October 21, 2026Data Analysis
Analytics maturity models are often presented as a ladder: every organisation should climb to the top. In practice, the levels answer different questions, and most organisations need several of them at once. The useful question is not "what is the highest level?" but "what decision are we trying to improve, and what level does it require?"
1. Descriptive analytics — what happened?
Descriptive analytics summarises the past: sales by month, churn by plan, defects by line, wait times by hour. It is the foundation, and it is where most reporting still lives.
The trap is volume without purpose. A dashboard with 40 charts is not maturity; it is a filing cabinet. Good descriptive work answers a specific operational question and makes the definition of each metric unambiguous.
2. Diagnostic analytics — why did it happen?
Diagnostic analysis looks for drivers: why did churn rise in one segment, why did a campaign underperform, why did defects spike on a particular shift? Techniques include segmentation, drill-downs, correlation analysis, cohort comparisons and root-cause investigation.
This is where statistical discipline matters. A plausible story is not a cause. Distinguishing a real driver from a coincidence usually requires careful comparison groups, time-series thinking or an experiment.
3. Predictive analytics — what is likely to happen?
Predictive models estimate future outcomes: which customers will churn, which accounts will default, which machines will fail, which patients are at risk of readmission. Techniques range from logistic regression and survival models to gradient boosting and neural networks.
The output is a probability, not a decision. A model becomes useful when it is calibrated, explained and connected to an action and a threshold. A high AUC that nobody operationalises is a research artefact.
4. Prescriptive analytics — what should we do?
Prescriptive analytics recommends an action: which offer to make, how much inventory to hold, which maintenance window to schedule, how to route a service call. It combines prediction with constraints, costs and objectives — often through optimisation, simulation or decision analysis.
This level is harder because it requires the business to state its objective explicitly. "Maximise retention" and "maximise profit" can lead to different policies. Prescriptive work also needs a feedback loop: the recommended action changes the system, which changes the data the next model sees.
5. Generative and agentic analytics — what can be created or automated?
The newest layer uses large language models and agents to generate content, interface with data and execute multi-step workflows. Examples include natural-language querying, cited answers over internal documents, automated report drafting, document extraction and routing, and assistants that call tools to complete a task.
Generative analytics is powerful, but it does not replace levels 1–4. It changes how users interact with them and how much routine work can be automated. An assistant that answers from a dashboard still depends on the dashboard's definitions. An agent that triggers a workflow still needs permissions, evaluation and an audit trail.
Where it fits
| Level | Core question | Typical output | AI's role |
|---|---|---|---|
| Descriptive | What happened? | Reports, dashboards | Natural-language querying, summarisation |
| Diagnostic | Why? | Driver analysis, segments | Faster exploration, anomaly explanation |
| Predictive | What next? | Scores, forecasts | Model selection, feature engineering, AutoML |
| Prescriptive | What should we do? | Recommendations, policies | Simulation support, optimisation assistance |
| Generative | What can be created or automated? | Assistants, generated content, workflows | The level itself |
Most organisations are stuck at level one — for a reason
It is common to find sophisticated models sitting next to unreliable reporting. The cause is usually not a lack of algorithms. It is weak foundations: inconsistent metric definitions, unclear data ownership, manual pipelines and no agreement on which decisions matter.
Moving up requires more than tooling:
- A named decision owner for each analytics product.
- Governed definitions of the core metrics, documented and enforced.
- Data quality and lineage that people trust enough to act on.
- A measurement loop so recommendations can be evaluated after the fact.
- Governance for models and AI, including monitoring and human review.
Assess yourself honestly
Pick three decisions your organisation makes every week. For each, ask: do we have the data to describe what happened, the analysis to explain why, a model to anticipate what comes next, a policy for what to do, and a way to automate the routine part? The gaps tell you more than a maturity score.
The goal is not to reach level five everywhere. It is to match the level to the decision, build the foundations once, and use AI where it removes work without removing accountability.
