in an enterprise context. Like any new pattern, there’s a ton of hype about Agentic AI but it’s clearly here to stay.

I have been thinking about this new paradigm and this new technology and my 25+ years of thinking about decision automation plays into it. Specifically, how you identify, integrate, design and build decision agents in an agentic framework. This is a hot topic among our clients at Blue Polaris, as you can imagine, and a space in which we are investing heavily.

While I was thinking about this, the folks at IBM introduced me to lightboards and their incredible lightboard channel – IBM Technology – YouTube. There’s a huge array of great videos on here – from very short to very long on everything from AI to data storage to REST to Docker.

I really wanted to be part of this great channel, so I recorded three lightboard videos for it:

The first is an overview of the role of Decision Agents in Agentic AI – How AI Agents and Decision Agents Combine Rules & ML in Automation. This discusses an agentic AI approach to loan origination and shows how Decision Agents can deliver efficient, consistent decision-making in an agentic framework.

The second drills into how a Decision Agent works – Building Decision Agents with LLMs & Machine Learning Models. This discusses how they complement LLMs and how decision agents can combine LLMs, machine learning models and business rules to enable scalable decision-making.

The final one shows specifically how you can design these Decision Agents visually using the Decision Model and Notation or DMN standard – Designing AI Decision Agents with DMN, Machine Learning & Analytics. It shows how DMN lets you think through the behavior you need in a decision agent and design an agent that leverages the right mix of technologies for a successful implementation. You can check out mine and Jan’s book Real-World Decision Modeling with DMN for more details on decision modeling.

I hope you enjoy them.

Practical Business Process Modeling and Analysis written by Jim Sinur, Zbigniew Misiak (two old friends) and BJ Biernatowski.

I’ve been working in decision automation a long time. While I’ve never been an expert in business process modeling, I’ve interacted with a lot of business process experts, worked on some projects that combined processes with decisions and even helped write a book (the MicroGuide I wrote with Tom Debevoise).

I recently got my hands on a great new book from Packt. I’ve interacted with Jim and Zbigniew many times over the years going back more than 20 years in Jim’s case and at least 10 in Zbigniew’s.

The new big has 10 chapters:

1. Winning at Digital Transformation with Process Modeling
   The role of process modeling in successful digital transformations
2. Pillars of a Successful Digital Transformation
   Key elements of the various kinds of digital transformation and how to lead them
3. The Wheel of BPM Driving Your Competitive Advantage
   Using BPM to drive transformations and start the improvement flywheel
4. Long term Trends and the Impact on Your Job
   Why digital change is so central
5. Business Process 101
   The basics
6. Establishing Process Architecture
   The architecture
7. Process Modeling Notations
   Why we ended up with BPMN
8. BPMN What You need to Know
   The core BPMN concepts
9. Advanced BPMN
   More advanced stuff and a little (too little) on DMN
10. Measuring the Business Value of Process Transformation
    How to measure value and calculate returns – business value assessments

There’s a lot of good stuff here. I really liked how the BPMN mechanics are wrapped with a focus on transformation and business value. The authors tie the act of process modeling to some key business concepts in a really useful way. There’s some good thinking about how AI will, and won’t, impact all of this, and plenty of focus on the human and organizational elements that are so important.

The book reads well and easily, with only small differences in style between the authors – each of whom brings a unique focus to their chapters and complement the others nicely. Its target audience – business professionals who need better ways to design and think about their business as they transform it – will find it easy to read and full of great insight.

Highly recommended. Preorder or buy here.

As 2024 wraps up, I thought I would share a few thoughts on Artificial Intelligence (AI) in the context of automating your business. It’s been a really interesting year for AI and I suspect 2025 will be more interesting yet.

• It’s important to remember that AI > Generative AI. Focus on applying the right mix of technologies to any artificial intelligence problem – business rules and decision automation, statistical models and predictive analytics, neural networks and machine learning, AND generative AI. Few problems only need one of these technologies, most are efficiently solved using a mixture. All these technologies are improving and evolving, and all will have a role to play in 2025. Make sure they’re all in your tech stack.
• Given you’ll need a mixture of technologies to solve problems, make sure to adopt a way to define requirements and do design that isn’t tied to a particular technology but let’s you focus on your automation goals. Decision modeling and process modeling are both great places to start. Far too many organizations have great plans for using new technology next year but no matching plan to change the way they plan and design their systems. Make sure you do.
• Continuous improvement will always matter. You’re not going to solve the whole problem with v1 so don’t assume you will. Plan for change, evolution and improvement. And capture the data about what you did and how well it worked out so you can do this. Good business-centric design and transparent capture of execution logs will be important next year – more so as your systems get more complex.
• Don’t let worries about generative AI stop you getting started now. These tools are getting really good at handling complex inputs (processing documents, supporting conversational interfaces) and explaining outcomes – you can’t afford to wait. Bring foundational models in house and add a good design so you control both how your AI is built and how it is used so you can ensure compliance and safety. Start soon.
• Plan for governance. Manage your process, your decisions and the way ML and AI are being used to support both. Even if you are just getting started, you’ll eventually get asked to prove you have everything under control so start early.

Finally, I’m reminded of Amar’s Law – we tend to overestimate the short-term impact of new technologies while underestimating their long-term effects. AI seems likely to follow this law – it’s not going to upend your business tomorrow, but it is going to radically reshape in the coming years.

Have a wonderful holiday season.

Gartner has increasingly been discussing Decision Intelligence as a concept, as well as Decision Intelligence Platforms. Recently, they published a Market Guide for Decision Intelligence Platforms, which is highly recommended reading.

