There’s a moment in every AI strategy discussion when someone inevitably suggests “starting small and seeing what works.” It sounds reasonable. Run pilots across departments, let innovation bloom organically, scale what succeeds. This approach has the comforting appearance of prudence—spreading risk, maintaining flexibility, avoiding premature commitment.

However, spreading AI efforts too thin increasingly leads to organizational irrelevance.

The uncomfortable truth emerging from the first wave of enterprise AI deployment is that marginal improvements scattered across many domains don’t add up to a competitive advantage. They compound into complexity, cost, and organizational fatigue. But the alternative—focusing resources intensively on a single domain—introduces a problem that most frameworks conveniently gloss over: How do you choose where to focus? And what if you choose wrong?

The Hidden Cost of Diffusion

Consider what happens when an organization decides to “democratize” AI across the enterprise. Marketing gets a content generation tool. Sales gets a lead scoring system. Customer service gets a chatbot. Engineering gets a code completion assistant. Each purchase is defensible. Each shows measurable time savings in isolation.

But despite these well-intentioned efforts, true transformation remains elusive.

Because each tool exists in its own silo, powered by generic models, producing outputs that must still be manually stitched into existing workflows. More insidiously, each tool requires adoption effort and workflow modification—but none delivers enough value to justify fundamentally reorganizing how work gets done. When economic pressure arrives, these are the first cuts.

A bank that deploys AI assistants for wealth advisors, mortgage processors, fraud analysts, and compliance officers finds that each group saves about thirty minutes per day. That’s real value—millions annually. But competitors capture the same millions with identical tools. The relative position hasn’t changed.

Compare this to an organization that uses those same resources to reinvent its underwriting process. Instead of giving underwriters an assistant, it rebuilds the entire workflow around AI, from data ingestion to risk assessment, decision explanation, and regulatory reporting. The AI doesn’t just augment human judgment; it restructures which questions are asked, which data are examined, and how quickly decisions are made.

In contrast, this focused, integrated approach creates a capability that competitors must fully replicate to keep up.

Why Shallow Deployments Fail

The economic argument for concentration has become sharper as organizations move from experimentation to production. Traditional enterprise software had fixed costs. Generative AI reverses this. Every query costs money. Every generated response consumes compute.

This creates a brutal reality: you cannot afford to deploy sophisticated AI everywhere. A chatbot serving 10,000 employees but answering trivial questions might cost hundreds of thousands monthly while delivering minimal value. Deep deployments processing high-value transactions, such as systems handling millions of inventory-optimization transactions, justify even the cost of expensive inference.

But there’s a deeper reason shallow deployments fail: they don’t generate data feedback loops. AI improves with use, but only if data from that use can be captured, structured, and fed back into the system. A shallow chatbot collects transcripts, but if those transcripts aren’t connected to resolution outcomes and product feedback loops, the data is noise, basically. A deep transformation rebuilds the entire architecture so that every interaction feeds into a unified knowledge system. The AI becomes smarter because the entire organizational system learns faster.

Scattered deployments break this flywheel. Each system optimizes its local function, but the organization as a whole does not learn.

Three Filters for Finding Your Target

To focus efforts effectively, apply three filters to guide your decision-making:

Infrastructure dependence. Domains requiring new infrastructure create natural moats. If transformation demands unifying fragmented data, rebuilding integration layers, or establishing new measurement systems, that upfront cost becomes a barrier to competition.

Manufacturing’s predictive maintenance exemplifies this. Success requires converging historically separate IT and operational technology systems, installing sensors on legacy equipment, and establishing real-time data pipelines. This is expensive and time-consuming. But once built, it creates platform advantages that extend beyond the initial use case. Contrast this with an off-the-shelf chatbot—competitors can replicate it in weeks.

Workflow centrality. Some domains are interconnected; others are isolated. Marketing is often central: insights from customer interactions inform product development, which shapes supply chain planning and influences financial forecasting. Accounting, while valuable, is often isolated. Better expense categorization doesn’t affect other departments.

Measurability timelines. Domains where value can be measured quickly create organizational momentum. Fraud detection in financial services is measurable almost immediately, creating political capital for longer-term investments. Conversely, domains where value only becomes apparent after years—such as product innovation and brand building—are risky because the organization may lose faith before results materialize.

Making It Actionable: A Four-Phase Approach

Here’s how to move from analysis to action:

Week 1-2: Rapid Domain Audit. Don’t spend months in analysis paralysis. Convene your leadership team and score your top five domains across the three filters. Use a simple 1-5 scale for each dimension. Any domain scoring 12+ (out of 15) merits serious consideration. This isn’t about finding the perfect choice—it’s about eliminating obviously wrong ones.

Week 3-4: The Readiness Reality Check. For your top two candidates, conduct a brutal assessment of organizational readiness. Does the data exist? Is it accessible? What infrastructure must be built first? Reckitt spent three months cataloging 300 discrete marketing tasks before deployment. That upfront mapping work is not optional—it’s the foundation.

