The Hidden Challenge Behind Corporate AI Investments

AI Investment Needs More Than One Measure of Success

Imagine a company reviewing three AI proposals: a customer service product, a research partnership, and an internal tool for resolving invoice disputes. Each requests funding and promises value. Yet the evidence that justifies backing one may tell managers little about the others.

A product needs customers and reliable delivery. A partnership needs useful collaboration. An internal tool needs to solve a specific operational problem. Evaluating all three with the same checklist can reward projects that produce convincing forecasts while overlooking those that require experimentation.

A study in Technovation, “Configuring innovation portfolio management practices in an era of emerging digital technologies,” offers a framework for understanding this mismatch.¹ Its implications extend beyond project management: investors assessing corporate AI spending should examine how companies decide what deserves more capital.

What the Digital Innovation Study Found

Researchers Kristian Hilmar Mikkelsen Wegener and Karl Joachim Breunig interviewed 26 people across seven Scandinavian organizations. Interviews took place in 2023, with supporting documents used to strengthen their analysis. The paper became available online in October 2026 and appears in the journal’s 2027 volume.

The organizations spanned telecommunications, assurance and risk management, transport solutions, brewing, consulting, healthcare products, and public research funding. The research covered emerging digital technologies broadly, including AI, machine learning, and the Internet of Things.

The central finding concerns differences in kind. Two projects can differ in budget or risk while still fitting the same management process. Other activities create value in fundamentally different ways, making common evaluation criteria inappropriate.

The authors identified three recurring configurations. These are patterns of management practices that can coexist within an organization, rather than permanent labels assigned to entire businesses.

How Different AI Projects Create Value

Applying this framework to AI makes the distinctions easier to understand. The following examples illustrate its potential use; they are not additional findings from the research.

AI Products Need Repeatable Delivery

Consider a software company building an AI assistant for paying customers. Its work involves product design, testing, deployment, and support. Structured reviews can assess whether the assistant meets quality requirements, attracts users, and can be delivered economically.

Here, management can reasonably request milestones and comparable commercial evidence. The central question is whether the company can repeatedly deliver something customers will pay for.

AI Partnerships Need Coordination

Now consider a manufacturer collaborating with a university and specialist suppliers on machine learning for equipment maintenance. Progress depends on access to data, complementary expertise, and agreement over how discoveries will be used.

A quarterly sales target may reveal little about whether that collaboration is becoming valuable. Better early questions concern data access, shared technical progress, and whether participants can combine their contributions into a usable solution.

This also clarifies the study’s use of “platform.” It describes facilitating exchange among participants, which need not mean selling a software platform.

Internal AI Tools Need Operational Evidence

An invoice dispute assistant presents a different challenge. Employees closest to the process understand exceptions, missing documents, and why apparently simple cases become complicated. Giving those specialists room to shape the tool may be more useful than enforcing specifications written before those complications were understood.

Success might mean fewer reopened cases or faster resolution without additional errors. Relevant market context includes Securities.io’s coverage of enterprise search and knowledge discovery and enterprise integration and workflow automation. Both areas connect AI ambitions to the information and processes businesses already use.

Measure AI Progress Before Comparing Returns

A useful implication is that companies need two layers of evaluation. The first asks whether an initiative is progressing on its own terms. The second asks whether continued spending remains justified relative to other uses of capital.

This avoids a false choice between demanding immediate revenue from every experiment and allowing experimentation to continue indefinitely. Different evidence can support funding decisions without removing financial discipline.

For example, an early research partnership might earn another funding round by demonstrating that a previously inaccessible dataset can support a useful prediction. Later, the resulting product should face stronger tests of reliability, adoption, and economics.

Managers can make that transition explicit with three questions:

• What uncertainty must the next funding stage resolve?
• What evidence would justify expanding, changing, or stopping the work?
• Who has the knowledge and authority to make that decision?

These questions are an application of the framework, rather than a tested formula from the paper. Their advantage is practical: they connect spending to learning and subsequent action.

Shared AI Infrastructure Can Distort Project Economics

Another implication concerns shared costs. Suppose several AI tools depend on the same data cleaning, permissions, and monitoring capabilities. Charging the entire setup cost to the first project could make it appear unattractive. Ignoring that cost across every project would exaggerate profitability.

A clearer approach separates reusable infrastructure from the additional expense of each application, while still accounting for both at company level. Investors can then distinguish an expensive foundation from an application that remains uneconomic after that foundation exists.

Security belongs in this calculation. Recent coverage of AI data security investment highlights the commercial attention surrounding agent access and sensitive information. For managers, the analytical lesson is to budget for controls alongside the capabilities they enable.

Decentralizing project decisions also does not require decentralizing every rule. Teams can choose suitable workflows while operating within shared access controls, spending limits, and accountability requirements.

Microsoft Offers Exposure Across Enterprise AI

These management challenges create demand for technology that supports both experimentation and repeatable deployment. Microsoft offers publicly traded exposure through enterprise software, cloud infrastructure, developer tools, and AI services.

In its 2026 annual report, Microsoft describes an AI offering spanning applications, models, infrastructure, and enterprise controls. Its relevance to this discussion lies in that breadth: customers can use the company’s technology for commercial products, collaborative development, and internal workflows.

Microsoft’s September announcement of Copilot capabilities and AI spending controls also connects directly to the problem. Alongside new capabilities, the company outlined tools intended to help organizations manage AI consumption and spending policies.

The investment argument is conditional. If customers turn experiments into sustained workloads, Microsoft could benefit from recurring software demand and infrastructure consumption. If deployments remain costly or operationally disappointing, expanding technical capabilities alone may not translate into attractive customer economics.

The study does not evaluate Microsoft, identify it as a participant, or establish that its products improve innovation management. The company is an illustrative investment connection. Assessing its shares still requires examining valuation, competition, infrastructure spending, and the profitability of growth.

Latest Microsoft (MSFT) News and Developments

What Investors Should Look for in AI Spending

The paper’s limitations matter. Its findings come from reported practices within a small group of Scandinavian organizations. It does not demonstrate higher returns, measure AI project failure rates, or prove that adopting the framework improves performance.

Nevertheless, it provides a useful lens for reading corporate AI strategies. Counts of pilots, partnerships, or deployed tools reveal activity. More informative disclosures explain the problems being addressed, evidence required for expansion, and how spending changes when results disappoint.

A company that stops an unproductive pilot may be allocating capital effectively. A company with many pilots may still lack a route to deployment. The distinction depends on decision quality and operational evidence.

As AI reaches more business functions, the ability to fund different kinds of innovation coherently could become an important competitive advantage. Better models expand what companies can attempt. Better management helps determine which attempts become durable businesses.

References:

¹ Wegener, K. H. M., & Breunig, K. J. (2027). Configuring innovation portfolio management practices in an era of emerging digital technologies. Technovation, 159, 103752. https://doi.org/10.1016/j.technovation.2026.103752