Everybody loves AI when it writes emails and summarizes meetings. The moment it starts telling a company how many units to stock, what price to charge, or when to shut down a machine for maintenance? Suddenly the room gets quiet.
That’s the picture in Spain right now: onlyone in four large companiesusesprescriptive AIinternally, according to a study. And even that “one in four” is doing it in two very different ways,19%are running it on targeted projects, while just5%have baked it into the business as a cross-company capability with a dedicated team.
The gap tells you everything: prescriptive AI isn’t a shiny add-on. It’s a power tool pointed straight at decision-making, and that means data headaches, system integration pain, and the uncomfortable question of who’s on the hook when the machine’s “best move” blows up.
Prescriptive AI, in plain English: stop predicting, start recommending
Think of AI in three layers.Descriptivetells you what happened.Predictivetells you what might happen.Prescriptivetells you what to do next, usually by optimizing a goal while respecting constraints.
In a real company, that can mean moving from dashboards to operational recommendations you can actually execute: tweak inventory levels, reshuffle schedules, choose between pricing options, or allocate resources across teams.
And that’s why the split matters. The19%“project mode” crowd is typically doing proof-of-value work: narrow scope, one business sponsor, short timeline, and heavy dependence on a few datasets that (hopefully) aren’t a mess.
The5%“fully integrated” group is playing a different sport. That requires governance, repeatable deployment, security standards, monitoring, and the ability to roll models out across departments without reinventing the wheel every time.
Why prescriptive AI is lagging behind generative AI
Generative AI spread fast because it slips neatly into low-stakes workflows: drafting text, searching internal knowledge bases, helping support teams respond faster. It’s visible, easy to demo, and usually doesn’t get anyone fired if it spits out something dumb.
Prescriptive AI goes for the jugular: procurement, pricing, staffing plans, maintenance schedules, logistics, budget allocation. When it’s wrong, the damage isn’t “oops, weird email.” It’s missed shipments, angry customers, wasted money, and a very human blame game.
Three blockers show up again and again:
1) Data quality and availability.Prescriptive recommendations need reliable history, consistent master data, real-time feeds in many cases, and shared definitions of KPIs. If “margin” means three different things in three departments, good luck optimizing it.
2) Integration with core systems.A recommendation trapped in a slide deck changes nothing. To matter, it has to connect to the systems that run the business, ERP, warehouse management, CRM, planning tools, so actions can be executed and tracked.
3) Human acceptance.Operators will ignore a recommendation they don’t trust, don’t understand, or that clashes with on-the-ground reality. And they’ll fight it if it improves a corporate KPI while wrecking their local metrics.
There’s also a fourth problem companies love to underestimate: prescriptive AI forces leadership to define what “good” even means. Lower cost, faster delivery, higher service levels, lower carbon footprint, higher margin, pick your poison. These goals collide. A prescriptive system encodes policy. If executives aren’t aligned, the model ends up optimizing the wrong thing with impressive mathematical confidence.
Where prescriptive AI actually pays off: logistics, pricing, maintenance
The best use cases live where constraints are clear and outcomes can be measured.
Inlogistics, prescriptive AI can recommend inventory and replenishment decisions based on demand, supplier lead times, promotions, and warehouse capacity.
Inplanning, it can propose staffing assignments or production sequences while accounting for skills, labor rules, and machine constraints.
Inpricingand commercial strategy, it can recommend actions like discount levels that protect margin while hitting volume targets, which products to push seasonally, or how to segment offers under profitability constraints.
Industrial maintenanceis another natural fit: pair failure prediction with recommendations on when to intervene, what to prioritize, and how to plan spare parts.
But here’s the catch: the value isn’t the recommendation. The value is executing it. Plenty of companies build fancy pilots and then stall out because approvals take too long, departments fight over goals, or nobody sets up a feedback loop to improve the model over time.
The fact that19%of Spain’s large companies are using prescriptive AI only in specific projects screams “cautious experimentation”, small pockets where ROI is easiest to prove. The5%who’ve integrated it fully are the ones with the boring-but-critical muscle: standardized data, model maintenance, and performance management that lasts longer than a quarter.
What “full integration” really means: a dedicated team and company-wide reach
When prescriptive AI is truly integrated, it can’t live as a side hustle inside IT or a lonely data lab. It needs an end-to-end pipeline: data collection, preparation, governance, training, deployment, monitoring, and continuous improvement.
In mature organizations, prescriptive AI starts to look like an internal product: versioning, performance metrics, escalation rules, documentation, the whole grown-up package.
That’s partly because prescriptive models age fast. Customer behavior shifts. Costs change. Suppliers miss deadlines. Regulations tighten. Without MLOps and monitoring, a model can quietly rot while everyone assumes it’s still “smart.” A dedicated team also prevents a graveyard of one-off tools nobody maintains.
Going cross-company also forces a painful cleanup: different departments often calculate the same metric differently. Prescriptive AI can’t optimize an indicator that isn’t stable.
And humans don’t disappear. In many deployments, it’s “human in the loop”: the system recommends, an operator approves. Automation rises only when trust, controls, and the stakes allow it. For high-impact decisions, human validation usually stays, if only to handle the weird exceptions the model can’t.
Regulation, traceability, and liability: the weight on prescriptive systems
Prescriptive AI touches decisions, and decisions come with responsibility. Pricing recommendations, credit allocation, case prioritization, workforce management, these can have immediate economic and social consequences.
So the demand fortraceabilityis higher than for “productivity AI.” Companies need to explain why an action was recommended, what data fed it, what rules applied, and what controls were in place.
Europe’s AI regulatory framework pushes harder on governance depending on risk level. For big companies, the message is blunt: prescriptive AI can’t be a duct-taped collection of models. It has to fit into compliance, security, internal controls, and risk management.
That’s why deployments often stick to areas where explainability is easier, data is less sensitive, and impacts on individuals are indirect. Once prescriptive AI starts affecting individual outcomes, the bar jumps: documentation, audits, bias management, access governance, and ways for people to challenge decisions.
What Spain’s numbers say about data maturity
Prescriptive AI is rarely step one. It usually comes after companies clean up data, unify master records, spread reliable KPIs, and industrialize predictive models. If those foundations aren’t there, prescriptive systems become fragile, contested, and expensive to keep alive.
The19% projectversus5% fully integratedsplit suggests Spain has real experimentation, but limited scaling. It may also reflect priorities: plenty of firms chased faster, flashier wins first, like document automation, customer service tooling, and internal assistants.
If companies want to speed up prescriptive adoption, the playbook isn’t mysterious: stronger data governance, business goals that don’t contradict each other, tighter integration with existing systems, and real monitoring for model performance.
But the biggest lever is organizational. Prescriptive AI only works when it plugs into an actual decision circuit, clear owners, guardrails, and a disciplined feedback loop. Generative AI made executives talk about AI. Prescriptive AI forces them to decide how much decision-making they’re willing to hand over, and how they’ll prove it keeps working.
