AI applications can be divided into applications providing decision intelligence or applications automating workflows.

Automating workflows is easy to describe and, at least conceptually, think about where the ROI comes from.

Decision Intelligence is a bit slipperier. But once you understand it, why it matters, and how to take advantage, you are capable of unlocking the tremendous value that AI promises. At the end of this article you’ll know what you need to know to make that happen.

‍

What is Decision Intelligence?

‍

Gartner defines it as “a practical discipline that advances decision making by explicitly understanding and engineering how decisions are made and how outcomes are evaluated, managed and improved via feedback."

Google’s Cassie Kozyrkov, ever the source of brevity, describes it this way: "Turning information into action in any setting at any scale.”

Decisions are everywhere: strategic, tactical, operational, so the hard part is not coming up with use cases. It is harder to pin down exactly where Decision Intelligence fits in as compared to say, data driven decision making. Manifold views the journey from data to ROI as two trips. The first is data to model. This is the focus of data engineering, data science, and MLOps. The end result is model that is in production with CI/CD in place to make sure its performance remains on par. The second trip is model to ROI. Here is where Decision Intelligence comes into play. A model is producing information, but information only has business value if it's acted upon. Actions need to be specified so causes of inertia are identified and fixed. Outcomes of those actions must be persisted so causal analysis can measure if value was produced, and by how much. Finally, the lessons learned need to implemented in feedback loops, and the process repeated.

Regardless of definition, the pain points are well known. No fewer than 3 Nobel Prize winners• have demonstrated humans are full of cognitive biases (especially when faced with choices under uncertainty) and routinely make logical errors that lead to bad decisions in experimental settings.

(ENDNOTE: Herbert Simon in 1978, Kahneman and Vernon Smith in 2002, and Richard Thaler in 2017.)

There’s also considerable evidence that realizing ROI from data, analytics, or AI/ML initiatives is very hard.  Survey after survey tends to cluster around 75% failure rates.

A host of reasons are reported for this, but a significant chunk of them are related to issues that decision intelligence focuses o

‍

Is Decision Intelligence New?

‍

In a word, no. But it is getting a lot more interest in recent years.

As a field of study, “decision intelligence” has been with us for a while.  The modern, normative, intellectual backbone on how to optimally use information to make decisions has been around since at least since Savage's work in the 1950's. (The short answer is that we all should be Bayesians.)

We also have had a good handle on how to set up and solve complex decision problems that evolve over time with imperfect information since about the same time frame. Bellman's equation remains the fundamental underpinning for almost all of Reinforcement Learning.

As a practical discipline, some flavor of decision intelligence has been in use in finance and algorithmic trading firms for some time.

A variety of capital budgeting decision systems have a long history, and real options approaches to project and strategy evaluation came on the scene in the late 1970s.

And commercial solvers such as Gurobi have progressed to the state where solving complex optimization problems in business supply chains and logistics is almost trivial.

So if decision intelligence has been around for decades, why the sudden increase in interest?

‍

Some of that is probably branding. Google, Gartner and others have started talking about it with increasing frequency, which leads other firms to follow suit.

But we also think there are a confluence of technologies that are making it more necessary in more situations. Things like:

• Cloud-scale compute and storage
• Modern AI/ML capabilities
• Embedded, automated feedback loops that measure real-world impact
• User-facing tools that bring decision models to the frontline, not just the C-suite

Our view of Decision Intelligence mirrors in many ways how Zhamak Dehghani describes Data Mesh.  It’s not a platform, or a particular architecture of some kind. It's a sociotechnical approach for producing what we call “Operationally Led – IT Enabled” Analytics.

‍

Three Key Decision Intelligence Principles

‍

An opinionated view of the three most important takeaways from Decision Intelligence are:

It focuses the attention of AI practitioners on the long, and sometimes messy, journey between model and ROI.

There are two journeys that must be taken to go from data to ROI. The first is going from data to production quality model serving. We accomplish this by data science and then MLops or AIops. The second journey is going from information provided by the model to ROI.

Decision Intelligence on this second journey.  Explicitly thinking through the desired actions and results from improved information. All the modeling in the world has no value if it is not acted on to achieve better outcomes.

‍

Scope of application.

There has been a tendency in “decision intelligence” to focus on large, one-off, events such as evaluating a M&A, deciding whether to go forward with a project, or corporate strategy exercises. A typical project consists of a lengthy period of requirements gathering, data sourcing, and analysis, recommendations and scenarios are presented, and then the analysis is tucked away in a folder somewhere and forgotten.

But in our view, the close relation of Decision Intelligence with modern data architectures, data science, AI, MLOps, and cloud compute is designed to democratize, and automate, the process to higher frequency, decision workflows with built in feedback loops

‍

Causality.

Decision Intelligence platforms like to treat decisions as first class citizens. The reason is to bring structure into the actual steps used to come to a decision with the hope of providing guard rails against the various ways the process can go off the rails.

