Decision intelligence
What is Decision Intelligence?
Decision Intelligence makes the decision itself the subject of design: what options do we have, what might happen, how much is each outcome worth to us—and how do we learn from the impact? An overview of the concept, methods and limitations, with a fully worked example.
At a glance
- Decision Intelligence treats the decision as the unit of analysis: signals, context, options, approval, outcomes and feedback are designed together, beyond data and dashboards alone.
- Each building block answers a different question: statistics (What do we know, and how certain are we?), simulation (What could happen?), optimization (What is the best feasible action?) and learning from feedback (What does the outcome teach us?).
- A good decision is judged by its process, not by a single outcome. Decision Intelligence therefore needs a baseline, criteria defined in advance and clear responsibilities.
Most data projects end with a number: a dashboard metric, a forecast or a score. The real work starts afterwards—someone must decide what to do with that number. Decision Intelligence starts at precisely this point. It treats the decision, rather than the data, as the object to be designed, measured and improved. This article puts the term in context, shows which method answers which part of the question, and works through a small example in full.
A working definition
Decision Intelligence is an approach that designs decisions as a repeatable process that can be evaluated. Lorien Pratt describes it as a discipline that explicitly connects data, actions and outcomes (Pratt 2019). The main novelty is the framework rather than its individual tools: statistics, operations research and decision analysis have existed for decades. Howard Raiffa described decision analysis as a discipline in its own right in the late 1960s (Raiffa 1968). Decision Intelligence connects these building blocks into an end-to-end cycle—and adds what many analytics projects lack: approval by a responsible person and feedback from the observed outcome.
The six steps in detail:
- Signals—measurements, reports and forecasts, each with its source and age.
- Context—the relationships that give a number meaning: which location, commitment or constraint does it relate to?
- Options—the actions actually available, including “wait” and “obtain more information”.
- Decision—evaluating the options against goals agreed in advance, with a designated role responsible for approval.
- Outcome—what actually happens, measured using the same quantities used in planning.
- Feedback—comparing expectations and results to improve models, assumptions and rules.
The central issue lies between steps 3 and 4. A forecast says what is likely to happen. The action that follows also depends on goals, costs, constraints and willingness to take risks. The same sales forecast can lead two companies to different order quantities, each well justified—the example below shows why.
From Business Intelligence to Decision Intelligence
Business analytics is often described as a series of stages, from standard reports through statistical analysis and forecasting to optimization (Davenport & Harris 2007). Decision Intelligence is not another stage on top; it connects these stages to the decision they serve:
| Guiding question | Typical methods | Output | Blind spot | |
|---|---|---|---|---|
| Business Intelligence (Descriptive Analytics) | What happened? | Metrics, reports, dashboards | Picture of the situation | Does not say what to do |
| Predictive Analytics | What is likely to happen? | Time-series models, classification, regression | Forecast, ideally as a distribution | A forecast is not yet an action |
| Prescriptive Analytics | What should we do? | Optimization, simulation, decision rules | Recommended action | Often ends with a recommendation, without feedback |
| Decision Intelligence | How do we decide—and what do we learn? | Combination of the building blocks, plus approval and outcome measurement | Decision process that can be evaluated | The effort pays off especially for recurring, consequential decisions |
The distinction from Prescriptive Analytics is small but important: Decision Intelligence does not end with a recommendation. Approval, implementation and measurement of the outcome belong to the system—as does the question of who is responsible for a decision.
The building blocks: which method answers which question
A reliable basis for decisions answers five distinct questions. Each has its own method:
| Question | Method | Explore further |
|---|---|---|
| What do we know—and how certain are we? | Statistics, Bayesian methods | Bayesian statistics, Bayesian A/B testing |
| How much is each outcome worth to us? | Decision theory | Decision theory |
| What could happen? | Simulation | Monte Carlo simulation |
| What is the best feasible action? | Optimization | Optimization |
| What does the outcome teach us? | Bandits, Reinforcement Learning | Multi-Armed Bandits, Reinforcement Learning |
What do we know—and how certain are we?
At its best, statistics provides a distribution of plausible values rather than a single number. Bayesian methods do this particularly directly: the posterior describes what is known about a quantity after observing the data. This is the raw material for decisions. A sales forecast of 1,000 units has little value if it leaves open whether 700 or 1,300 are equally plausible. A prediction interval makes that range visible (Hyndman & Athanasopoulos 2021); its calibration across many forecasts shows whether it can be trusted.
How much is each outcome worth to us?
Decision theory connects probabilities with valuations: how good is each outcome, and how heavily does a loss weigh against a gain of the same size? Maximizing expected profit alone means acting as a risk-neutral decision-maker—an assumption, not a given. The value of information follows from the same logic: additional data is worthwhile only if it can change a decision. Douglas Hubbard therefore recommends first asking which uncertainty actually affects the choice, and measuring precisely that (Hubbard 2014).
What could happen?