Decision intelligence platforms combine explicit decision modeling, AI, analytics and related capabilities to support, augment or automate decision making, driving business outcomes. Data and analytics leaders can use this guide for investing in DIPs to make their organization more decision centric.

Gartner’s focus on Decision Intelligence represents an effort to consolidate fragmented tools and approaches aimed at enhancing decision-making. They correctly argue that the growing complexity of business environments, regulatory demands, and adoption of consumer technology necessitate improved automation and management of decision-making processes. The proliferation of AI and advanced machine learning technologies is further accelerating this trend.

Our longstanding advocacy for building a comprehensive decision management capability—by focusing on platforms and techniques that allow broad application of decisioning technology rather than piecemeal solutions—aligns well with the report’s emphasis. It is also encouraging to see AI, simulation, data science, and business rules highlighted collectively, a stance we have maintained for more than 15 years and is one of our three best practices.

In addition to technology, Gartner is talking more about modeling decision-making explicitly. This aligns with another of our three best practices – DecisionsFirst as we like to say. However, Gartner still promotes flow-based models for decision-making, whereas the industry has shifted towards declarative models like those based on the Decision Model and Notation standard. These models facilitate quicker requirements gathering, greater business engagement, increased reuse, and better integration of rules with AI and machine learning.

Their emphasis on monitoring and continuously enhancing decision-making is crucial and constitutes the third of our three best practices. Since decision-making is dynamic, it is essential to build systems with flexibility for change and improvement. Selecting Decision Intelligence Platforms based on their support for maintenance and enhancement is therefore key.

As for the technology vendors themselves, they’re a real mix. Some are very focused on specific pieces of the platform (simulation or collaborative decision-making) while others are “legacy” vendors with business rules, optimization or analytic platforms. Some though really do have a complete platform – certainly for the kind of high-volume, operational decisions we specialize in.

If you are looking to adopt a Decision Intelligence Platform or need assistance in evolving your existing system to provide more comprehensive support for robust decision-making capabilities, please reach out.

We’ve been focusing on this for quite some time – long before it was cool enough to have a Gartner market name! ?

Cross-posted from our company blog

As regular contacts of mine will have noticed, we have a new company name and a new brand

Artificial Intelligence (AI) is changing the way companies approach technology, and we are broadening our palette of technologies and services to match. A new name, and a new focus, are required. Blue Polaris will still be helping companies apply machine learning (ML), AI, business rules and decision modeling but we’ll be helping think about how AI changes their workflow and document management, how governance needs to evolve to bring ML/AI into compliance, how to capture all the knowledge they have and much more.

It’s been 15 years since I founded Decision Management Solutions and I’m super-excited about this next chapter.

Onward and upward.

You may have noticed articles about a chatbot recently that got a little out of line – an airline’s chatbot misstated the rules for a fare class (see The Guardian‘s article or The Washington Post‘s). The airline has, of course, been held accountable for its chatbot – just as it would have been for an employee. Two key lessons can be learned from this outcome:

• You are responsible for everything your chatbots say, even their hallucinations and errors.
  I would have thought this was obvious but apparently the airline’s lawyers thought that blaming the chatbot might work!
• You don’t want your chatbot making decisions about things like eligibility, pricing, discounts – decisions that are regulated, based on complex and published policies, and that impact customers.

The airline’s intent here was a good one I think – use a chatbot to make it easier for people to get answers to questions about the notoriously complex topic of fares. The power of Large Language Models (LLMs) and Generative AI (GenAI) to power more interactive chatbots is real and is going to change how consumers use your website and understand your intent. They can dramatically improve explicability, making your website/systems easier to use, easier to understand and fundamentally less technical to access.

But there are issues. What AI chatbots say is not always reproduceable. They may hallucinate – sometimes spectacularly and with references! How they work is largely inexplicable – especially to regulators. And even bad answers look like good ones. And, as this story shows, you’re going to be held accountable for them.

The solution is not to dump LLMs/GenAI from your roadmap but to recognize that this technology has no sense of the truth or facts and simply generates the most likely content – it’s not prescriptive. You need to add prescription so you can precisely define what the chatbot should do in which circumstances that is based on ground truth and factual content. While LLMs and GenAI are great for interacting with customers and explaining results, they can’t be trusted to prescriptively make regulated or policy-based decisions.

Adding decisioning based on business rules – explicit decision logic – grounds their behavior in facts and rules. Modern decisioning platforms are great for transparency and consistency, especially when decision modeling is used to manage the logic. Using a decisioning platform to automate decisions like eligibility (for a fare, product, service or benefit), dynamic or complex pricing, risk assessment gives you precise business control over your decisions. Unlike a chatbot, the logic is explicit, explainable and managed.

So why not JUST use decisioning? Decisioning platforms deliver APIs aimed at internal systems. The decisions are compliant, precise, transparent – but not accessible to a customer. Typically, you have to put all the data needed for the decision into forms and processes before you can get an answer. Adding LLMs/GenAI to handle the interaction provides a customer-friendly interface to the decisioning APIs and delivers both a great interaction and reliable, compliant decisions.

This was a topic of a webinar we did with IBM recently – How to achieve more trustworthy Generative AI with Decision Automation [free registration required]. See also this post about using AI to improve interactions and this one on using ML/AI to improve the operational decisioning itself.

If you are interested in learning more about how you can combine AI-driven decisioning with chatbots, drop us a line. Or, if you are based in the NYC area, register for our upcoming event April 10: Unlocking the Power of Automated Decisions: Harnessing the Power of AI/ML for Intelligent Rules