Predictive ML
Models trained on your own history that score risk, rank work or forecast demand, so people can act earlier.
What it is
Predictive ML learns patterns from your organisation's own history and uses them to score new cases. It answers narrow, repeated questions: how likely is this patient to miss the appointment, which machine is likely to fail next, how many units will this store sell next week, which payments look unusual. It usually isn't generative AI at all. Gradient-boosted trees and classical time-series models still do most of this work well.
It works as a pipeline. Historical data is cleaned and turned into features, such as 'missed appointments in the last year' or 'days since last service'. A model is trained on part of the history and tested on a later period it hasn't seen. The approved model is versioned and deployed behind an endpoint or run as a batch job, and every score is stored with the model version that produced it.
In an enterprise, the score is an input to a process, not a decision. It reorders a worklist, triggers a reminder or adds a case to a review queue. The hard parts are getting reliable data out of operational systems, agreeing what a good outcome is, and noticing when the world changes and the model's picture of it goes out of date.
Use it when, and think twice when
Use it when
- You have a few years of reliable history with the outcome recorded
- The question is narrow and repeated thousands of times
- Acting earlier on a score has a clear benefit
- You can find out later whether each score was right
Think twice when
- The outcome is rare, poorly recorded or defined differently by each team
- A simple rule or report would get most of the benefit
- The score could disadvantage a protected group and you can't test for it
- Nobody will retrain or monitor the model after launch
Reference architecture
01Channels
- Transactions and events
- Scheduled batch runs
- External data feeds
02Intake
- AWS GlueData Factory in Microsoft FabricApache Airflow, or your existing ETLData extracts and pipelines
- Amazon S3Azure Data Lake StorageCeph, or IONOS S3 Object StorageData lake
03Intelligence
- SageMaker Feature StoreAzure Machine Learning managed feature storeFeast on KubernetesFeatures
- Amazon SageMaker AIAzure Machine LearningMLflow and Kubeflow on KubernetesTrain and register models
- SageMaker AI endpoint or batch transformAzure Machine Learning online or batch endpointKServe, or a batch job on KubernetesScore: real time or batch
04Orchestration
- SageMaker AI PipelinesAzure Machine Learning pipelinesKubeflow PipelinesML pipelines and retraining
- Amazon EventBridgeEvent GridApache KafkaScores out as events
05Systems of record
- ERP, EPR or core system
- Data warehouse
06People
- Person
Act on the ranked worklist - Person
Model owner reviews drift and fairness
- Person
- Cloud or platform service (AWSAzuresovereign / on-prem)
- AI agent
- A person decides
- Your existing systems
- Outside your organisation
Components, side by side
| Layer | AWS | Azure | Sovereign / on-prem |
|---|---|---|---|
| Data extractionUsually most of the effort. Agree data definitions with the business before modelling. | AWS Glue | Data Factory in Microsoft Fabric | Apache Airflow, or your existing ETL |
| Data lake and warehouse | Amazon S3, Amazon Redshift | Azure Data Lake Storage, Microsoft Fabric | Ceph or IONOS S3 Object Storage; PostgreSQL or your warehouse |
| Feature storeThe same feature definitions in training and scoring. | SageMaker Feature Store | Azure Machine Learning managed feature store | Feast |
| Training and model registryBoth clouds support MLflow tracking, which keeps models portable. | Amazon SageMaker AI | Azure Machine Learning | MLflow and Kubeflow on Kubernetes |
| ScoringBatch is cheaper and simpler unless you need the score in seconds. | SageMaker AI real-time endpoint or batch transform | Azure Machine Learning online or batch endpoint | KServe, or a scheduled job on Kubernetes |
| Pipelines and retraining | SageMaker AI Pipelines | Azure Machine Learning pipelines | Kubeflow Pipelines |
| Drift and fairness checksSageMaker Model Monitor and Clarify closed to new customers on 30 July 2026. | Custom checks in SageMaker AI Pipelines, using open-source libraries (e.g. Evidently, Fairlearn) | Azure Machine Learning model monitoring and the Responsible AI dashboard | Open-source libraries (e.g. Evidently, Fairlearn) in Kubeflow Pipelines |
| ExplanationsGive the person acting on a score its top reasons, in plain words. | Open-source SHAP | Responsible AI dashboard, or open-source SHAP | Open-source SHAP |
| Monitoring and auditStore every score with the model version that produced it. | CloudWatch, CloudTrail | Azure Monitor, Log Analytics | Prometheus, Grafana, OpenTelemetry |
On a phone, the table shows the platform you pick. Sovereign / on-prem means open, portable components you can run in your own data centre or with a UK or European provider; IONOS is shown as one example, and others exist. What sovereign cloud means →
Operations
Keeping it reliable
AI fails quietly. These three habits make failure visible before customers see it.