Simulation plays out actions across many scenarios. Monte Carlo simulation draws thousands of variants of uncertain inputs and shows the distribution of outcomes—including the uncomfortable tails that an average hides.
What is the best feasible action?
Optimization searches all actions that respect capacities, budgets and rules for the one with the best objective value. It is only as good as the formulation of the objective and constraints: an unnoticed change in a weight shifts priorities without anyone having agreed to it. Objective functions and constraints therefore need to be documented and agreed, not merely implemented.
What does the outcome teach us?
When decisions recur, a system can learn from their consequences. A/B tests are a controlled form of this learning. Multi-Armed Bandits allocate trials so that learning incurs relatively little cost along the way; Reinforcement Learning extends the idea to sequences of decisions whose effects only become visible later.
Games offer a precise testing ground for these questions because the rules, game state and outcome are clearly defined. Through DeepSquare, NyxAI investigates how search, evaluation and time budgets interact in chess decisions under time pressure—as an explainable chess AI with human playing styles, available to play on deepsquare.ai. What transfers most directly is the method; applications beyond chess require their own data, assumptions and tests.
You can try several of these building blocks on a small scale in the interactive explainers.
A worked example: how much should we order?
Before the season, a retailer decides how many units of a seasonal product to buy; replenishment is not possible. Each unit sold earns €40, purchasing costs €22 per unit, and unsold stock can be cleared after the season for €10. There are also €5,000 in fixed listing and campaign costs. In inventory theory, this pattern is known as the newsvendor problem (Porteus 2002).
Step 1: the forecast as a distribution
The sales forecast is 1,000 units—with uncertainty. For simplicity, assume that demand is normally distributed with a mean of 1,000 and a standard deviation of 300. The 80% prediction interval therefore runs from approximately 616 to 1,384 units. This range alone shows that “order 1,000” is not an obvious answer.
Step 2: options and valuation
Profit depends on the order quantity and demand :
We compare a cautious quantity (800), the point forecast (1,000), a generous quantity (1,200) and the mathematically optimal quantity from step 4.
Step 3: simulation
For each option, two million demand values are drawn from the predictive distribution and the profit is calculated for each. The simulated profits provide the expected value, quantiles and probability of a loss (rounded figures):
| Order quantity | Expected profit | 10th percentile of profit | P(loss) | Expected surplus | Expected shortage |
|---|---|---|---|---|---|
| 800 | €8,040 | €3,850 | 4.3% | 45 | 245 |
| 1,000 | €9,410 | €1,450 | 7.4% | 120 | 120 |
| 1,076 | €9,520 | €540 | 9.0% | 162 | 86 |
| 1,200 | €9,240 | −€950 | 11.9% | 245 | 45 |
The values agree with the closed-form formulas for the normal distribution. Simulation has the advantage that it also works with skewed distributions, capacity constraints or multiple products.
Step 4: optimization
Which quantity maximizes expected profit? A missing unit costs €18 in foregone contribution margin (), while an extra unit costs €12 in write-downs (). At the optimum, the probability that the order quantity is sufficient must equal the ratio of these costs—the critical fractile of the newsvendor model:
This gives . Because a missing unit costs more than a leftover unit, the optimum is above the point forecast.
Step 5: the decision
The table shows what the point forecast alone conceals. Ordering 1,076 rather than 1,000 units adds only a little over €110 in expected profit, but the probability of a loss rises from 7.4% to 9.0%. Ordering 800 units yields around €1,500 less in expected profit than the optimum, but the probability of a loss is only half as high. The right option depends on how much risk the business can and wants to bear—a valuation question, not a forecasting question. A retailer with many similar, independent products can take a risk-neutral approach and choose 1,076; for a small business that would be hit hard by a loss, 1,000 or even 800 units may be the better-justified choice. The calculation makes the trade-off visible; the people responsible must decide.
Step 6: feedback
After the season, compare expectations with reality: was demand within the prediction interval, and how far apart were expected and actual profit? After several seasons, it becomes clear whether the 80% intervals actually covered demand in roughly eight out of ten cases. If they miss substantially more often, the forecast is overconfident—and so is every optimization based on it. This comparison is what closes the loop.
Decision quality is not outcome quality
Suppose demand this season is only 700 units. Under the assumed distribution, demand of 700 units or less is not unusual: it occurs roughly once every six seasons. An order of 1,076 units leaves a profit of €3,088; ordering 800 units would have yielded €6,400. In hindsight, the smaller order looks wiser. Before the season, under the agreed criterion of expected profit, it was not.
Annie Duke calls the mistake of judging a decision by a single outcome “resulting” (Duke 2018). Decision quality is assessed through the process: were the relevant options known, was uncertainty quantified honestly, and was the valuation agreed in advance? Across many decisions, the outcomes will also show whether the process works. Three rules follow for evaluating a decision system:
- Establish a baseline. How were decisions made previously, and with what results? Without that baseline, any improvement is merely a claim.
- Set criteria in advance. Which quantity matters—profit, service level, probability of loss—and over what period? A decision rule formulated only after the outcome justifies rather than tests.
