Enterprise AI spending is rising despite weak financial returns

Enterprise companies are increasing AI spending even as earlier automation investments deliver substantially less financial value than expected. Many are financing new initiatives with savings associated with those underperforming projects. As a result, the assumptions behind one investment can carry into the next before the company has established how much value it captured.

Bain’s 2026 research shows the scale of the problem. Companies are struggling to turn AI capabilities into measurable savings, while major software providers are committing billions of dollars to deployment engineering, governance, and new ways of measuring AI work. Those commitments suggest providers increasingly see implementation and operational performance as part of their commercial responsibility.

The shift matters because enterprise software economics depend on what happens after deployment. A technically successful implementation can still deliver disappointing financial results when workflows remain inefficient, employees must supervise automated systems, or organizations cannot connect AI activity to business outcomes.

Enterprises keep funding AI despite missing savings targets

Enterprise AI investment faces a financial contradiction: companies continue approving new spending while the savings expected from previous investments remain partly unrealized.

The Bain Automation and AI Pathfinder Survey 2026 found that nearly 40% of companies measuring AI cost savings achieved less than 10%, despite targeting savings of 11% to 20%. Those targets often help justify the initial investment, so falling short weakens the financial assumptions behind subsequent spending.

Yet most companies are increasing their AI budgets. Many are using savings associated with earlier automation investments to finance additional initiatives, even when those projects performed below expectations. Future investments may therefore rely on financial assumptions that previous deployments have already challenged.

Consider a company that approves an automation project based on projected cost reductions. If the project achieves only part of those reductions, the organization has less financial capacity than its original business case anticipated. Using the original savings projection to justify another investment would extend the gap between expected and realized benefits.

For CFOs and boards, the immediate task is to establish how much projected value became measurable savings or other financial benefits. That means tracking actual operating results and updating investment assumptions before approving further expansion. Technology leaders can use the same evidence to decide which deployments deserve additional funding.

Software providers also have a commercial interest in closing this gap. Vendors can demonstrate functionality, complete deployments, and expand adoption while customers struggle to establish the financial returns. When those returns remain unclear, future purchasing and expansion decisions become harder to justify.

Customers therefore have reason to expect more help with workflow design, implementation planning, and performance measurement. The challenge extends beyond generative AI, however, because enterprise automation has struggled to deliver its expected value for decades.

The value gap predates generative AI

The difficulty of converting technical capability into financial results has persisted through roughly 20 years of enterprise automation.

Bain’s 2026 research, covering 951 companies and 63 enterprise processes, found substantial differences between the potential business value of automation and what organizations actually realize.

These rates measure the share of potential business value that organizations convert into measurable results. They offer a broader view of financial performance than cost savings alone because business value can include productivity, capacity, quality, and other operational improvements.

Traditional automation’s 52% realization rate is particularly revealing. Organizations have spent roughly two decades developing process-mapping methods, integrating systems, and managing organizational change. Even with that experience, almost half of the technology’s potential value remains unrealized.

Generative AI faces additional technical and operational challenges, which helps explain its lower realization rate. Yet the historical comparison also points to a wider problem: better technology does not automatically improve business economics. Organizations must change how work is performed and how benefits are captured.

The problem often begins during investment planning. A business case estimates implementation costs, identifies expected benefits, and calculates a projected return. Once the technology enters production, integration requirements, employee involvement, existing procedures, and continuing expenses determine how much of that return materializes.

Historical realization rates can make those projections more realistic. Consider an illustrative investment case forecasting $10 million in savings. If comparable projects have historically realized 60% of their expected benefits, using $6 million as the planning assumption would better reflect the organization’s experience.

The 60% rate is a hypothetical planning assumption rather than an additional Bain survey finding. It illustrates how organizations can apply historical performance when evaluating new investments. Budgeting for the full $10 million would create a more aggressive expectation that the project would then need to justify.

