presentsUNAIVERSE
The UN AI timeline

    Learn AI /AI Foundations /Predictive AI

    Predictive AI

    Also called: predictive analytics.

    Detail

    In five seconds

    What it is

    A system that uses past data to estimate what is likely to happen next, such as a future value, category or level of risk.

    In plain words

    Predictive AI looks at past examples and uses them to estimate what may happen next.

    An everyday example

    A bus station manager in Addis Ababa looks at records from the last six Monday mornings. The records show rain, school holidays and how many passengers arrived between 7:00 and 8:00. For next Monday, the manager sees that rain is expected and schools are open, so she estimates that about 420 passengers will arrive. She then assigns an extra bus before the rush begins.

    Where the story stops working

    A human manager may use local knowledge that was never written down, such as a road closure or a festival. Many predictive systems only use the variables they were given and may miss important context. Also, real models can process far more examples and variables than a person can, so the analogy helps with the idea of prediction but not with the scale or mathematics.

    Remember thisPredictive AI learns relationships between past inputs and known outcomes, then estimates a value, category or probability for new inputs. The estimate is evidence for a decision, not advance knowledge of the result.

    This term is contested

    The boundary is somewhat unsettled. Some people use 'predictive analytics' for statistical forecasting without calling it AI, while others include many forecasting and scoring systems under AI. The core idea of estimating future or unknown outcomes from data is more stable than the label.

    Understand it

    Understand it

    Why it matters

    Many important choices depend on expectations about the near future: stock levels, staffing, maintenance, fraud checks, disease surveillance and transport planning. Predictive AI can help people spot patterns too large or complex to review by hand. That can improve speed and consistency. But a prediction can also mislead. If the training data is incomplete, biased or outdated, the output may be systematically wrong for some groups or situations. If people treat a score as a fact, they may make unfair or unsafe decisions. For a non-specialist, the key question is not whether the system sounds advanced. It is whether the prediction is fit for purpose, checked against reality, and used with appropriate human oversight.

    How predictive AI turns past records into a forecast

    After viewing this, a reader should be able to say what goes into predictive AI, what the model does, and what comes out.

    1. 01

      1. Past records

      The system starts with historical data that includes both input details and the outcome that later happened.

      Draw a box labelled 'Past records' with mini rows such as 'rain, weekday, 390 passengers' and 'dry, holiday, 240 passengers'.

    2. 02

      2. Build the model

      A training process finds patterns linking the inputs to the known outcomes.

      Draw a box labelled 'Build model' with a simple icon of dots connected by lines or a formula symbol. Add a caption: 'finds useful patterns'.

    3. 03

      3. New case arrives

      The model receives fresh input data for a case whose outcome is not yet known.

      Draw a box labelled 'New input' showing example fields like 'rain expected, schools open, Monday'.

    4. 04

      4. Prediction

      The model returns an estimate such as a number, category or probability that can support planning or review.

      Draw a box labelled 'Predicted result' with examples: '420 passengers' and 'high risk of delay'. Use a note: 'estimate, not certainty'.

    A left-to-right flow diagram with four boxes. Read from past records, to model building, to new case input, to predicted result. Small labels under each box show examples of inputs and outputs. Arrows show that historical data is used to build the model, while new data is used later to get a prediction.

    Hover or tap each box to reveal one sentence with an example input or output; arrows highlight in sequence from left to right.

    Describe this without the diagram

    This diagram shows predictive AI as a four-step flow. First, past records contain earlier situations and the outcomes that followed. Second, a model is built by finding patterns in those records. Third, new input data for a current case is entered. Fourth, the model produces a predicted result, such as a number or a risk category. The main idea is that the system learns from past examples and then applies that pattern to a new case. The prediction is presented as an estimate, not a certainty.

    With reduced motion: A static four-box flow is shown from left to right: Past records, Build model, New input, Predicted result. Each box includes one short example and a note that the output is an estimate, not a certainty.

    Predictive AI uses examples from the past to estimate a future outcome. The input is usually data: recorded facts such as dates, amounts, locations, sensor readings or past choices. The output is a prediction, such as a number, a category or a probability.

    A common process has two stages. First, a model is built from historical data. A model is a mathematical pattern learned from examples. Then the model receives new data and produces an estimate, such as “high risk of delay” or “likely demand next week: 2,000 units”.