If your highest-scoring domain requires two years of data remediation before AI deployment, you face a choice: commit to the infrastructure build as a multi-year bet, or choose a more modest domain where you’re better prepared. Both are valid. What’s invalid is pretending the infrastructure work doesn’t exist.

Month 2: The One-Page Business Case. Force yourself to articulate the bet in one page: What workflow are we reinventing? What’s the three-year value hypothesis? What infrastructure must we build first? What’s our kill criterion if this doesn’t work? This discipline prevents scope creep and creates accountability.

Month 3: The Pilot That Matters. Don’t pilot the technology; pilot the transformation. Choose one sub-process in your target domain, rebuild it around AI, and measure whether the new architecture works. L’Oréal didn’t pilot a chatbot; it rebuilt its entire customer diagnostic process. The pilot should test whether your organization can execute a deep transformation, not just whether AI works in theory.

Risk Mitigation: What If You Choose Wrong?

The fear of choosing the wrong domain paralyzes more organizations than actual wrong choices. Here’s how to structure the bet to limit risk:

Build reversible infrastructure first. Data unification, API layers, and measurement systems have value regardless of which AI applications ultimately succeed. A manufacturer that unifies IT/OT systems for predictive maintenance gains infrastructure that enables quality control, production optimization, and supply chain visibility—even if maintenance proves less valuable than expected.

Set explicit kill criteria before you start. Define what “not working” means with specific metrics and timelines. “If we haven’t reduced downtime by 15% within 18 months,” or “if adoption among target users remains below 60% after training,” gives you permission to exit without the sunk cost fallacy taking over.

Stage the investment. Don’t fund three years upfront. Fund infrastructure build first (6-12 months), then initial deployment (6 months), then scaling (12+ months). Each stage requires demonstrating progress to unlock the next round.

Monitor the ecosystem. If competitors are succeeding in a different domain, that’s data. One major bank focused on lending transformation while competitors succeeded with wealth management AI. After 18 months, they pivoted—the infrastructure they’d built for lending proved 60% reusable for wealth management. The “wasted” investment was actually the cost of validated learning.

The Metrics That Actually Matter

Traditional ROI metrics can be misleading in the early stages of deep transformation. Instead, track the following:

Year 1: Infrastructure completion rate. Are you hitting milestones for data integration, system connectivity, and foundational capabilities? A manufacturer deploying predictive maintenance should measure “percentage of critical assets with real-time data feeds,” not “downtime reduced” (which won’t move yet).

Year 1-2: Adoption velocity. What percentage of target users are actively using the system weekly? Is usage increasing? Low adoption despite technical success is your kill signal. When GitHub deployed Copilot internally, they tracked not just code-completion rates but also whether developers kept it enabled—that’s the real adoption metric.

Year 2+: Operational metrics shifting. Now, traditional measures should move. Manufacturing: downtime reduction, quality improvements. Banking: approval time reduction, loss rates. Retail: inventory turns, conversion rates. If infrastructure is complete and adoption is high, but these haven’t shifted, something is fundamentally wrong.

Year 2+: Competitive positioning. Are you pulling away from competitors on specific dimensions? Can you offer service levels at price points they cannot match? This is the ultimate test of whether concentration is building a moat.

Sustaining Commitment Through the Valley

The most challenging period is months 6 to 18, when costs are tangible, but benefits have not yet been realized. How can you maintain commitment during this phase?

Storytelling, not just scorecards. Share concrete examples of how the new system changes daily work. A maintenance technician solves a complex problem in ten minutes instead of two hours. An underwriter approves a complicated case without calling three supervisors. These stories sustain belief when dashboards still show red.

Celebrate infrastructure milestones. Treat completing data integration or reaching 80% user adoption as wins worth recognizing. If you only celebrate ROI, you’ll lose momentum during the build phase.

Visible executive commitment. The CEO or business unit leader should reference this initiative in every quarterly address. Sustained executive attention signals this is not a side project that might get cut.

Competitor intelligence. Track what competitors are doing. When they announce shallow deployments across twelve functions, reinforce why your concentrated approach will prove more durable. When they struggle to scale pilots, explain why your infrastructure investment avoids that trap.

The Uncomfortable Truth

The hardest advice remains: choosing where to concentrate is less important than committing to concentration itself.

Organizations spend months analyzing which domain is optimal, but the marginal benefit of choosing Domain A over Domain B is often smaller than the cost of delay. What matters more is marshaling the organization behind a single direction and sustaining that commitment through inevitable difficulties.

The analysis is necessary to avoid catastrophic choices: domains where you lack data, where regulation makes AI impossible, or where the value hypothesis is fundamentally flawed. But once you’re in the reasonable range, execution and commitment matter more than optimization.

The organizations that succeed are not those that make the perfect choice. They are those who chose decisively, invested fully, and persisted through the messy middle. The diffused approach is appealing because it avoids this test of conviction. You are never wrong because you never commit. You are also never right because you never concentrate enough to find out.

Begin by applying the three filters and conducting a rapid assessment. Make a decision, build the necessary infrastructure, track relevant metrics, and maintain visible commitment throughout the process. While there is no guaranteed answer to the concentration question, following a proven process is sufficient to start.