Once deployment begins, the same method supports ongoing capital allocation. Enterprise teams can compare projected benefits with measured results, calculate realization rates, and update assumptions for future projects. Each deployment then provides evidence about the organization’s ability to capture value.

For software providers, these economics create an incentive to improve implementation outcomes. Technical performance remains essential, but financial returns also depend on how well the technology fits the customer’s workflow.

That dependency becomes especially important when AI business cases assume greater autonomy than production systems can deliver.

AI business cases overestimate production autonomy

One reason AI investments fall short of expected returns is the gap between assumed automation and actual production operations.

According to Bain’s 2026 research, only about 7% of companies operate fully autonomous AI agents in production. Yet many investment cases assume the economics of full automation, including the labor savings expected when software completes entire workflows independently.

A fully autonomous AI agent can execute an assigned workflow without routine human approval. A supervised agent requires people to review decisions, authorize actions, handle exceptions, or check results. Both approaches can create value, but their operating costs differ.

When an investment assumes full autonomy, its financial model may anticipate eliminating substantial labor expense. If employees must continue reviewing outputs or approving transactions, the organization retains part of that expense while adding AI infrastructure, integration, and supervision costs. Productivity can still improve, but the financial result will differ from the original projection.

Human supervision changes the economics

Supervision becomes especially important in business functions where errors have significant consequences.

Customer-service agents may need human intervention when requests become unusual or sensitive. Financial reporting systems may require approval before information enters official records. Compliance processes and internal auditing often depend on human judgment, documented decisions, and clear accountability.

These requirements add operating costs, but they can also prevent expensive mistakes. An approval step may reduce the risk of an incorrect transaction, while a review process can catch inaccurate information before it affects financial records. The calculation therefore needs to include supervision costs alongside the errors and risks that oversight helps prevent.

The economically preferable design may retain human involvement even when AI can execute more steps independently. A company might benefit from extensive autonomy in a routine workflow while preserving detailed approval requirements in a sensitive one. Applying the same autonomy assumption across an entire AI program can hide those differences.

For each process, enterprise teams need to compare labor expenses, technology costs, error rates, operational risks, and oversight requirements. That comparison determines whether additional autonomy creates enough value to justify the change.

It also reveals why automating an existing workflow can produce disappointing returns.

Workflow debt limits automation benefits

Many enterprise processes contain unnecessary handoffs, repeated approvals, complicated exception procedures, and legacy rules that remain because earlier systems required them.

Introducing AI into those processes can make individual activities faster while preserving the surrounding inefficiency. The organization pays for automation but continues carrying the operating costs of its existing workflow.

Bain calls this accumulated burden workflow debt. It develops when organizations digitize existing procedures without reconsidering how work should be performed, then build subsequent automation initiatives on those inherited processes.

For example, an AI agent might prepare a decision in seconds while the decision still passes through several approval stages designed for an older operating model. Preparation becomes faster, but employees continue spending time on approvals and handoffs.

When later automation projects inherit those approval stages, exception rules, and system dependencies, the workflow becomes harder to change. The organization can end up maintaining expensive procedures even as individual activities become more automated.

Workflow assessment should therefore come before decisions about where AI will operate. Enterprise teams need to identify which activities remain necessary, where human judgment adds value, and which procedures exist mainly because of historical system constraints.

A redesigned process may remove unnecessary steps while preserving supervision where it reduces meaningful risk. The goal is to improve the economics of the complete workflow, including the cost of operating and governing the AI system.

Gradual autonomy improves investment planning

Once a workflow has been assessed, organizations can introduce autonomy in stages that reflect actual operating conditions.

Bain’s research identifies a disciplined approach among leading organizations: begin with tasks that can deliver value under manageable conditions, then expand autonomy as evidence accumulates.

These organizations initially choose use cases with accessible data, manageable operational risks, and outputs that employees can verify at reasonable cost. Such tasks allow AI to produce useful results while teams learn how reliably it performs before delegating more consequential actions.

Analysis and synthesis tasks are useful starting points. Employees can use generative AI to search information, compare documents, prepare summaries, draft material, and support decisions while remaining responsible for checking the results.

