Across 951 companies and 63 enterprise processes in Bain’s research, organizations are converting roughly 38% of AI’s potential into deployed, measurable value. We call this the realization rate: the share of expected value that ultimately translates into measurable business results.

For software providers, it may become one of the most important numbers in the AI market. Customers are no longer asking only what AI can automate; increasingly, they want partners who help them capture the value that they were promised. That shift is already reshaping how the largest vendors compete, and it’s turning the customer’s realization gap into the provider’s problem to solve.

The challenge isn’t unique to AI. After two decades of process mapping, integration work, and change management, traditional automation has realized only 52% of its potential value, whereas generative AI is realizing 38% across functions. AI is new technology, but it has inherited a familiar challenge: converting technical capability into business value.

Realization rates should drive AI investment

Most AI business cases are written like capital projects: a defined investment, a projected return, and a payback period. But generative AI and agents don’t work that way. They introduce ongoing costs, require integration and oversight, and often depend on workflows that were never redesigned for the technology.

The result is a mismatch between economic expectations and operating reality. Nearly 40% of companies that measured AI cost savings realized less than 10% despite targeting savings as high as 20% (see Figure 1). Even so, most are increasing AI spending, with many funding new initiatives using savings from earlier automation efforts that fell short of expectations.

Figure 1

A better approach is to evaluate AI investments against realization rates. If the last wave delivered 60% of expected savings, the next wave shouldn’t be funded as if it delivered 100%. If a GenAI business case assumes full automation but production still needs human approval, the expectations should reflect that. Increasingly, CFOs and boards will ask one question: How much of the expected value was actually realized? Companies that can’t answer may be trying to fund new AI investments with savings that were never achieved.

Avoid workflow debt by redesigning processes for AI

Similar to automation before it, AI doesn’t fix a broken process; it locks it in, speeds it up, and makes it more expensive to unwind. This workflow debt accumulates when companies digitize work without redesigning it, then treat the resulting process as the baseline for the next investment case.

Instead of asking where to deploy AI, teams should be considering how they would redesign a process from scratch given AI’s capabilities. That creates a path for real change and new savings. Otherwise, AI becomes a very expensive way to preserve the past.

Most AI is supervised, but the business cases aren’t

Only about 7% of companies run fully autonomous agents in production, yet many investment cases are built on the economics of full autonomy (see Figure 2).

Figure 2

In practice, most agentic processes still require human oversight. That’s often the right choice, but the cost should be reflected in the business case. Supervised agents can still deliver meaningful value through faster execution, better quality, and stronger decision support, even if they don’t deliver the economics of full autonomy.

This is why realization rates should be measured at the process level, not at the program level. Customer interactions, financial reporting, compliance, and internal auditing may all warrant keeping people in the loop. Those decisions can improve outcomes and reduce risk. But the financial implications should be clear. Otherwise, the company may end up approving one model while operating another.

What the winners did differently

Leading companies didn’t begin by automating their most complex processes. Instead, they started where data was accessible, operational risk was manageable, and people could check the work without eliminating the economic benefits. They used generative AI for analysis and synthesis before moving toward operational autonomy.

That sequence matters because enterprises tend to trust generative AI first when it helps people understand, summarize, compare, search, draft, and decide. Trust is earned through repeated use in bounded contexts. Once teams can see where the technology is reliable, where controls are needed, and where data gaps remain, they can move to higher-stakes workflows with more confidence.

Rather than treating generative AI as a connection of disconnected pilots, leading companies treated it as a progression. They captured realized value at every stage, reinvesting returns to fund the next set of use cases and basing each step on evidence from the one before.

Competing on realization

For technology providers, the realization gap is both a warning and an opening. Customers will keep spending on AI, but the buying conversation is changing. Boards and CFOs are becoming less interested in theoretical automation potential and more interested in realized business outcomes. Providers who help customers close the gap will be harder to displace than those still selling on projected value.

The market has already begun pricing this in. Over a period of roughly 10 weeks during the middle of 2026, the four largest AI vendors committed more than $9 billion to embedding their own engineers inside customer organizations, not to build better models but to make deployment work. AWS launched a $1 billion forward deployed engineering unit on June 30; Microsoft followed on July 2 with its $2.5 billion frontier company of roughly 6,000 engineers. The signal is clear: Once model capability converged, the source of advantage moved from the model to deployment.

Five imperatives separate the providers capturing this advantage from those still selling capability.

First, build workflow diagnostics into the sales motion. Before selling technology, help customers find the workflow debt (broken handoffs, approval loops, exception paths, and legacy rules no one would design today), and ask if the process would still make sense if built from scratch today. The providers institutionalizing a lightweight workflow-readiness assessment before implementation are the ones generating the outcome data that wins the next deal. In a market where capability claims are converging, that outcome data is the most powerful asset a provider has.

Second, build investment cases around what will actually run in production. With so few companies running fully autonomous agents today, business cases anchored in full-autonomy economics set up customers to miss their own targets. Help CFOs model staged autonomy, cost of people in the loop, governance requirements, and realistic adoption curves. Honest economics may shrink the first business case, but they raise the odds of a second. Even the largest vendors are designing for this reality: Microsoft’s 2026 Copilot architecture, governed through Agent 365, is built around permission scopes, approval workflows, and execution logging—an explicit acknowledgment that control and capability have to scale together.

Third, make realized outcomes a native part of the product. Program-level metrics (licenses deployed, workflows automated, hours saved) are no longer enough. Build outcome measurement into the platform: cost per transaction, cycle-time reduction, error rates, decision throughput, and financial impact. Salesforce has leaned hard in this direction, pricing its Agentforce agents on consumption rather than seats and reporting 3.8 billion agentic work units (its own measure of tasks completed by AI agents) delivered across Agentforce and Slack. The metric itself reframes the relationship around work delivered rather than seats sold, shifting the relationship from software access to documented outcomes. When CFOs begin auditing what automation actually returned, the providers with native outcome tracking will own the renewal conversation.

Fourth, meet customers where their data is. Data access and integration remain the biggest barriers to AI adoption. The providers capturing market share are not waiting for customers to modernize their data systems. They offer lightweight connectors, data-normalization tooling, and deployment patterns that let customers start with the data they have and expand over time. ServiceNow’s AI Control Tower reflects this approach. Rather than treating data readiness as a prerequisite, it provides a measurement and governance layer that works with customers’ existing environments. Waiting for the customer’s data to be perfect could mean allowing competitors to move in first.

Fifth, make governance part of the value proposition. As AI agents move into critical workflows, buyers are less interested in potential value and more interested in what happens when something goes wrong and who is accountable. Providers who build audit trails, escalation paths, and real-time oversight into the product accelerate the legal and compliance review that otherwise stalls late-stage deals. At its Knowledge 2026 event, ServiceNow demonstrated its AI Control Tower detecting a prompt-injection attack on a pricing agent and shutting the agent down in real time, a concrete picture of enterprise-grade governance as a product capability rather than a promise. Increasingly, that capability is a point of differentiation at the C-suite level.

The next phase of AI won’t turn on model performance; it will be won on realization. The enterprises in this survey will keep funding AI, but they will increasingly favor the leaders who can turn ambition into measurable results and who treat their customers’ realization gap as their own commercial problem to solve.