- Document the comparison. Record the recommendation, the actual decision, the reasoning and the outcome. Only then can you distinguish whether a model, an assumption or a deliberate deviation explains the result.
Human judgments are not the gold standard against which models should be measured. Kahneman, Sibony and Sunstein show how widely experts’ judgments of identical cases vary—a noise that occurs alongside systematic biases (Kahneman et al. 2021; Kahneman 2011). Structured decision processes reduce both by treating comparable cases according to the same criteria.
Governance: who decides?
A decision system does not transfer responsibility to a model. Governance determines who owns a system, who approves recommendations, how risks are measured and how a decision can be reconstructed afterwards. The NIST AI Risk Management Framework organizes these tasks into four functions: Govern, Map, Measure and Manage (NIST 2023).
A graduated approach works well in practice. A system can execute routine cases within clear boundaries itself; cases outside the agreed boundaries go, with an explanation, to a person who reviews and approves them. This principle is called Human-in-the-Loop. It only works if a recommendation reveals the assumptions behind it—otherwise approval becomes a formality.
Common mistakes
- Confusing forecasts and decisions. The order quantity does not follow from the forecast alone, but from the forecast, costs and willingness to take risks.
- Passing on point estimates instead of distributions. Reporting only “1,000 units” hides precisely the information that determines the risk.
- Leaving goals and constraints implicit. Optimization without explicit objectives and constraints optimizes the wrong thing—with great precision.
- Failing to plan a feedback channel. Without comparing expectations and outcomes, model errors remain undetected and recur.
- Automating too early. Automating a decision process before it is understood and measured also scales its weaknesses.
- Leaving responsibility unresolved. If nobody has to approve a recommendation, nobody may feel responsible when things go wrong.
Start small
The first use case does not have to be the largest. A more valuable choice is one where, after a few weeks, you can establish reliably whether the new basis for decisions is better than the old one—and where those involved can understand the recommendations.
Conclusion
Decision Intelligence is less a new technology than a new framework for proven methods. Statistics quantifies what we know; simulation shows what could happen; decision theory and optimization determine the best choice under our own goals and constraints; learning from feedback improves the next decision. Together, they create a process that remains understandable—and in which people make decisions on a better basis.
Frequently asked questions
What is Decision Intelligence, in simple terms?
Decision Intelligence designs decisions as a process that can be evaluated: signals and context are captured, actions are compared using statistics, simulation and optimization, a responsible person approves the choice, and the observed outcome informs the next decision.
What is the difference between Decision Intelligence and Business Intelligence?
Business Intelligence describes what has happened through reports and dashboards. Decision Intelligence starts where that information needs to become an action: it compares options under uncertainty, makes goals and constraints explicit, and measures whether the decision had the intended effect.
Does Decision Intelligence replace human decision-making?
No. It improves the basis for decisions and can automate routine cases within clear boundaries. Governance determines who owns a system, who approves recommendations and how exceptions are handled. Responsibility remains with people.
Which methods belong to Decision Intelligence?
The main methods are statistics and Bayesian approaches for uncertainty, decision theory for valuing outcomes, simulation for possible trajectories, optimization for the best feasible action, and learning methods such as bandits or Reinforcement Learning for feedback from outcomes.
Sources and further reading
- BookPratt, L. (2019): Link: How Decision Intelligence Connects Data, Actions, and Outcomes for a Better World. Emerald Publishing.
- BookRaiffa, H. (1968): Decision Analysis: Introductory Lectures on Choices under Uncertainty. Addison-Wesley.A classic introduction to decision analysis, one of the roots of Decision Intelligence.
- BookDavenport, T. H. and Harris, J. G. (2007): Competing on Analytics: The New Science of Winning. Harvard Business School Press.
- BookHyndman, R. J. and Athanasopoulos, G. (2021): Forecasting: Principles and Practice (3rd edition). OTexts.Freely available textbook covering predictive distributions and prediction intervals, among other topics.
- BookPorteus, E. L. (2002): Foundations of Stochastic Inventory Theory. Stanford University Press.Covers the newsvendor model and its critical-fractile condition.
- BookHubbard, D. W. (2014): How to Measure Anything: Finding the Value of Intangibles in Business (3rd edition). Wiley.Measurement as a reduction in uncertainty, calibrated estimates and the value of information.
- BookDuke, A. (2018): Thinking in Bets: Making Smarter Decisions When You Don't Have All the Facts. Portfolio/Penguin.
- BookKahneman, D. (2011): Thinking, Fast and Slow. Farrar, Straus and Giroux.
- BookKahneman, D., Sibony, O. and Sunstein, C. R. (2021): Noise: A Flaw in Human Judgment. Little, Brown Spark.
- ReportNational Institute of Standards and Technology (NIST) (2023): Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1. U.S. Department of Commerce.
Content last reviewed on 25 September 2026. Please let us know about any errors: impressum@nyxai.com.