How the organisation's data supports predictive analytics, AI, and machine learning, including transparency, bias mitigation, and model governance.
INT-CON-01
Access to predictive insight
To what extent do you have access to, and actually use, predictive or forward-looking analysis — forecasts, propensity scores, risk indicators, anomaly alerts — rather than only historical reporting?
Maturity level descriptions
Only descriptive, historical reporting is available; the Data Consumer has no access to predictive or forward-looking analysis. Decisions about future activity are made from past-period reports and professional judgement alone.
Predictive analysis has been produced occasionally as a one-off exercise, not made routinely available or maintained. The Data Consumer recalls a forecast or model produced once for a specific project, which was not repeated or sustained.
Defined predictive outputs are produced and made available to relevant business users on a regular basis, with their purpose explained. The Data Consumer regularly receives forecasts or risk indicators relevant to their work and understands what they are for.
Use and value of predictive outputs are measured, accuracy is reported to business users, and outputs are refined based on business feedback. The Data Consumer sees how accurate previous predictions proved to be, and their feedback influences how outputs are produced.
Predictive insight is embedded in business processes at the point of decision and continuously improved, with business value measured as an outcome. The Data Consumer encounters predictive insight within their normal workflow rather than as a separate report, and its business value is measured.
INT-CON-02
Decision Intelligence in the workflow
To what extent are recommendations, prioritisations or automated decisions generated by analytical or AI systems part of your everyday workflow, and how clearly is their role defined?
Maturity level descriptions
No analytical or AI-generated recommendations feature in the Data Consumer’s work; all decisions are made unaided. Work is prioritised and decided entirely on individual judgement, with no system-generated support.
Recommendations appear occasionally through pilots or trials, with no defined role and unclear status. The Data Consumer encounters a system-generated recommendation without knowing whether it is advisory, authoritative or experimental.
Decision support is deployed with a defined role — advisory or determinative — documented and communicated to the business users who receive it. The Data Consumer knows whether a recommendation is advisory or binding, and what they are expected to do with it.
The effect of decision support on business outcomes is measured, override rates are tracked, and the scope of automation is reviewed on a defined cycle. The Data Consumer’s acceptance and override of recommendations is measured and reviewed, informing where automation is appropriate.
Decision Intelligence is continuously optimised against measured business outcomes, with the balance of human and automated decision-making actively and transparently managed. The Data Consumer works within a decision model whose human/automated boundary is deliberately set, measured and adjusted over time.
INT-CON-03
Explainability of model outputs
When an analytical or AI system produces a score, recommendation or classification that affects your work, how well can you understand why it reached that result?
Maturity level descriptions
No explanation is available; the Data Consumer receives an output with no indication of what drove it. A score or recommendation is presented as a bare number or label, with no accessible reasoning.
Explanations can sometimes be obtained by asking a technical specialist, and are expressed in technical rather than business terms. The Data Consumer must approach an analyst for an explanation, which is difficult to relate to business circumstances.
Business-meaningful explanation is provided with model outputs as a defined standard — the main contributing factors expressed in terms the Data Consumer understands. The Data Consumer sees the principal factors behind a score alongside the score itself, in business language.
Explanation quality is measured, business-user comprehension is assessed, and explanations are improved where they are not understood. The Data Consumer’s understanding of explanations is tested and acted upon, rather than assumed.
Explanation is interactive and case-specific, allowing the Data Consumer to explore why an individual result occurred and what would change it. The Data Consumer can examine an individual case, see the factors that determined the outcome, and explore what would alter it.
INT-CON-04
Appropriate reliance and override
How clearly do you understand when a model output should be relied upon and when your own judgement should take precedence — and how straightforward is it to record an override?
Maturity level descriptions
No guidance exists on when to rely on or override a model output; the Data Consumer decides case by case with no framework. Reliance on model outputs varies entirely by individual disposition, with no guidance and no record of overrides.