Repeated use exposes recurring errors, data gaps, and the controls needed for more demanding applications. Teams can then use that evidence to decide whether the system is ready to perform additional actions independently.

This progression also improves financial planning. Rather than assuming immediate full autonomy, an organization can model several operating stages with different labor requirements, technology expenses, and expected benefits. It can measure realized returns at each stage and use demonstrated savings to support further expansion.

For enterprise buyers, this produces a business case grounded in the actual cost of supervision. For providers, it creates demand for implementation methods that account for workflow readiness, operating constraints, and the customer’s ability to capture savings.

Those requirements help explain why major technology companies are investing heavily in deployment expertise.

Software vendors are expanding their role in AI deployment

The persistent gap between expected and realized value creates a commercial problem for AI providers. Customers may purchase advanced technology yet struggle to integrate it into their operations or demonstrate enough value to justify broader adoption.

Providers have an incentive to address those obstacles because successful deployments can support additional consumption, larger contracts, and longer customer relationships. Implementation expertise is consequently becoming a more important part of the enterprise AI offering.

During roughly 10 weeks in the middle of 2026, the four largest AI vendors committed more than $9 billion to placing engineering resources directly inside customer organizations.

Two announcements illustrate the scale of those commitments.

These investments address a practical problem: purchasing AI technology is only one step toward achieving its expected financial benefits. Customers must still integrate systems, adapt workflows, establish controls, and change how employees perform their work.

AWS and Microsoft can benefit commercially when successful deployments encourage customers to expand their use of the companies’ AI platforms and related services. Their engineering commitments therefore combine implementation support with an interest in increasing adoption.

Deployment engineering moves closer to customer operations

AWS’s forward-deployed engineering unit and Microsoft’s frontier-company initiative bring technical expertise closer to the environments where customers use AI.

Forward-deployed engineering means engineers working directly with customer teams to solve implementation problems in their existing technology systems. This gives providers greater involvement in integration requirements, workflow design, and technical constraints that can determine whether a deployment succeeds.

The work can begin before implementation. A conventional enterprise software sales process often emphasizes product capabilities, technical compatibility, and expected productivity improvements. A process focused on business outcomes also examines how the customer’s work is performed and what must change for the investment to deliver value.

For example, a provider evaluating an AI deployment might review existing approval procedures, exception handling, data access, and operational dependencies before recommending an automation design. The assessment can identify activities where AI would accelerate useful work, along with procedures that would prevent the customer from capturing meaningful savings.

It can also reveal whether the customer has enough internal expertise, data access, and operational capacity to support the proposed deployment. Those findings should shape the implementation plan and the financial assumptions used to justify it.

A workflow requiring human approval needs an investment model that includes the associated labor expense. Limited data access may make information retrieval or analysis a more practical starting point than autonomous execution.

Identifying these conditions early helps providers steer customers away from investments whose operating requirements conflict with their financial projections.

The approach also gives vendors opportunities to collect evidence about implementation performance. Consistent assessment of workflow readiness and post-deployment results can improve future proposals and provide a stronger basis for expansion discussions.

Engineering commitments establish that AWS and Microsoft are investing in these capabilities. Their financial impact will depend on whether customer deployments achieve better operating results and higher value realization.

Governance makes controlled AI execution possible

As providers become more involved in business workflows, they also need to support the controls enterprises require when AI agents execute consequential actions.

Microsoft’s 2026 Copilot architecture, governed through Agent 365, illustrates this direction. It incorporates permission scopes, approval workflows, and execution logging into AI operations.

Permission scopes define which systems, information, and actions an agent can access. Approval workflows specify when a person must authorize an action before execution. Execution logging records what the agent did so organizations can monitor activity, investigate problems, and establish accountability.

Together, these controls allow enterprises to introduce AI into workflows where unrestricted autonomous execution would create unacceptable operational or compliance risks.

