From Predictive Models to Autonomous Decisions: AI Use Cases

The 2021 version of this article asked what you can foresee with predictive modeling. The answer has not changed much: demand, delivery times, patient deterioration, machine failure, fraud. What changed is what happens next. A forecast that lands in a dashboard and waits for someone to read it is a report. A forecast that triggers a replenishment order, reroutes a shipment or opens a maintenance work order, inside limits a person has set, is a decision.

That shift, from predictions to autonomous decisions, is what the industry now calls decision intelligence. This article covers how the loop works, what still makes it hard, and where it is running in supply chain, healthcare, manufacturing, retail and insurance, including a shipment ETA model we built on 304K shipments.

From a prediction to a decision: the loop

A decision system has four stages. Predictive modeling is the first.

1. Predict. A numeric model estimates what will happen: the ETA of a shipment, the probability a patient deteriorates in the next 12 hours, the remaining useful life of a bearing. These are still gradient-boosted trees, regression, time-series and survival models. Large language models do not do this part well, and do not need to.
2. Decide. Rules and optimization turn the prediction into an action: if the ETA slips past the customer’s dock window, which alternative carrier, at what cost, against which service-level penalty. This is where business policy lives, and it is deterministic.
3. Act. The action is executed in a system of record: the TMS, the EHR, the maintenance system. In 2026 this is increasingly an AI agent with tools, orchestration and an escalation path, which is what makes the loop autonomous rather than advisory.
4. Learn. The outcome is recorded and fed back: was the shipment late, did the patient deteriorate, did the bearing fail. The model is retrained, the thresholds tuned, the exceptions reviewed.

The language model’s role is stages 3 and 4: executing multi-step actions, explaining the decision to the person who owns it, and drafting the exception for review. The forecast stays numeric. The agent makes it operational. Gartner expects 60% of supply chain disruptions to be resolved without human intervention by 2031, and forecasts supply chain software with agentic AI growing from under $2B in 2025 to $53B by 2030. Both numbers describe this loop, not a better forecast.

What still makes it hard

The challenges section of the 2021 article was about learning techniques: supervised versus unsupervised, one-shot, transfer learning. The hard problems now sit around the model.

Data quality decides the model. In our shipment ETA project, geocoding accuracy mattered more than algorithm choice. Missing coordinates and wrong city-to-country mappings cut the first usable dataset to roughly 4,400 shipments. After cleaning, the production models trained on 304,441. Feature engineering (lanes, distances, business-day indicators, service codes) had more effect than model complexity.

Explainability becomes mandatory once the system acts. When a prediction only informs, a black box is tolerable. When it triggers a purchase order or a clinical alert, the owner has to be able to say why. In the ETA project, linear regression and XGBoost reached similar error, about 21 to 24 hours RMSE. The linear model is easier to explain in a logistics operation, which is a reason to prefer it. Simpler models can win.

Drift is permanent. Lanes change, suppliers change, patient populations change. A model trained once and left alone degrades. Retraining pipelines with evaluation gates, and monitoring of both inputs and outcomes, are part of the system, not a maintenance task.

Acting on uncertainty. A prediction is a distribution, not a number. The decision layer has to know the cost of each error: an ETA that is 6 hours early costs less than one that is 6 hours late; a missed deterioration costs more than a false alarm. Thresholds are set on those costs, per action, and tuned on outcomes.

Governance of autonomous decisions. Who may set thresholds, who reviews exceptions, what is logged, what an auditor sees. Deloitte found only 1 in 5 organizations has a mature governance model for autonomous agents. It is the part most pilots skip and most production reviews require.

Bias. A model reproduces the biases in its training data. In healthcare, insurance and hiring the consequence is unequal treatment. Error rates are measured per segment, and segments with worse performance get a higher bar before automation.

What decision intelligence delivers

The benefits list from 2021 (demand, workforce, competitor, risk and churn prediction) is still right. The value moved from the prediction to the action it triggers:

• Demand forecast → replenishment. Orders drafted and, within limits, released, instead of a report a planner reads on Monday.
• ETA prediction → exception management. Customers notified and shipments rerouted before the dock window is missed.
• Failure prediction → scheduled maintenance. A work order with parts and a technician, created before the line stops.
• Deterioration risk → escalation. A nurse paged with a score and the reason, not a chart to review.
• Churn risk → retention action. An offer or a call, triggered at the right moment, with the reason logged.
• Fraud risk → hold and review. A claim held at first contact, with the evidence assembled for the adjuster.

Use cases by industry

Supply chain: predictive ETA and exception management

Predictive ETA is now an expected feature of transportation visibility platforms. BCG puts adoption near 50%, and Gartner lists it as an essential capability of real-time transportation visibility platforms. Shippers use it to plan dock labor, set customer expectations and decide when to intervene.

Erbis use case: shipment ETA on AWS SageMaker. For a global agricultural equipment manufacturer, served through a transportation management SaaS, we built ETA models that predict delivery time from historical transit data. Linear regression and XGBoost models, trained on Amazon SageMaker, on 304,441 unique shipments across 60+ international lane combinations. The linear model reached an RMSE of 21.4 hours on 80,889 test shipments. Predictions are served from SageMaker endpoints behind a Lambda function that routes requests by account, so the same pipeline serves multiple customers. Pay-as-you-go SageMaker avoided $10K to $50K of GPU capital expenditure. Read the full case study.

