Solutions · Analysis and decision
Decision intelligence
Preparing decisions with signals, scenarios, constraints and visible consequences.
On one page
Models and systems for classifying, prioritising, anticipating, comparing scenarios and providing evidence in concrete, evaluable situations.
- 01Models that help to decide without pretending to replace the decision. We build systems for classification, prioritisation, prediction and scenario comparison around concrete decisions.
- 02From prediction to decision. The model estimates; the system decides how to use that estimate.
- 03Optimisation. When several alternatives and constraints exist, anticipating what will happen is not enough.
Models that help to decide without pretending to replace the decision.
We build systems for classification, prioritisation, prediction and scenario comparison around concrete decisions.
Before selecting a method we represent the situation the organisation wants to improve:
From prediction to decision
The model estimates; the system decides how to use that estimate. It can order cases, compare scenarios, propose an action, trigger a review or abstain. That translation needs objectives, error costs, constraints and a subsequent signal with which to evaluate the result.
Optimisation
When several alternatives and constraints exist, anticipating what will happen is not enough. It may also be necessary to choose an allocation, a sequence or a trade-off between objectives. We integrate prediction and optimisation when both functions are part of the same decision.
A useful decision does not end when a score is produced. It needs a threshold, a policy, a person or a process that knows what to do with it, and a subsequent signal that makes it possible to learn whether the result improved.
Starting from the decision taken today
We ask what decision is taken today, who takes it, what alternatives exist, what information is known at that moment, what information arrives afterwards, what objective is to be improved, what constraints must be respected, what cost each type of error carries, what capacity the process has to act on the recommendation and what result will make learning possible.
This representation prevents the project from optimising a technical metric disconnected from the use. A more accurate prediction may turn out to be less useful if it arrives late, if it demands data that does not yet exist, if it produces too many cases to review or if it concentrates the errors in high-impact situations.
Baselines that represent the current work
The main comparison is the way the organisation decides today: a rule, a threshold, a heuristic, a historical average, a model already deployed, a team’s judgement or an allocation policy.
The new system is compared under the same conditions of information, time and capacity. A technical improvement acquires value when it translates into a decision that is better prepared, more consistent, faster, cheaper or safer.
Uncertainty and calibration
Uncertainty is part of the output. We work so that a probability of seventy per cent has a consistent empirical meaning within the scope evaluated, and we observe calibration by segment, horizon and period.
Uncertainty can be used to adjust thresholds, hand over cases, request information, choose another model, show an interval, narrow the scope, abstain or ask for human judgement.
Policies, human capacity and the subsequent result
The policy defines which action corresponds to each range, which constraints prevail, which cases need review, what capacity exists, what happens when data is missing, when it is updated, what is recorded and how it is reverted. It stays examinable and versioned: a change in the model does not silently modify the authority or the path.
After deployment we observe predictive quality, calibration, drift, latency, cost, human decisions, corrections, rate of use and operational effect. Optimisation makes objectives and constraints explicit and makes it possible to compare the consequences of different priorities.
This solution applies machine learning, optimisation and decision models within operational intelligence: the signal enters a path that preserves state, authority and verification.
What we build
- 01DefineWe represent the decision, its alternatives, its constraints and its time horizon.
- 02LearnWe use machine learning, statistics or rules according to the type of information and the nature of the problem.
- 03CompareWe evaluate alternatives, scenarios, uncertainty and sensitivity.
- 04IntegrateWe embed the model within the process where its results can be used.
- 05ObserveWe check how it behaves with new data and how it affects the real decision.
Usual capabilities
Classification, ranking, forecasting, anomaly detection, scoring, recommendation, optimisation and scenario analysis.
We choose rules, statistical models, machine learning or optimisation according to the nature of the data, the stability of the phenomenon, the cost of each error and the way the decision will be able to be evaluated.
- 01which decision needs to improve?
- 02what information is available when it is taken?
- 03what does each error cost?
- 04which part requires human judgement?
- 05how will we know whether the system really improves the result?
The system from the inside
Maturity · Operational- How it is organised
- We first represent the decision: alternatives, information actually available at the moment of deciding and cost of each type of error. On that representation we train tabular models — usually gradient boosting — or apply rules when the condition is explicit. The output enters the process as a signal accompanied by its uncertainty, not as an order.
- What can be examined
- We can show the evaluation notebook over public or synthetic data, with calibration curves, cost matrix and variable importance. Also the diagram of where the prediction enters the flow and who decides after it.
- How it is checked
- We evaluate with temporal validation so as not to leak future information, and we always compare against the rule or the criterion applied today, not against chance. After going live we track calibration and drift with the new data.
Conditions and limits
A model trained on past decisions inherits their biases, so we keep human review in the highest-cost cases. When the phenomenon changes regime, performance falls before the metrics signal it. The system describes associations: it does not produce causal explanations.