Informal advice circulates about when outputs are unreliable, based on anecdote rather than analysis, and overrides are unrecorded. The Data Consumer follows colleagues’ informal views about when a model “gets it wrong”, without evidence or record.
Documented guidance describes the intended use, known limitations and override expectations for models used by business users, and overrides are recorded. The Data Consumer can consult a statement of a model’s intended use and limitations, and records their reason when overriding it.
Override rates and reasons are measured, reviewed against model performance, and used to determine where reliance is appropriate. The Data Consumer’s overrides are analysed alongside outcomes, and reliance guidance is adjusted on the evidence.
Reliance is continuously calibrated — models express confidence, guidance adapts to measured performance, and overrides feed model improvement directly. The Data Consumer sees a confidence indication with each output and knows that their overrides feed back into model refinement.
INT-CON-05
Awareness of approved AI capability
How clearly do you know which AI and GenAI capabilities you are permitted to use in your work, and what each is approved for?
Maturity level descriptions
No information exists about which AI capabilities are approved; the Data Consumer uses or avoids AI tools entirely on personal judgement. Business users adopt AI tools independently, with no organisational position on what is permitted.
General messages have been issued about AI, without specifying which tools are approved or for what purposes. The Data Consumer has heard broad encouragement or caution about AI, without a list of approved tools or uses.
A documented register of approved AI capabilities and their approved uses is available and communicated to business users. The Data Consumer can consult a list of approved AI tools and what each may be used for.
The register is maintained on a defined cycle, adoption and unapproved use are monitored, and business users are supported in using approved capability well. The Data Consumer sees a current register, receives support in using approved tools, and unapproved use is detected and addressed.
Approved capability is surfaced in context and continuously updated as tools change, with business users informed before new capability is enabled. The Data Consumer sees approved capability presented within their working environment and is informed ahead of changes.
INT-CON-06
Awareness of bias and fairness
For model or AI outputs that affect people — clients, staff or citizens — how well are you informed about known bias and fairness limitations you should take into account?
Maturity level descriptions
No information about bias or fairness limitations reaches business users; the Data Consumer acts on outputs unaware of any such consideration. Outputs affecting people are acted upon with no knowledge of whether the model performs differently across groups.
Bias is acknowledged in general organisational messaging, without anything specific to the models a Data Consumer actually uses. The Data Consumer is aware that AI can be biased in principle but knows nothing about the specific models they act on.
Known bias and fairness limitations are documented for models affecting people and communicated to the business users who act on their outputs. The Data Consumer can consult a statement of a model’s known limitations across groups and what to be alert to.
Fairness is monitored in production, results are reported to business users, and concerns raised by them are investigated and tracked. The Data Consumer receives periodic fairness monitoring results and can raise a fairness concern that is formally investigated.
Fairness is continuously monitored with automated detection, business-user observations are a recognised signal, and findings drive model or process change. The Data Consumer’s observations of unfair outcomes are treated as a formal monitoring input, with demonstrable resulting change.
INT-CON-07
Feedback on model performance
When a model output turns out to be wrong or unhelpful in practice, how effectively does that experience reach the people responsible for the model?
Maturity level descriptions
No route exists for reporting model errors; the Data Consumer simply disregards outputs they find unreliable. Business users stop using a model without anyone responsible learning that it is failing.
Problems are mentioned informally to whoever is nearby, with no record and no expectation of action. The Data Consumer mentions a wrong recommendation to a colleague, and nothing is recorded or actioned.
A defined route exists for business users to report model performance problems, with a named responder and documented process. The Data Consumer reports a wrong output through a named channel and it reaches the model owner.
Reports are logged, tracked to resolution and analysed alongside monitoring data, with outcomes communicated back to the reporter. The Data Consumer receives an outcome on reported problems, and their reports are analysed with technical monitoring data.
Business-user feedback is a standing input to model monitoring and retraining decisions, with observable improvement resulting. The Data Consumer sees model behaviour improve as a result of feedback they and colleagues provided, within a defined retraining cycle.