Not quite automatic: Decision optimization in the AI age is complex

Decision optimization technology uses mathematical models and algorithms to make critical business decisions based on constraints (e.g., budgets or time) and objectives (e.g., risk reduction or profit maximization). In the AI era, decision optimization has become more prescriptive, with AI-generated recommendations guiding optimal actions. It has also become easier to automate, since those actions can be triggered autonomously. This could lead to the assumption that enterprises can now use AI to automate many business decisions. However, the reality is more nuanced.

As findings from 451 Research's Voice of the Enterprise: AI & Machine Learning, Use Cases 2026 survey illustrate, organizations are now using AI to automate decision-making. But that does not necessarily mean the whole process is automated. The most common approach, according to the survey data, is a blend of human involvement and AI, where AI informs decisions and workflows only. This suggests organizations are predominantly using AI for prescriptive analytics only. Furthermore, the survey data indicates that the current popular narrative — that AI-automated business decisions will supplant data-driven decisions made by humans — is fundamentally flawed. There are a variety of technical, economic, cultural and legal reasons why AI-driven automated decisions are not desirable or feasible for many use cases. However, a certain class of decisions that are repetitive, low-risk and low-complexity can significantly benefit from automation, and this is where the growth in adoption of this approach will occur.

Context

AI is playing a pivotal role in decision-making by enabling systems to make decisions with minimal or no human involvement based on AI-generated recommendations, promising speed, efficiency, cost reduction and productivity gains. But successful decision automation relies on meeting a whole host of criteria, making it inappropriate or unrealistic for many decisions that call for human involvement.

Our findings show that almost half (46%) of organizations are using AI for decision optimization, according to our Voice of the Enterprise: AI & Machine Learning, Use Cases 2026 survey. Furthermore, 43% plan to adopt it in the next 12 months, reflecting its current popularity and future growth potential. That said, less than a quarter (22%) of organizations have a fully automated decision optimization deployment that executes with minimal or no routine human involvement.

Barriers to decision automation

A lack of confidence in AI-generated predictions is one likely reason why organizations rely on a human in the loop. Only 17% of enterprises fully trust predictions made by their own AI apps/algorithms, while 16% completely trust predictions made by AI apps/algorithms from other organizations. Furthermore, confidence in AI-generated predictions has hit a new low since we started tracking it in 2023, suggesting that, to some degree, human involvement will continue to prevail. In our Voice of the Enterprise: AI & Machine Learning, Use Cases 2023 survey, a quarter of enterprises fully trusted AI predictions from homegrown AI apps/algorithms and 24% fully trusted third-party model responses. A lack of explainability is a major reason why organizations are hesitant to trust automated AI-generated predictions. The concern is that if the decision is wrong and the AI is not defensible due to "black box" decision-making, disastrous ramifications could ensue, such as significant customer attrition or damaged brand reputation.

There are also many technical hurdles to jump to make decision automation effective. For example, a well-defined AI model with clearly defined objectives and constraints, fed with diverse, high-quality data, is critical to ensuring accurate, optimal outputs. A model that has a prevalence of missing, noisy or biased data will produce unreliable and potentially unfair responses. Many decisions require up-to-date data, which is not easy to achieve either. The AI model must also be able to incorporate feedback loops effectively for improvement. Successful decision automation also relies on an organization having the appropriate skills to implement the technology, integrate it with relevant systems already in place, and address interoperability issues.

Budgetary constraints can also impede the automation of decisions. It requires parsing a high volume of enterprise data to deliver it effectively, which makes data processing costs expensive. Furthermore, the largest reasoning models will be necessary as they will be the most effective in delivering correct decisions, thereby adding another significant expense. Investments in other new technologies and infrastructure will also be required, such as AI infrastructure for inference, which is a major expense due to significant horsepower demands. Furthermore, the more AI infrastructure is required, the higher the energy costs involved. Maintenance expenses must also be factored into the budget. Labor costs to address skills gaps are another potential expense.

Finally, compliance with AI and data regulations may prevent decision automation. For example, the EU AI Act and GDPR Article 22 require meaningful human oversight and clear explanations for high-impact decisions. The EU AI Act also bans certain forms of decision automation, such as social scoring. Additionally, a company's own data privacy and security mandates might make decision automation a "no-go." If the organization is highly regulated and thus operates under audits, reporting and strict operational controls, it is likely to have stringent data privacy and security mandates that could be breached by automating certain decisions.

Decision automation criteria, use cases

That said, certain classes of decisions are safe to automate. They fall into the low-risk, low-complexity, high-volume category. These types of simple, repetitive decisions are also easier to explain, so organizations can justify them to users, auditors or regulators, if required. They are also typically reversible through a refund, reprioritization or reassignment, so if the AI makes a faulty decision, it can be easily and quickly corrected.

Our survey findings show that financial transactions are the most popular decisions to automate, as 36% of enterprises use an automated end-to-end approach with minimal or no routine human involvement, according to our Use Cases 2026 survey. This use case is closely followed by financial planning and analysis (33%) and asset valuation (31%). These data-driven decisions do not rely on subjective judgment that only humans can provide, but are determined numerically, which is a major reason they are automated. Inventory stocking based on a demand forecast, approving an expense claim under a certain amount, or autoscaling servers using API rate limits and capped energy demand responses are other use cases that are safe to automate for similar reasons.

Strategic and high-impact decisions clearly carry significant risk and are complex and difficult to explain, making them poorly suited to any degree of automation. Decisions that are ethically sensitive or require emotional intelligence are not good candidates for automation either. If a decision affects an individual's rights, livelihood or health, it will require human judgment, which means it cannot be fully automated.

Indeed, a plethora of use cases can be partially automated as the criteria involved apply to a variety of decisions, suggesting this approach will become more commonplace. Partial automation is the best choice for ambiguous decisions in which data is unable to capture all nuances; sensitive stakeholder decisions involving ethics, fairness and reputational risk; and decisions made in frequently changing conditions, such as rapidly changing markets. Insurance claims and underwriting, contract review and negotiation, and IT security response investigation are popular examples.

S&P Global Market Intelligence 451 Research is a technology research group within S&P Global Market Intelligence. For more about the group, please refer to the 451 Research overview and contact page
This article was published by S&P Global Market Intelligence and not by S&P Global Ratings, which is a separately managed division of S&P Global.