Evaluation
Test on a later period
Train on older data and test on the most recent months, the way the model will be used. Compare it with the rule you use today, not with nothing. Report accuracy by customer group and site, not only overall.
Monitoring
Watch the inputs and the outcomes
Track whether incoming data still looks like the training data, and compare scores with real outcomes as they arrive. Set thresholds that trigger a review, and retrain on a schedule as well as on alert.
Cost
Score in batch where you can
Most predictive models are cheap to run. Training and always-on endpoints are where costs grow, so score in batch unless you need an answer in seconds, and switch off idle endpoints. The biggest cost is usually the data engineering, so budget for it.
Security and failure risks
Model drift
Behaviour changes after a price change, a new process or a shock to demand, and the model keeps scoring confidently on a world that no longer exists.
How the design avoids it
Drift checks on inputs and outcomes, alert thresholds with a named owner, scheduled retraining, and a fallback to the previous rule.
Unfair outcomes through proxies
A postcode or buying pattern stands in for a protected characteristic, and one group is consistently scored worse.
How the design avoids it
Test outcomes by group before launch and every month after, remove or constrain proxy features, and keep a person on every adverse decision.
Leakage in training
A feature quietly contains the answer, such as a field filled in after the outcome, so test results look excellent and live results don't.
How the design avoids it
Build features only from data available at the moment of scoring, test on a later period, and treat suspiciously good results as a bug.
Data poisoning and probing
Someone alters training data to skew the model, or probes a scoring endpoint to learn how to beat it.
How the design avoids it
Control who can write to training data, version datasets, keep endpoints private behind the API layer, and rate-limit and log every call.
Playbooks that use this pattern
All 33 playbooks using Predictive ML →Payment fraud and APP scam detection
Risky payments scored in milliseconds, only the highest-risk band held automatically under rules a person approved, every other alert reviewed by an analyst before any hold, and people making every decision about a customer's payment or claim.
- Predictive ML
- Agents with tools
- Human in the loop
- Event-driven integration
Reconciliation and exception handling
Millions of transactions matched every day, the breaks investigated before anyone opens a spreadsheet, and a person approving every posting, with the audit evidence built as the work is done.
- Predictive ML
- Agents with tools
- Human in the loop
- Event-driven integration
Credit decisioning support
Every referred application arrives with the evidence gathered, affordability worked out and the score explained in plain English. A person makes every referred decline, clear-cut cases stay with the existing decision engine and its safeguards, and fairness is checked every month.
- Predictive ML
- Document AI
- Agents with tools
- Human in the loop
Spotting vulnerable customers
Possible signs of vulnerability noticed the first time a customer mentions them, a trained person offered the right support, and outcomes checked so no group of customers is left worse off.
- Retrieval (RAG)
- Predictive ML
- Speech and voice
- Human in the loop
Renewal pricing fairness monitoring
Every renewal price checked against the new-business price before it goes out, every pricing change tested for fair value and proxy discrimination, and clear evidence for the senior manager who signs the attestation. People own every pricing decision.
- Predictive ML
- Agents with tools
- Human in the loop
- Event-driven integration
Appointment no-show prediction
Fewer empty clinic slots, because the patients most likely to miss an appointment get help to attend or rebook in good time. Prediction leads to support, never to penalties, and people make every decision about a patient's care.
- Predictive ML
- Agents with tools
- Human in the loop
- Event-driven integration
Waiting-list validation
A waiting list that matches reality: data errors found every night, patients asked whether they still need care, and a person confirming every change. Clinical validation stays with clinicians.
- Predictive ML
- Document AI
- Agents with tools
- Human in the loop
- Event-driven integration
Care notes and incident analysis
Better care notes in less time, often by voice, and early warning of patterns such as falls, weight loss, missed medicines and safeguarding signs, put in front of a manager who decides what to do.
- Speech and voice
- Predictive ML
- Human in the loop
- Retrieval (RAG)
Demand forecasting and replenishment
Demand forecast by product, store and channel every night, with promotions, weather and local events built in. Suggested orders and allocations ready each morning, buyers and merchandisers approving them, and less fresh food thrown away.
- Predictive ML
- Event-driven integration
- Agents with tools
- Human in the loop
Patterns
All eight building blocks
Document AI
Read forms, letters and PDFs into structured data.
Retrieval (RAG)
Answer questions from your own documents, with sources.
Agents with tools
AI that takes small, controlled actions across your systems.
Human in the loop
Put people on the decisions that matter, with the evidence.
Predictive ML
Score, rank and forecast from your own historical data.
Event-driven integration
Connect AI to core systems safely, through events and queues.
Computer vision
Inspect photos, video and camera feeds for defects and damage.
Speech and voice
Turn calls and conversations into text, notes and actions.
Thinking about predictive ml in your organisation?
Talk through where it fits, what it would plug into, and which decisions should stay with people.
