Businesses Are Investing in AI but Struggling to Measure Its Returns

Businesses are spending time and money testing artificial intelligence, but determining whether those investments produce lasting value can be harder than launching a pilot. A demonstration may impress employees and senior leaders without showing whether a process has become cheaper, more accurate or more useful to customers. For organisations moving beyond experimentation, the central challenge is to connect AI projects to measurable business outcomes.

The first difficulty is defining what “return” means. A tool that drafts a report more quickly may save employee time, but that time has financial value only if it is redirected to productive work or reduces a real cost. A customer-service assistant may handle more enquiries, but a higher volume does not necessarily mean better service if customers must contact the company repeatedly to resolve the same issue. Companies need to decide in advance which outcomes matter and how they will be measured.

A useful evaluation begins with a baseline. Before changing a process, teams should record its current cost, cycle time, error rate, throughput and relevant customer outcomes. They should then compare results after deployment, allowing for changes in workload, staffing or seasonality. Where possible, a controlled trial or comparison group can help distinguish the effect of the AI system from other changes happening at the same time.

The full cost of implementation is often broader than the price of a software subscription. Businesses may need to pay for integration, data preparation, staff training, security reviews, monitoring and human oversight. They may also have to redesign workflows or correct mistakes generated by the system. If these costs are excluded, a project can appear profitable on paper while consuming more resources than expected.

Accuracy and reliability should be measured alongside speed. For some tasks, a small error rate may be acceptable if a person reviews every result. For others, a single error could create financial loss, regulatory exposure or damage to customer trust. Teams should examine the types of mistakes that occur, not only an average accuracy score, and should track how frequently employees must intervene. The cost of handling exceptions can change the economics of a project substantially.

AI performance can also vary over time. Models and software products are updated, business data changes and employees discover new ways of using a tool. A successful pilot may not scale cleanly to a larger group or a different customer population. Companies should therefore treat evaluation as an ongoing process, with clear owners, regular reviews and a way to pause a system when performance falls below agreed standards.

A balanced scorecard can combine financial and operational measures. These might include cost per completed task, time saved, error and rework rates, customer satisfaction, employee adoption, system availability and the number of cases escalated to a human. Not every project needs every metric. The goal is to choose a small set of indicators that reflect the original business problem and reveal unintended consequences.

Comparisons between projects also require care. A customer-support tool, a forecasting model and an internal coding assistant may create value in different ways and over different time horizons. A simple ranking based only on immediate cost savings can undervalue projects that improve quality or enable new services. At the same time, strategic potential should not become an excuse to avoid accountability: long-term projects still need milestones and evidence that progress is being made.

Governance is part of the return calculation. A tool that exposes confidential information, produces discriminatory outcomes or makes unsupported claims may create costs that outweigh its benefits. Companies should assess privacy, security, legal and reputational risks before deployment and decide which uses require human approval. Clear documentation also helps leaders understand why a project was approved and whether it remains appropriate as conditions change.

The strongest AI investment decisions are likely to be disciplined rather than dramatic. Businesses should begin with a defined problem, test the smallest useful solution, compare results with a baseline and scale only when the evidence supports it. The question is not whether AI can perform a task; it is whether the complete process produces a better outcome at an acceptable cost and risk.

This is a general analysis draft. To publish it as a current business report, add verified survey results, company examples and attributed comments from finance or technology leaders.