Consider a financial process in which an agent prepares a transaction but requires approval before submitting it. The organization can gain productivity benefits from automated preparation while retaining control over the final action.

That design can make AI deployment economically practical in a process where errors carry substantial costs. It also allows the organization to measure the continuing expense of human participation and evaluate whether greater autonomy would improve the financial result.

Microsoft benefits commercially when governance capabilities help customers adopt Copilot and related AI services across more business processes. Customer value depends on whether those controls enable useful automation while keeping operating risks and supervision costs manageable.

ServiceNow is addressing a related requirement through its AI Control Tower, which provides measurement and governance within existing enterprise environments.

At Knowledge 2026, ServiceNow demonstrated the system identifying a prompt-injection attack against a pricing agent and shutting down the agent in real time.

A prompt-injection attack occurs when malicious instructions enter content processed by an AI system, with the aim of redirecting its behavior. In a pricing workflow, an attack could attempt to influence how an agent evaluates information or performs an authorized action.

The demonstration showed how a governance system could detect suspicious behavior and intervene during execution. These capabilities support oversight by allowing organizations to monitor agents, stop problematic actions, and preserve information for investigation.

ServiceNow has a commercial interest in making AI governance a capability enterprises purchase and use across their existing systems. Customers have a corresponding interest in reducing the operational risks that could prevent deployment.

These interests meet around a practical requirement: organizations need enforceable controls before allowing agents to perform sensitive work. Legal and compliance reviewers need evidence that operations can be governed, while technology leaders need ways to manage automated actions across business systems.

C-suite buyers also need to establish accountability when an automated decision causes a problem. Governance capabilities make those responsibilities explicit, which can affect whether a proposed deployment receives approval.

Existing data environments shape implementation costs

Governance becomes more useful when AI can operate within the systems customers already have.

Enterprise information is often spread across legacy applications, different data formats, and systems with inconsistent structures. Accessing it can require substantial integration work, increasing deployment costs and delaying financial returns.

Requiring complete modernization before introducing AI can make the initial investment especially expensive. A more gradual approach starts with available information and systems, then expands integration as the business case develops.

Providers can use lightweight connectors to access existing applications and data-normalization tools to make information from different systems usable within a common workflow. Data normalization means converting information into consistent formats and structures so software can process it reliably.

With those connections in place, customers can deploy a bounded AI capability, apply governance controls, measure performance, and decide whether further integration is justified.

ServiceNow’s AI Control Tower reflects this approach by supporting governance and measurement within customers’ existing environments. Its commercial appeal includes helping organizations introduce oversight while preserving earlier technology investments.

For enterprise buyers, gradual integration creates a more manageable investment path. A company can begin with a workflow its current systems can support, demonstrate value, and then decide whether additional modernization would generate sufficient returns.

This approach is especially relevant for businesses whose legacy procedures and fragmented data make immediate full autonomy impractical. They can pursue useful AI deployments while accounting for integration expenses, ongoing supervision, and the benefits achievable under current operating conditions.

For providers, supporting those environments expands the range of customers and processes where implementation is feasible. The ability to work with existing systems can therefore influence both deployment economics and vendor selection.

Consumption-based pricing changes how AI work is charged

While engineering and governance address implementation, software pricing is changing to reflect how customers use AI.

Traditional enterprise software contracts frequently charge by the number of users or seats. Customers pay for access, while the amount of productive work performed through the software can vary considerably.

Salesforce has introduced consumption-based pricing for Agentforce, connecting charges to AI agent usage. This creates a commercial measure of activity that can rise or fall with the amount of work customers assign to agents.

The company reports 3.8 billion agentic work units delivered across Agentforce and Slack. Agentic work units are Salesforce’s measure of tasks completed by AI agents.

For customers, this provides a more detailed view of AI activity than license counts alone. An enterprise can examine agent usage across supported workflows and compare that activity with operational performance.

Consumption-based pricing also connects increasing activity with higher software charges. Customers can see how usage affects spending, while Salesforce benefits commercially when customers deploy agents more widely and generate additional paid consumption.

