Market and product details are as of 22 September 2026. INVISOM has no commercial relationship with any vendor named here.
Better reports have not produced better decisions
Most large organizations have spent two decades building business intelligence. They have warehouses, governed models, self-service tools and hundreds or thousands of reports. The evidence that this has made them better at deciding is surprisingly thin.
In McKinsey's 2019 survey on decision making, only 20% of respondents said their organizations excel at it, and a majority said that at least half of the time they spend on decisions is used ineffectively. Gartner's survey of 566 data and analytics leaders, published in March 2023, found that only 44% said their team was effective at providing value to the organization. Both numbers were collected in organizations that already had mature reporting.
Figure 1
Mature reporting has not produced confident decision making
Selected survey and census evidence on decisions and analytics value
McKinsey survey, 2019
Gartner survey of 566 D&A leaders, published March 2023
When it is real, it spreads fast
Share of US manufacturing plants using data-driven decision making
Nearly tripled in five years. The change was organizational, not just technical.
The research on dashboards themselves points the same way. A 2021 systematic review of patient-safety dashboards in BMJ Health & Care Informatics screened 4,624 publications, kept 33, and concluded that while dashboard use had grown for 15 years, their impact remained poorly understood. Healthcare is one of the few fields that measures this at all; most enterprises never check whether a dashboard changed what anyone did.
INVISOM's argument is simple: the gap is not a tooling gap. BI was designed to make information available. It was never designed to make a specific choice better, faster or more consistent. Decision intelligence is the discipline that does that, and it starts by changing what you design.
What decision intelligence actually is
Two framings matter, and they agree more than they differ.
Gartner defines decision intelligence as a practical discipline for improving decision making by "explicitly understanding and engineering how decisions are made", and by evaluating, managing and improving outcomes through feedback. Three words in that definition do the work: explicit (the decision is written down, not implied by a report), engineering (it is designed, with inputs, logic and owners), and feedback (outcomes are measured and fed back).
Cassie Kozyrkov, Google's first Chief Decision Scientist, gave the discipline its most widely quoted practitioner framing. On the Google Cloud Platform Podcast in May 2018 she described it as data science "plus-plus": the applied side of data science, augmented with the behavioral and managerial sciences, so that data actually drives decisions rather than merely informing them. Her emphasis was on process: who the decision maker is, what they would do under each possible answer, and how a question gets from idea to a reliable, sometimes automated, decision.
Put together, the practical definition is this: decision intelligence treats a recurring business decision as a designed asset, with a named owner, explicit logic, the data and models it needs, the action it triggers, and a measured outcome that improves the next decision.
The market has followed. Gartner published its first Magic Quadrant for Decision Intelligence Platforms in January 2026, describing platforms that combine decision modeling, analytics and AI to augment and automate decisions. SD Times reported several Gartner predictions alongside it, including that by 2028 a quarter of chief data and analytics officers' vision statements will be "decision-centric" rather than "data-driven". Treat predictions as directional. The underlying shift, from data as the product to decisions as the product, is the useful part.
Why dashboards rarely change decisions
Dashboards fail to change decisions for structural reasons, not because they are badly built.
Figure 2
BI informs. Decision intelligence is designed to change what happens next.
How the two disciplines differ in practice
- Unit of designBI: Report or dashboardDI: A named, recurring decision
- Starting questionBI: What data do we have?DI: What choice must be made, by whom, how often?
- OutputBI: Information on a screenDI: A recommendation or action, within limits
- LogicBI: Implicit, in each reader's headDI: Explicit, versioned decision model
- MemoryBI: Metric historyDI: Decision taken, options rejected, outcome
- Success measureBI: Report views and adoptionDI: Change in the business outcome
They are designed around data, not choices. A sales dashboard shows revenue by region because revenue by region is available. It does not know that the regional director's real decision each Monday is whether to move field resources, and that the decision depends on pipeline coverage, not last month's revenue.