But this is over-engineered in many cases. The added overhead gets in the way of usability, and the time and effort involved runs counter to the spirit of more agile real-time decision making.

However, timestamping decisions and storing metadata is crucial for continuous improvement.  This is where causal analysis comes into play. All Decision Intelligence systems should be fully automated to capture the necessary data and measure improvements that can be causally attributed to the decision process itself.

‍

Achieving Decision Intelligence through Smart Analytics

‍

Decisions are everywhere: strategic, tactical, operational. Use cases are primarily limited by data and imagination. But the most fundamental application of the Decision Intelligence approach is in Smart Analytics.

“Smart” typically means that AI/ML is embedded in the analytics system - either to help in self servicing, or to construct the actual metrics.

Analytics tends to focus on the automated, high touchpoint, higher frequency applications where Decision Intelligence is most valuable.

How does Decision Intelligence apply to modern analytics?

1. Measure what’s meaningful, not what’s easy. Doing so is part a business question and part a technology question. The technology end includes modern data foundations and data pipelines. The business case requires a clear view of organizational goals. If the goals manifest as latent concepts such as brand recognition, satisfaction, etc. AI/ML modeling is needed to construct the metrics.
2. Work through the actions to be taken with analytics users. What happens if a KPI crosses a threshold? Who is on the hook for monitoring? Who is responsible for initiating corrective measures?
3. Develop the architecture and processes for continual improvement. Call it A/B testing, or experiments, or causal analysis. The goal is the same. Actions were taken, results were recorded, and procedures in place to measure if improvement there was and by how much.
4. Usability, transparency, context are key. Data and AI implementations fail primarily due to people and process challenges, not technical limitations. The shift from intuition-based to algorithmic decision-making requires comprehensive change management addressing:
   1. Cultural resistance to data-driven decisions.
   2. More and more, Fear of job displacement
   3. Skepticism about AI recommendations over human expertise.
   4. The transition from "HiPPO" (Highest Paid Person's Opinion) to systematic decision frameworks

‍

Common Decision Intelligence Pitfalls

‍

Common pitfalls include:

• Over-scoping early implementations: The most successful implementations began with one or two high-leverage decision workflows.
• Insufficient validation and feedback loops: Organizations that skipped shadow mode or A/B testing risked unvalidated decision models reaching critical business processes, eroding trust in the system.
• Unrealistic expectations: Expecting comprehensive enterprise-wide transformation without acknowledging that only 33% of decisions are currently made based on data versus intuition creates inevitable disappointment.
• Failure to assign ownership: Without a dedicated steward or center of excellence, frameworks become outdated, questions go unanswered, and decision models grow stale.

In short, all the normal rules of change management with digital transformation initiatives apply, with the added complexity of persuading people/organizations to change how they make decisions.

‍

The Decision Intelligence Rollout Process

‍

A typical rollout might follow this arc:

‍

Assess & Prioritize with Realistic Expectations.

Identify 1–2 business-critical processes where better information integration can improve outcomes or reduce time-to-decision. Start with a proof-of-concept budget ($25K-$100K) rather than enterprise commitments.

‍

Connect Data Sources with Quality-First Approach.

Begin with core systems (CRM, ERP, financial systems), ensuring data quality, governance, and access controls are properly configured. Implement continuous quality checks integrated into decision pipelines rather than batch validation.

‍

Establish trust.

There are often business rules and built in institutional constraints that must be obeyed. Introduce advanced analytics and ML for more complex decision optimization only after a clear understanding of the setting is established and human trust and adoption of the strategic objectives.

‍

Deploy with Comprehensive Governance and Guardrails.

Launch with human oversight, A/B testing against current processes, and clear regulatory compliance frameworks. Monitor decision quality, track business outcomes, and gradually increase automation where appropriate with proper ethical oversight.

‍

Document, Iterate, Scale with Risk Mitigation.

Scale by department with realistic timelines (for example, a pilot implementation of $100K-$500K before scaled rollout) and measure causal impact continuously. Plan for 15-25% annual maintenance costs and continuous model retraining.

‍

Looking Ahead: Infrastructure, Not Dashboards

‍

When deployed thoughtfully with realistic expectations, achieving decision intelligence in analytic systems enables a new operating model for modern organizations. Information becomes actionable, decisions become faster and more informed, choice processes become structured and auditable, and leaders focus more on strategy than analysis.

However, success requires acknowledging that this represents organizational transformation rather than IT project deployment. The focus must be on creating sustainable decision-making capabilities, not implementing technology platforms.

‍

Ready to Explore?

‍

We offer focused, half-day workshops designed to help you:

Evaluate decision maturity and organizational readiness

Map your current decision processes and data landscape

Launch a realistic pilot with measurable decision outcomes

Build a blueprint for enterprise-wide decision intelligence that acknowledges both potential and limitations

Contact us to learn more.

‍