AI ROI Starts With the Process

AI initiatives can create direct productivity gains, but the process review they trigger may reveal a larger pool of value. This article looks at how to redesign the work first, use conventional automation for predictable steps, and apply AI where language, unstructured information, pattern recognition, or flexible reasoning creates a measurable advantage.

Aug 14, 2026 · Blaize Stewart · #ai-adoption #ai-roi #process-redesign #workflow-redesign #operating-model #automation #measurement

Most AI investment discussions begin with the technology. Teams compare model quality, license costs, token consumption, developer productivity, and the number of minutes an assistant can save on a task and tout these savings are true ROI. Those measures can be useful, but they can also narrow the problem too early. The larger economic opportunity more often appears when the AI initiative forces an organization to reopen the design of the process itself.

Most business processes were not designed once and left untouched. They accumulated steps and kruft. A manual review was added after an error. Before long, another handoff appeared when responsibilities shifted between teams. Over time, each change may have been rational on its own while the end-to-end process became slower, harder to understand, and more expensive to operate.

Dropping AI into that process can improve one or two steps without improving the process very much. A model may draft a document faster while the document still waits two days for an approval. An assistant may summarize a case in seconds while the case still passes through three queues. Local productivity may improve. but the constraint that determines cycle time, cost, or throughput remains somewhere else.

The real value appears when the entire process is gutted and remodeled. McKinsey's 2025 global AI survey tested 25 organizational attributes against reported EBIT impact from generative AI. Workflow redesign had the largest effect on whether respondents reported financial impact. Yet only 21 percent of respondents whose organizations used generative AI said they had fundamentally redesigned at least some workflows. The gap is significant. AI adoption is moving faster than the redesign of the work that surrounds it. Deloitte found a similar pattern in research on human and AI work design. Fifty-nine percent of surveyed organizations described their AI investment approach as technology focused, while only 16 percent reported fully redesigning roles, processes, and operating models. Technology-focused organizations were 1.6 times more likely to report that AI investments were not exceeding expectations. Deloitte's 2026 Human Capital Trends research also reports that organizations prioritizing intentional work design are twice as likely to exceed AI ROI expectations.

One of Deloitte's examples makes the mechanism easier to see. A European telecommunications company introduced an AI expert into customer service while leaving the broader workflow largely intact. Productivity improved by roughly 5 percent. During a broader rollout, the company spent much more effort redesigning how people and AI worked together. It changed workflows, trust thresholds, escalation paths, and training. Productivity improvement reached roughly 30 percent. One company cannot establish a universal benchmark, but the difference illustrates what process redesign can unlock. The AI capability contributed value in both cases, but the redesigned process created a much larger opportunity to capture it.

That revelation changes how an AI opportunity should be approached. The first question should concern the business outcome and the current process that produces it. Cycle time, cost per completed outcome, first-pass quality, rework, throughput, conversion, service level, and customer effort provide a better baseline than model usage. Once the process is mapped, unnecessary work becomes much more visible. Some approvals may no longer be needed. Some handoffs can disappear. Some inputs can be standardized. Some decisions can move closer to the work. Some exceptions can be handled earlier. Some steps can be removed completely.

This is also where conventional automation deserves more attention. AI is flexible, which makes it attractive, but flexibility is not always useful. A deterministic step with clear rules often belongs in ordinary software, a workflow engine, an API, a database constraint, or a rules engine. Those mechanisms are easier to test, cheaper to operate, and more predictable. AI earns its place where language, unstructured information, pattern recognition, variable context, or probabilistic reasoning creates an advantage that deterministic automation cannot provide economically.

Human work should be treated with the same scrutiny. Some decisions need accountable judgment. Some interactions need empathy, negotiation, or context that should remain with a person while some AI outputs may be suitable for direct execution at low risk, while others require review or escalation. Process redesign creates the opportunity to define those boundaries more explicitly instead of inheriting them from old workflows.

The ROI calculation also changes when the process becomes the unit of value. Ten minutes saved on a task does not automatically become ten minutes of economic value. The saved time may be absorbed by another bottleneck. A process-level measure captures what happened to the completed business outcome so it can show whether the redesign reduced actually total elapsed time, increased throughput, improved quality, lowered cost, released capacity, or improved the customer experience.

This approach can also expose value that has nothing to do with AI. A process review may discover that a report can be completely eliminated, that two approvals can be combined, or that an existing integration can remove an entire handoff. Those gains still belong in the business case even though no model produced them. The AI initiative created the reason to examine the work, but the examination uncovered a better process.

A practical way to apply this is to take one AI candidate before model selection and map the process end to end.

Establish the baseline outcome measures. Identify waiting, rework, handoffs, approvals, duplicated data entry, and unnecessary steps.
Remove or simplify what can be changed without AI. Automate predictable work with predictable mechanisms.
Assign AI only to the remaining steps where its capabilities create a measurable advantage.

The resulting design should be able to explain why every major step still exists, who owns it, how success is measured, and why AI belongs where it has been placed.

That exercise produces something more useful than an AI use case. It produces a better process with a clearer economic model. When AI adds value inside that process, the technology becomes part of the return instead of being asked to carry the entire business case.

Further reading