They leave the hardest step to the reader. Going from "this number is red" to "therefore I will do X" requires a threshold, a set of options and some knowledge of what each option tends to cause. A dashboard supplies none of these. Each reader improvises, so the same signal produces different actions, or none.
They have no memory of what happened next. BI records the metric. It does not record the decision taken, the alternatives rejected, or the outcome. Without that record an organization cannot learn which decisions were good, only which periods were lucky.
They measure consumption, not consequence. BI teams count report views. A report opened 500 times a month that changes nothing is a cost, not an asset.
Data-driven decision making does pay when it is real. Erik Brynjolfsson and Kristina McElheran found, using US Census data, that adoption of data-driven decision making in US manufacturing plants nearly tripled between 2005 and 2010, from 11% to 30%, and that the pattern is consistent with productivity gains. The lesson is not that data does not help. It is that adoption of data-driven decision making is an organizational change, and buying reporting tools is not the same thing.
Decision modeling: make the decision explicit
The core technique of decision intelligence is the decision model: a structured description of a decision, its inputs, its logic and its dependencies. There is a mature, vendor-neutral standard for this.
Decision Model and Notation (DMN) is maintained by the Object Management Group, the body behind BPMN and UML. Version 1.0 was adopted in September 2015; version 1.6 was formally adopted in September 2026, with 1.7 in beta. DMN provides two things: a graphical notation, the decision requirements diagram, that shows how a decision depends on sub-decisions, input data and business knowledge; and an expression language, FEEL, with decision tables and boxed expressions precise enough to execute. OMG explicitly designed the notation to be readable by business owners and analysts as well as developers, and to sit alongside BPMN process models.
Figure 3Illustrative
A decision model makes logic, data and ownership visible
Simplified DMN-style decision requirements view for a credit-limit increase
Input data
Sub-decision
Risk band
Uses: Bureau score, payment history
Knowledge: Risk policy table (owner: Credit Risk)
Sub-decision
Affordability
Uses: Income, current exposure
Knowledge: Affordability rule (owner: Compliance)
Decision · owner: Head of Cards
Increase this customer’s credit limit, and by how much?
Each box can be owned, tested and changed separately. A dashboard filter cannot.
You do not need a DMN engine to benefit from DMN thinking. The value for most organizations is the discipline it imposes:
- Name the decision as a question with a finite set of answers. "Should we increase this customer's credit limit, and by how much?" not "credit risk".
- Separate sub-decisions. Risk band and affordability are decided separately and combined. Each can be tested and owned.
- Distinguish data from knowledge. Bureau score is data. The policy table that maps score bands to limits is knowledge, owned by a policy holder, versioned and auditable.
- Make the logic inspectable. A decision table can be reviewed by risk, compliance and the business in a way that model code and dashboard filters cannot.
Where decisions are high-volume and rule-heavy, as in credit, claims, pricing and eligibility, the model can be executed directly in a decision engine. Where decisions are strategic and infrequent, the same structure works as a decision brief that is agreed before data is pulled.
Connecting BI to action, and the cost of skipping the model
Connecting analytics to action is where the value is, and where the risk sits. Automating a decision without an explicit model, limits and a feedback loop magnifies whatever is wrong with it.
Zillow is the clearest public case. Its iBuying business, Zillow Offers, used pricing models to make offers on homes at scale. On 2 November 2021 the company announced it would wind the business down, reported an inventory write-down of about $304 million because it had bought homes in the quarter at prices above its current estimates of future selling prices, and said the wind-down would cut its workforce by about 25%. Its chief executive said the unpredictability in forecasting home prices far exceeded what the company had anticipated.
The forecasting model was only one part of what failed. The decision system around it, meaning how much forecast uncertainty was tolerable per offer, what volume limits applied, and how quickly realized sale prices fed back into offer logic, was where the exposure accumulated. That is exactly the territory decision intelligence asks you to design explicitly: decision logic, guardrails, and outcome feedback, not only the prediction.
