Amdahl’s Law states that improving a single part of a system can only speed up the whole system so much. Even if one step of a process becomes exponentially faster, the other steps impose a ceiling unless they, too, are accelerated.

This concept may explain some of the disappointment companies are experiencing with artificial intelligence. An AI assistant can draft a proposal in seconds, but it then sits for days waiting on reviews, pricing and approval. Similarly, a coding tool can produce software in minutes, but testing, security and deployment still take days or weeks.

Companies have accelerated specific tasks without reengineering the systems that surround them. This gap between inflated expectations and modest returns is what some are calling AI’s “valley of despair.”

After a period of vigorous adoption and experimentation, executives are now being confronted with ever-growing costs and limited proof of bottom-line impact. Analyst firm Forrester Research Inc. recently found that only 15% of AI decision-makers reported an increase in earnings before interest, taxes, depreciation and amortization over the past 12 months. Fewer than one-third could clearly map the value of AI back to bottom-line changes.

This skepticism is healthy; AI investments should face the same scrutiny as any other capital expense. Leaders need to be clear on expected returns, timelines and milestones that show whether their investment delivers measurable business results.

The productivity paradox

AI does have business-transforming potential, but small, localized productivity improvements aren’t game changers. Though giving employees access to general-purpose AI tools improves discrete tasks and builds fluency in the technology, this value is diffuse and disconnected from the metrics that determine business performance. This is Amdahl’s Law playing out at scale.

A more mature approach requires reassessing and reengineering how the business operates, applying AI to create fundamentally new operating processes. Whereas automation asks how existing tasks can be performed faster, reengineering asks whether that task should even exist.  Organizations often seek to automate what they already do rather than reinvent it.

Time saved doesn’t automatically become revenue or margin, so what does AI ROI actually look like? Generally speaking, organizations should target positive, measurable change in one or more of the following areas: revenue, efficiency or risk. A quantifiable business goal might be to reduce customer churn, shorten the lead-to-proposal cycle, improve forecast accuracy or detect and manage fraud faster.

Enterprises should reasonably expect firm ROI data within about six months. Though leading indicators such as scenario testing and pilot programs can be delivered quickly, large-scale, immediate returns are unrealistic. A thoughtful business process redesign can take 90 to 120 days to design and implement, followed by acceptance testing and, crucially, focused change management.

These initiatives are most likely to succeed when people accountable for business results are charged with delivering a specific outcome rather than simply achieving a technology go-live.

Costs also deserve more sophisticated treatment, since the expense of various agentic processes can vary drastically based on factors like use case complexity, model calls and context. Companies should measure and track the distribution of cost per outcome rather than relying on a pre-deployment average.

This is an important nuance: Early adopters find that monitoring how much each AI-handled case costs and how long it takes, and then refining their strategy to keep both numbers within an acceptable range, is crucial for successfully redesigning business processes with AI.

Avoiding the ‘valley of despair’

Treat AI as a series of targeted business transformations, not solely a technology deployment. That means pairing clear outcomes with disciplined execution.

Take a closer look at business processes. Examine how work actually happens within the organization, not just how a procedure “should” happen. Identify bottlenecks, handoffs, approvals, exceptions and rework sources before deciding where to apply AI. Expect to find substantial differences between documented processes and how work actually gets done.

Tie every initiative to a clear outcome. Name the revenue, efficiency or risk metric the AI initiative is intended to impact, establish a baseline, give a business leader ownership of the result and set a hard timeline for impact data.

Design pilots for production. AI pilots should test the path to sustained use, not merely demonstrate impressive capabilities under controlled conditions. Include operational data, integration, governance, security and change-management requirements from the beginning.

Measure economics, not activity. Track metrics such as cycle time, quality, adoption, cost per outcome and financial impact rather than seats, prompts or demos. For agentic systems, monitor both the range and the average of operating costs, since complex cases can consume far more resources than simpler ones.

Review and adjust regularly. Review regularly to decide whether to move forward, narrow the scope, redesign the workflow or stop. Exploratory work won’t always succeed, but it should continually produce evidence that improves the next capital allocation.

AI’s “valley of despair” isn’t inevitable. Companies that understand the value of investing time and capital to rethink and reengineer their business processes can turn AI spend into tangible outcomes.

Jeremiah Stone is chief technology officer at SnapLogic Inc. He wrote this opinion article for SiliconANGLE.

Image: SiliconANGLE/Reve