    This is different from generative AI. Generative AI produces new content such as text or images. Predictive AI mainly scores, classifies or forecasts. Predictive systems may use machine learning or statistical methods. Fixed if-then rules alone form a rule-based system.

    The prediction is not a fact and not a guarantee. It depends on the quality of the data, the design of the model and whether the future still resembles the past. Human review is often needed, especially when mistakes could affect people’s rights, safety or access to services.

    How “Estimating bus crowding from past mornings” maps onto the real thing

    In the storyIn the AI system
    Past Monday records with weather and school statusHistorical data used to find patterns
    The manager comparing today’s conditions with earlier morningsA model applying learned relationships to new input data
    Estimate of 420 passengersPredicted output such as a number or risk score
    Assigning an extra busOperational action taken using the prediction

    A worked example

    An information management officer in a regional bureau is finalising the weekly workplan and needs to forecast the volume of support requests likely to reach the reporting portal helpdesk.

    Starting from: Approved historical records of daily request totals, day of the week, reporting deadlines, scheduled portal releases and recent service interruptions.

    1. Past records are paired with the known number of support requests received on each day.
    2. A forecasting model is fitted to learn how request volumes have varied with deadlines, releases, weekdays and interruptions.
    3. For the coming week, the officer enters the current reporting calendar, release schedule and service status.
    4. The model returns an estimated request volume for each day. The officer checks whether recent workflow changes make the historical pattern less reliable.
    5. The forecast informs the duty roster, while the officer retains contingency cover because the estimates are not certainties.

    Result: The workplan places additional helpdesk coverage on the days expected to be busiest. Actual request volumes are then compared with the forecast so that errors can be monitored and the model reviewed.

    In a United Nations workplace

    An ICT officer in a Resident Coordinator's Office drafts the note for the file after a bilateral with a vendor. The vendor's tool claims to predict whether inputs to the next UN Country Team progress report will miss clearance, using past submission times, revision counts and the interval between first draft and deadline.

    The model maps those historical features to an outcome, such as on-time or late, and returns a probability for each new input. The office could use the estimate to schedule reminders or additional editorial support, rather than waiting for the clearance log to become self-explanatory at close of business.

    Before you try thisDo not upload draft reports, personal performance information or other non-public material to a tool that has not been approved for it. A high-risk score must not be treated as a judgement about a colleague or office. Reporting arrangements change, and historical delays may reflect circumstances the model cannot see.

    Where you may hear it

    • At an ICT governance board review of a forecasting pilot
    • During a vendor demo in a procurement meeting
    • In an evaluation office discussion of model accuracy
    • While clearing a workplan that uses risk scores to prioritise follow-up

    Avoid confusion

    Avoid confusion

    Often confused with

    Side-by-side

    Common misconceptions

    • Predictive AI tells the future.

      It does not know the future. It estimates what is likely based on patterns in past data and current inputs. Unexpected events, missing variables or changed conditions can make the prediction wrong.

    • If a model is accurate overall, it is reliable for every group and every case.

      An average accuracy score can hide weaker performance for certain regions, languages, seasons or population groups. Systems should be tested on the real contexts where they will be used, not only on a single headline metric.

    • Predictive AI always means advanced machine learning.

      Some predictive systems use machine learning, but others use statistics or even rule-based methods. The important feature is that they estimate an outcome; the underlying method can differ.

    • A risk score makes the decision objective.

      A score can support a decision, but it reflects choices about data, labels, thresholds and goals. Human review is still needed, especially when consequences are serious.

    Quick check

    Which example best fits predictive AI?

    Go deeper

    Go deeper

    The simple version

    Predictive AI looks at past examples and uses them to estimate what may happen next.

    The technical version

    Predictive AI maps input variables, often called features, to an expected target outcome using patterns derived from historical data. During development, a method such as regression, classification, time-series forecasting or another supervised learning approach is fitted on examples where the outcome is already known. After deployment, the trained model receives new feature values and returns an estimate such as a continuous value, class label or probability score. Depending on the use case, the method may be a statistical model, a machine-learning model or a hybrid with rule-based components. Performance depends on data quality, representativeness, feature selection, model choice and whether the deployment environment matches the conditions seen during training.

    Reviewed external resources

    Where this came from

    Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.

    The one thing to remember

    Predictive AI learns relationships between past inputs and known outcomes, then estimates a value, category or probability for new inputs. The estimate is evidence for a decision, not advance knowledge of the result.