INVISOM's practical guidance is to think in three levels of connection, and to choose deliberately:
| Level | What the system does | Who decides | Good fit |
|---|---|---|---|
| Decision support | Presents relevant data and context for a named decision | A person, unaided by a recommendation | Strategic, infrequent, ambiguous decisions |
| Decision augmentation | Recommends an option, with reasons and confidence | A person, who can accept, change or reject | Frequent decisions with material risk or judgment |
| Decision automation | Executes the decision within explicit limits | The system, with human oversight of exceptions | High-volume, well-understood, reversible decisions |
Moving a decision up a level should require evidence from the level below: a recorded history of recommendations and outcomes, known error rates, and a defined stop condition. Our companion framework, Closing the Gap Between Dashboards and Decisions, sets out how to run that in practice.
AI-assisted analytics needs a semantic layer, and a decision to serve
Generative AI has made "ask your data a question" the default promise of every BI vendor. It is a real improvement in access. It does not, on its own, improve decisions, and it fails badly without governed business definitions underneath it.
The research is consistent on the second point. In a 2023 benchmark by Juan Sequeda, Dean Allemang and Bryon Jacob at data.world, GPT-4 answered enterprise insurance questions correctly 16% of the time when it queried the SQL database directly, and 54% of the time when the same questions were posed over a knowledge-graph representation of the business semantics. Spider 2.0, an enterprise text-to-SQL benchmark presented at ICLR 2025, found that the best model at the time solved 21.3% of its 632 real-world tasks, against 91.2% on the original Spider benchmark and 73.0% on BIRD. Model scores have risen since, but the gap between clean academic schemas and real enterprise schemas has not disappeared.
Figure 4
AI answers over enterprise data depend on the business semantics underneath
Accuracy of language models answering questions over databases, in two published benchmarks
Benchmark 1 · Same model, same questions
Benchmark 2 · Best model, by benchmark realism
The industry response is to move business meaning into a governed semantic layer: metrics, dimensions, relationships, synonyms and descriptions defined once and consumed by every tool and every AI assistant. Three developments show the direction:
- Open Semantic Interchange (OSI). Snowflake announced on 27 January 2026 that the first OSI specifications were finalized under the Apache 2.0 licence, with participants including Databricks, dbt Labs, Salesforce, ThoughtSpot, Qlik, Sigma and others.
- dbt MetricFlow. dbt Labs moved its metric engine to an Apache 2.0 licence in October 2025, with the stated aim of deterministic metric definitions for AI systems.
- Power BI "Prep data for AI". Microsoft's guidance for Copilot asks model owners to add AI data schemas, verified answers and AI instructions to the semantic model, and to mark a model "Approved for Copilot" before broad use. Microsoft also notes that this preparation cannot guarantee a specific output every time.
For Gulf organizations, the semantic layer is also where Arabic belongs. Synonyms, descriptions and verified answers in both Arabic and English, maintained under the same governance as the metric definitions, are what make bilingual natural-language analytics dependable. That is a design task, not a model feature. For the platform-level trade-offs, see Snowflake, Databricks, warehouses and marts.
The decision-intelligence point is the first one, though. A semantic layer makes answers consistent. Only a decision model makes them consequential. An AI assistant that answers "what was churn in Riyadh last quarter?" correctly is useful. One that is wired into the retention decision, knows the threshold that triggers an intervention, and records whether the intervention worked is decision intelligence.
A practical test: is this BI or decision intelligence?
Use these questions on any analytics product, existing or proposed. Each "no" is a gap.
| Question | BI answer | Decision intelligence answer |
|---|---|---|
| Which decision does this serve? | "Sales performance" | "Weekly reallocation of field sales capacity across territories" |
| Who owns the decision? | The report owner | A named business role, with decision rights |
| What options exist? | Not stated | Three defined actions, plus "do nothing" |
| What triggers action? | A colour change | A documented threshold with a response playbook |
| How is logic captured? | In filters and the reader's head | In a decision table or model, versioned |
| What is recorded? | Metric values | The decision taken, the alternatives, the outcome |
| How is success measured? | Usage of the report | Change in the business outcome the decision affects |
If most of your portfolio sits in the left column, the next investment should not be another BI tool or an AI assistant. It should be a decision inventory.
