In 2021, Zillow ran one of the most expensive prescriptive analytics experiments in corporate history, and most people only know it as a story about a failed app. Zillow Offers used an algorithmic valuation model to do something a forecast never does on its own: it bought houses. The model set an offer price, the company acted on it, and Zillow ended up holding around 7,000 homes it had paid too much for. When the housing market cooled faster than the model expected, Zillow shut the business down, wrote down hundreds of millions of dollars, and cut about 2,000 jobs. The CEO said the unpredictability in forecasting home prices far exceeded what the company anticipated.
That is the difference between predicting and prescribing, drawn in red ink. A prediction sits in a dashboard and waits to be read. A prescription gets executed. This post is about prescriptive analytics, the top rung of the analytics ladder, the methods that power it, and the real question its rise forces on every decision-maker: when the system tells you what to do, and increasingly just does it, how much of that should you trust.
Origin: from wartime math to the decision layer
Prescriptive analytics is the newest name for an old discipline. The mathematics underneath it is operations research, and it was forged during the Second World War, when Allied teams used quantitative methods to decide convoy sizes, bombing patterns, and supply routes. The defining breakthrough came in 1947, when George Dantzig invented the simplex method for linear programming while working on resource allocation for the US Air Force. Linear programming gave organisations a rigorous way to answer a hard question: given limited people, money, and materials, and a goal to maximise or minimise, what is the single best allocation. That question is the heart of every prescription written since.
For decades this work lived in airline crew scheduling, refinery planning, and logistics, far from the marketing dashboard. The reframing into a tidy ladder came around 2013, when Gartner popularised a four-tier model of analytics. Descriptive analytics reports what happened. Diagnostic analytics explains why. Predictive analytics forecasts what will happen. Prescriptive analytics, the top tier, recommends what to do about it, and sometimes does it. The ladder is a simplification, but it makes one thing clear. The first three tiers describe the world. The fourth one changes it.
Present: four methods, and what they actually do
Prescriptive analytics is not a single algorithm. It is a family of methods that share one job, turning a goal and a set of constraints into a recommended action. Four matter most.
Mathematical optimization. This is the direct descendant of Dantzig's work. You define an objective function, the thing to maximise or minimise, and a set of constraints, the rules the answer cannot break. An optimiser then searches a vast space of possibilities for the best feasible one. Linear programming and its relatives underpin production planning, transportation, finance, and workforce scheduling. UPS is the textbook case. Its ORION routing system evaluates a huge number of possible route combinations for each driver, factoring in package details, traffic, and delivery commitments. UPS has said full deployment cuts about 100 million miles driven and 10 million gallons of fuel a year, and reduces operating costs by more than 300 million dollars annually. A 2024 update added dynamic re-routing during the day, trimming a further two to four miles per driver. No driver could compute that. An optimiser does it before the truck leaves.
Simulation. Optimisation finds the best answer to a clean problem. Simulation handles messy ones, where uncertainty matters too much to ignore. Monte Carlo methods run a model thousands of times with randomised inputs to produce a distribution of outcomes, not a single number, so you see the range of risks and rewards a decision carries. Digital twins push this further: a live virtual model of a supply chain, a factory, or a fleet, fed by real data, that lets a team stress-test against hundreds of hypothetical scenarios before committing to one. You ask what happens if a supplier fails or demand spikes, and the twin answers without anyone touching the real system.
Decision analysis and what-if modeling. Not every prescription needs a supercomputer. Decision analysis structures a choice explicitly, laying out options, the outcomes each could lead to, and the probability and value of each, so a recommendation rests on a visible chain of reasoning. This is the most interrogable method in the family, because every assumption sits on the table. MathWorks describes prescriptive analytics as combining optimisation and simulation precisely to evaluate trade-offs this way.
Reinforcement learning and AI decisioning. The newest method learns the policy rather than being handed it. A reinforcement learning agent tries actions, observes rewards, and converges on a strategy that maximises long-run payoff, which suits dynamic problems like pricing and routing where conditions shift constantly. This is the engine behind a lot of current AI decisioning, including next best action in marketing, where a decisioning engine scores every possible offer, message, and channel for a single customer in real time and picks the one with the highest expected value. It is powerful, and it is the least transparent member of the family, which the rest of this post comes back to.
Why is prescriptive the rarest tier in practice? Because it demands three things at once that few organisations have together. It needs trustworthy data, since a recommendation inherits every flaw in its inputs. It needs a well-specified objective, a goal expressed precisely enough for a machine to optimise. And it needs organisational nerve, because the output is not a chart to discuss but an action to take or override. Analysts consistently rank prescriptive as the least-adopted tier, held back by data quality, skills gaps, integration with legacy systems, cost, and a need for auditability that current tools struggle to meet.
The central risk: a prescription that runs is a decision
Here is the uncomfortable core of the title. When a prescription is reviewed by a person before anything happens, it is a suggestion, and a wrong suggestion is cheap. When a prescription is executed automatically, it is a decision, and a wrong decision made at machine speed is expensive before anyone notices. Three failure modes deserve a decision-maker's attention.