The decision layer sits on top. When a predicted ETA crosses a customer’s threshold, the exception is raised, the customer notice drafted and the reroute options priced. That is the step from a forecast to an operation, and it is where the agentic AI engineering work goes.

Healthcare: early warning and escalation

Predictive models in care settings estimate risk of deterioration, readmission and no-show, and increasingly feed revenue-cycle decisions such as prior authorization. Providers have moved past strategy: Bain and KLAS found 70% of providers have an AI strategy, and fewer than 5% say deployed AI misses expectations.

Ysbyty Gwynedd use case: early symptom detection. Ysbyty Gwynedd, the general hospital in Bangor, Wales, uses Philips IntelliVue Guardian to reduce manual observation rounds. The software computes an Early Warning Score from vital signs and alerts nurses when a patient’s condition is likely to worsen, so intervention happens earlier.

The decision-intelligence version of this is the escalation path: the score, the reason, the suggested action and the page to the right clinician, with every alert and response logged. The human stays in the loop by design. The system removes the delay, not the judgment.

Manufacturing: predictive maintenance and quality

Manufacturing replaces real estate in this list because it is where predictive models most often trigger actions today. Industry surveys report 47% of manufacturers use AI in quality processes, up from 33% in 2025 (Metrology News), and Deloitte cites 30% to 50% less unplanned downtime from predictive maintenance.

The loop is concrete. Vibration, temperature and current sensors feed a remaining-useful-life model per asset. When the predicted failure window crosses the planning horizon, a work order is created in the maintenance system with the parts list and a proposed slot in the production schedule. A planner approves the slot. The rest is automatic. In quality, a computer-vision model flags a defect, and the decision layer decides whether to divert the unit, stop the line or sample more, based on the defect class and the cost of each error.

Retail: demand forecasting and replenishment

Retailers forecast demand per SKU and store and use the forecast to replenish, price and staff.

Carrefour use case: inventory management. Carrefour uses AI-driven demand forecasting to plan supply and stock stores, reducing both shortages and overstock. The autonomous version closes the loop: replenishment orders released within limits, markdowns proposed for slow stock, and exceptions (promotions, weather, supply disruptions) routed to a planner with the reason.

Insurance: fraud screening at first contact

Fraud models score claims and applications against past patterns and flag anomalies.

Lemonade use case: claims processing. Lemonade handles first contact through an AI assistant that checks customer history, reads the claim and flags possible fraud for review, so suspicious claims are held before payment rather than investigated after.

The governance point matters here more than anywhere. A hold is an action against a customer. The threshold, the review path and the logging are what regulators ask about, and a person makes the final call on a denial.

Where the human stays

Autonomy is a policy, not a switch. For every decision the loop can make, write down three things:

• Decision rights. Which actions the system may take alone, which need approval, which are never automated. Money thresholds, reversibility and regulatory exposure are the usual criteria.
• Thresholds and their owners. Who sets the confidence threshold for each action, on what evidence, and who can change it.
• Review and audit. What share of automated decisions a person samples, how exceptions are handled, and what the log contains: inputs, prediction, decision, action, outcome, identity.

Early in production most actions need approval: human in the loop. As outcomes accumulate and error rates are known, approval moves to exceptions and the person reviews samples: human on the loop. The move is made on numbers, per action, not on a date.

How to start

Begin with one decision, not one model. Pick an action with a clear cost, a measurable outcome and a system of record it can be written to. Build the evaluation set from your own history. Ship the prediction, the decision rule and the action together, with the human checkpoint in place, at a bounded scope. Measure against the current process. Then decide.

That is the shape of our AI Pilot in 4 Weeks: fixed fee, fixed timeline, your data, and a go/no-go decision with a production plan at the end. The data platform underneath (lakehouse, feature pipelines, retraining and monitoring on AWS or Databricks) is the work of our data and AI platforms practice.

Erbis has built software since 2012 and production AI since 2017, with 200 engineers, $110M+ in documented client savings and ~95% AI accuracy in production.

FAQ

What is decision intelligence?

The practice of connecting a prediction to an action through explicit decision rules, executing the action in a system of record, and learning from the outcome. It adds a decision layer and an execution layer, increasingly an AI agent, to conventional predictive modeling.

Do large language models replace predictive models?

No. Numeric forecasting is still done by regression, gradient-boosted trees, time-series and survival models, which are cheaper, more accurate and easier to explain for that task. Language models execute multi-step actions on the prediction, explain decisions and draft exceptions for review.

How accurate does a prediction need to be before you automate a decision?

Accurate enough that the expected cost of acting on it is lower than the cost of the current process, for that specific action. The threshold is set per action from the asymmetric costs of each error, tested on an evaluation set from your own history, and tightened for segments where the model performs worse.

Which industries are furthest along?

Supply chain (predictive ETA, replenishment, disruption response), manufacturing (predictive maintenance, visual quality inspection) and healthcare (deterioration and readmission risk, revenue-cycle decisions) have the most production deployments where predictions trigger actions. Retail and insurance have mature forecasting and fraud models and are automating the actions behind them.

If you have forecasts that people read and act on by hand, book a discovery call. We will tell you which decision to automate first, and whether your data is ready for it.