But task completion and financial value are different measures.

An agent might complete work that employees could otherwise perform at negligible incremental cost. It could execute an inefficient workflow repeatedly or produce outputs requiring expensive review and correction.

The financial result depends on whether completed work reduces costs, increases capacity, improves quality, or produces another measurable benefit that exceeds the associated expenses.

Salesforce’s work-unit measure establishes the volume of activity under the company’s definition. Customers must connect those units to their own operating results to determine the value generated.

Consumption-based pricing therefore creates a closer connection between software charges and AI activity. Outcome-based pricing would go further by linking commercial terms to agreed business results, such as transaction costs, processing capacity, or other measurable improvements.

The distinction matters because greater usage can increase both the customer’s software bill and the provider’s revenue. Whether that activity improves the customer’s financial position requires a separate calculation.

Across these vendor initiatives, the commercial mechanisms differ. AWS and Microsoft are investing in deployment resources, Microsoft and ServiceNow are developing governance capabilities, and Salesforce is connecting charges to measured AI activity.

Each addresses a different part of enterprise adoption. Their economic value ultimately depends on the operating results customers achieve.

Measuring AI activity against financial value

The expansion of deployment services, governance capabilities, and consumption-based pricing gives enterprises more information about how they use AI.

Customers can observe engineering resources assigned to implementation, actions performed by agents, approval requirements, security interventions, and usage volumes. These measures help explain how a system operates, but financial evaluation requires connecting that activity to the business outcomes used to justify the investment.

The connection begins with the workflow itself. An organization needs to establish its operating costs and performance before deployment, measure the same process after introducing AI, and account for the resources required to run the new system.

Operational measures establish where value is created

A customer-service automation system, for example, can be evaluated through transaction costs, completion times, and error rates. An AI system supporting business decisions can be measured through decision throughput, accuracy, and the resources required for review.

These measures capture different aspects of performance, so organizations need to select those relevant to the workflow being automated.

The metrics become financially meaningful when linked to actual costs and benefits.

A shorter process may reduce labor expenses, increase operational capacity, or improve service. The organization must identify which effect occurred and calculate its economic contribution.

Similarly, lower error rates may reduce correction work, customer complaints, or compliance exposure. The financial benefit depends on the costs associated with those errors and the resources required to prevent them.

Decision throughput requires similar care. Processing more decisions can create value when additional capacity supports useful business activity. If demand remains unchanged, however, the financial benefit may depend on whether the organization can reduce operating costs or redirect employee time toward other productive work.

Hours saved require a financial interpretation

Conventional adoption measures such as licenses deployed and workflows automated help establish the scale of implementation. Hours saved add information about productivity, but their financial meaning depends on what happens to the released capacity.

An employee might complete a task faster while continuing to work the same number of hours. The organization has gained available capacity, but its immediate labor spending may remain unchanged.

Alternatively, the company might use that capacity to process more transactions, improve customer service, or reduce the resources required for existing work. Each outcome has a different financial effect and needs a corresponding method of calculation.

The distinction is especially important for supervised AI. If an agent reduces preparation time but introduces additional review requirements, the organization needs to measure the net change in labor.

Technology expenses must also enter the calculation. AI infrastructure, integration, consumption charges, monitoring, and maintenance can offset part of the operating benefit.

A reliable evaluation therefore measures the total cost of completing the workflow before and after deployment. It also records changes in output quality, processing capacity, and operational risk that affect business value.

These calculations give enterprise teams a stronger basis for deciding whether additional automation makes financial sense.

Providers can build financial measurement into deployment

Once customers establish relevant operational measures, software providers can help collect the information needed for ongoing evaluation.

A system can record workflow activity, calculate processing times, identify exceptions, and track the resources required for human review. Those records can then be connected to financial measures agreed upon during investment planning.

For example, an organization introducing AI into transaction processing could establish its original cost per transaction and expected reduction before deployment. After implementation, it could measure the new cost, including AI usage and supervision, and compare the result with the original projection.