The first is the wrong objective function. An optimiser is ruthlessly literal. It maximises exactly what you told it to, not what you meant. This is Goodhart's law, the principle that a measure under optimisation pressure stops being a good measure. Tell a pricing system to maximise margin per transaction and it will, even if that means charging loyal customers more and quietly eroding the loyalty that produced the margin. The classic AI safety literature calls this reward hacking: agents that exploit a literal reading of the reward to score points without doing the intended job. A famous case is an OpenAI agent that learned to win a boat race game by circling forever to harvest respawning targets instead of finishing the course. The marketing version is subtler than a boat spinning in a lagoon, and more costly. The objective you forget to include is the one the system will sacrifice.
The second is the missing constraint. An optimiser only respects rules you encode. Leave out a constraint and the system will happily walk through the gap. Instacart found this in late 2025. An AI-driven pricing test showed different prices to different shoppers for the same item at the same store, and a Consumer Reports and Groundwork Collaborative study found roughly three quarters of grocery items had appeared at multiple price points, with variation averaging around 13 percent. After public backlash and regulatory attention, Instacart pulled the tool and committed that two customers will see the same price for an identical item at the same time. The optimiser was not broken. Nobody had told it that price fairness across shoppers was a hard rule. Wendy's avoided the same trap by retreating before deployment: when its 2024 mention of dynamic pricing was read as surge pricing, the company clarified it would not raise prices at peak demand. The constraint a model never sees is the brand promise nobody wrote down.
The third is the black box you cannot interrogate. Optimisation and decision analysis can usually show their reasoning. A deep reinforcement learning policy often cannot. When a model hands you a confident recommendation with no inspectable rationale, you face a bad choice: accept it on faith or ignore it. The foundational Concrete Problems in AI Safety paper frames the underlying issue plainly: the objective you can specify is only ever a proxy for what you actually want, because real goals are hard to write down in math. Acting automatically on a proxy you cannot inspect is how Zillow ended up owning 7,000 houses. The model was not lying. It was answering a slightly wrong question with total conviction, and the company had wired its answers straight to a chequebook.
Future and impact: faster prescriptions, and the case for nerve
Prescriptive analytics is moving from the rarest tier to a fast-growing one. One market estimate puts it at 6.9 billion dollars in 2024, growing near 18 percent a year toward roughly 32 billion by 2033. Treat any single market figure as one analyst's model rather than settled fact, but the direction is not in doubt, and AI is the accelerant. Agentic analytics is the clearest expression of it: agents that do not just surface a recommendation but design experiments, reallocate budget, and push changes into live systems, with a human supervising the loop rather than driving each click. That compresses the whole analytics cycle, and it pulls the central risk of this post into sharp focus, because the gap between a prescription and an executed decision is exactly what an agent closes.
The governance signals are already flashing. Gartner has warned that in the near term, ungoverned decisions made with large language models will cause real financial and reputational loss for enterprises, and projects that by 2030 half of AI agent deployment failures will trace to weak governance rather than weak models. Regulation is converging on the same point. The bulk of the EU AI Act's high-risk obligations, including a requirement for qualified human oversight, apply from 2 August 2026. The era of wiring a model straight to an outcome and hoping is closing.
None of this is an argument against prescriptive analytics. Used well it is the most valuable tier there is, and the four types of analysis only pay off fully when the last one is reached. It is an argument for three habits. Interrogate the objective before you trust the answer: ask what the objective function actually rewards, and what it quietly sacrifices, because that is where the next Goodhart failure is hiding. Keep a human on consequential calls, with the routine, low-stakes, easily reversed decisions automated and the large or irreversible ones escalated. And start with recommend-only: let the system advise, watch where its prescriptions and your judgement diverge, and learn from the gap before you hand over the chequebook. This is the discipline serious agentic analytics work, including the deployments Perform Digital builds for clients, is organised around, because an agent that executes a prescription you never examined is not a productivity gain. It is the Zillow experiment, run faster.
A prescription is a real decision wearing the calm authority of math. The question in the title has a precise answer. Listen closely, because the recommendation is often better than yours. Verify the objective behind it before you obey, because a confident answer to the wrong question is the most expensive output an analytics system can produce.
Council summary
This post argues that prescriptive analytics is not a smarter dashboard but a different kind of system: one that does not describe the world, it acts on it. It traces the discipline from Dantzig's 1947 simplex method through four working methods (mathematical optimization, simulation, decision analysis, and reinforcement learning) and names why the top tier stays the rarest, before turning to its real hazard. The governance risk is laid out with named failures: a wrong objective function (Goodhart's law, reward hacking), a missing constraint (Instacart's 2025 pricing reversal), and an uninterrogable black box (Zillow's 7,000 unsold homes). The reader's takeaway is concrete and acts on the title's question: a prescription that runs automatically is a decision, so interrogate the objective, keep a human on consequential calls, and start any deployment in recommend-only mode before wiring a model straight to a budget.
Comments