What to do in the next 90 days
- List the 20 to 30 recurring decisions that drive most of your value or risk, written as questions with finite answers. Start in pricing, credit, inventory, workforce and service operations, where decisions repeat.
- Name an owner for each decision, not for each dashboard, and agree their decision rights.
- Model three decisions explicitly, using DMN decision requirements diagrams and decision tables, even on a whiteboard. Have business, risk and data teams review the logic together.
- Map every existing report to a decision. Reports that serve no named decision become candidates for retirement.
- Move the metrics those decisions depend on into a governed semantic layer, with definitions, synonyms and descriptions that an AI assistant can use, in Arabic as well as English where relevant.
- Start recording decisions and outcomes for the three modeled decisions, so the next quarter's review is based on evidence rather than recollection.
- Set an explicit level for each decision: support, augmentation or automation, and the evidence needed to move it up a level.
The value case discipline in From AI Experimentation to Enterprise Value applies directly: a decision-intelligence initiative should name the decision, the baseline and the outcome it expects to move.
Frequently asked questions
What is the difference between business intelligence and decision intelligence?
Business intelligence makes information available, usually through reports and dashboards. Decision intelligence designs specific decisions: their owner, inputs, logic, triggering thresholds, the action taken and the measured outcome, so that decisions can be improved through feedback.
How does Gartner define decision intelligence?
Gartner describes it as a practical discipline that improves decision making by explicitly understanding and engineering how decisions are made, and how outcomes are evaluated, managed and improved through feedback.
What is DMN and do we need it?
Decision Model and Notation is an Object Management Group standard for modeling decisions, with a diagram notation and an executable expression language including decision tables. You do not need a DMN engine to start, but its structure is a strong way to make decisions explicit and reviewable.
Will AI assistants for analytics replace decision intelligence?
No. Natural-language analytics improves access to data and depends on a governed semantic layer for accuracy. Decision intelligence adds what assistants lack: a defined decision, thresholds, options, ownership and outcome feedback.
Sources and further reading
- Decision Intelligence, Gartner IT Glossary.
- Episode 128: Decision Intelligence with Cassie Kozyrkov, Google Cloud Platform Podcast, May 2018.
- Decision Model and Notation (DMN), Object Management Group. Version 1.6 adopted September 2026.
- Gartner Survey Reveals Less Than Half of Data and Analytics Teams Effectively Provide Value to the Organization, Gartner, 21 March 2023. Survey of 566 data and analytics leaders, September to November 2022.
- Decision making in the age of urgency, McKinsey & Company, 2019.
- The Rapid Adoption of Data-Driven Decision-Making, Brynjolfsson and McElheran, American Economic Review, May 2016.
- Dashboards for visual display of patient safety data: a systematic review, Murphy et al., BMJ Health & Care Informatics, 2021.
- Zillow Group Q3 2021 results and plan to wind down Zillow Offers, Zillow Group, Form 8-K Exhibit 99.1, 2 November 2021.
- Gartner acknowledges growth of decision intelligence platforms with inaugural Magic Quadrant, SD Times, 3 March 2026.
- A Benchmark to Understand the Role of Knowledge Graphs on LLM Accuracy for Enterprise SQL, Sequeda, Allemang and Jacob, arXiv, November 2023.
- Spider 2.0: Evaluating Language Models on Real-World Enterprise Text-to-SQL Workflows, Lei et al., arXiv, November 2024; ICLR 2025.
- Open Semantic Interchange specifications finalized, Snowflake, 27 January 2026 (vendor material).
- Open-sourcing MetricFlow, dbt Labs, 14 October 2025 (vendor material).
- Prepare your data for AI to improve Copilot results, Microsoft Learn.