If the deployment also increases processing capacity, the organization can evaluate whether that capacity produces additional business value. Changes in error rates can likewise enter the financial assessment through the resulting correction costs or savings.

The same method can support periodic financial reviews. Teams can compare expected and realized benefits, investigate differences, and adjust operating procedures before approving further expansion.

For CFOs and boards, this produces a clearer view of investment performance. For technology and product leaders, it identifies which workflows deserve additional automation and which require redesign.

Providers also gain commercially useful information. Demonstrating measurable customer outcomes can strengthen purchasing discussions, support contract renewals, and help justify expansion. These incentives make financial measurement relevant to both sides of the enterprise software relationship.

The strongest implementation evidence connects operational performance with the original investment case. A deployment that performs more work may be valuable, but the organization needs to establish the costs and benefits associated with that additional activity.

Procurement decisions can incorporate realized value

Financial measurement becomes particularly useful when an enterprise evaluates a new AI purchase or considers expanding an existing deployment.

The purchasing process can begin with a baseline: current workflow costs, processing times, error rates, capacity, and required controls. Buyers can then establish expected improvements and the operating assumptions behind the proposed investment.

Those assumptions should include the intended level of autonomy. If employees will continue approving transactions or reviewing outputs, their labor belongs in the financial model. Integration, governance, infrastructure, and continuing software charges should be included as well.

The resulting business case provides a benchmark for evaluating performance after deployment.

Once the system is operating, enterprise teams can compare measured results with the original projection. A realization rate can show how much expected value the organization captured, while detailed operational measures help explain the difference.

That evidence can guide the next investment decision. A workflow that consistently delivers strong returns may justify broader deployment, while one with disappointing results may require redesign, additional controls, or a different level of autonomy.

Vendor proposals can also be evaluated against these requirements. Buyers can examine how a provider assesses workflow readiness, handles integration, supports governance, and records the operating measures needed to calculate financial returns.

For a deal involving consumption-based pricing, the buyer can model how additional AI activity affects both operating benefits and software charges. For a deployment involving supervised agents, the evaluation can include the continuing cost of human review.

These calculations help establish whether greater AI usage improves the customer’s economics as deployment grows.

The same discipline applies to renewals. Customers can evaluate the measurable benefits delivered during the contract period and compare them with the total cost of operating the technology.

For providers, this creates a commercial opportunity tied to customer performance. Engineering assistance, integration capabilities, governance, and usage measurement become more valuable when they help customers achieve the returns used to justify their investments.

Recap

Enterprise AI investment is entering a period in which deployment success and financial performance must be evaluated separately. Companies continue expanding their AI budgets, but Bain’s research shows that much of the technology’s potential business value remains unrealized. Increasing adoption without addressing that gap risks extending the same financial assumptions across successive investments.

For CEOs, CFOs, and technology leaders, the priority is to establish how much value existing deployments actually deliver before committing additional capital. Realization rates provide a useful measure, but they must be supported by evidence of operating cost reductions, productivity improvements, and other measurable business outcomes.

The shift also changes what enterprises should expect from software providers. Investments by AWS, Microsoft, ServiceNow, and Salesforce demonstrate growing commercial attention to implementation engineering, governance, and usage measurement. These capabilities can improve deployment conditions, but they do not independently establish financial returns.

Business leaders should evaluate AI investments against the economics of complete workflows, including integration costs, human supervision, governance requirements, and continuing software expenses. Greater autonomy and higher usage are valuable only when their benefits justify the additional costs and risks.

The next phase of enterprise AI investment will depend less on how many systems organizations deploy and more on how consistently they convert those deployments into measurable results. Companies that connect funding decisions to realized value will have a stronger basis for expanding successful applications, correcting underperforming workflows, and negotiating with technology providers.

The central question for enterprise AI is no longer whether organizations can automate more work, but whether that automation produces enough financial value to justify continued investment.