What Matters Most When Scaling AI in 2026

ReadStats & TrendsBest PracticesInfographic
Written By Jeff Winter

On May 21, 2026, I posted a LinkedIn poll asking what I knew was an unfair question:

What matters MOST when scaling AI? Not what matters. What matters most.

That distinction matters because the obvious answer is that all five options matter. Clean, contextualized data matters. Business ownership matters. Trust, governance, and risk matter. Workflow readiness matters. Reusable architecture matters.

Yes, you need all five. Thank you for attending the world’s shortest strategy session. But the poll was never about creating a complete list of requirements. It was about forcing a choice. In strategy, the hard question is rarely, “What is important?” The hard question is, “Which constraint is creating the most drag on the system right now?”

The poll received 466 votes:

  • Clean, contextualized data: 43.6%

  • Clear business value owners: 22.5%

  • Trust, governance, and risk: 20.8%

  • Workflows ready for AI: 8.4%

  • Reusable AI architecture: 4.7%

Before interpreting the results, one clarification is essential: this is not a ranking of overall importance. It is a ranking of what people selected as most important. Reusable AI architecture was not voted “fifth most important.” It was simply selected least often as the single most important factor. If I had asked respondents to rank all five, architecture could have easily landed much higher. That is why I find the result interesting. Each answer likely reflects a different organizational condition, maturity level, industry context, or painful lesson learned.

Clean, Contextualized Data: 43.6%

This was the clear top response, and I believe it was selected by people who have experienced the practical limits of AI inside messy enterprise environments. In my experience, companies rarely suffer from a total absence of data. They usually suffer from fragmented data, inconsistent definitions, uneven ownership, and missing context.

That distinction matters. Data existing somewhere in the enterprise is not the same as data being usable for AI. An ERP system may understand orders, inventory, and financial transactions. A manufacturing execution system may understand production activity. A CRM may understand customers and opportunities. A historian may capture time-series signals. A quality system may hold inspection results. But AI does not automatically know how those pieces relate to each other, which definitions are trusted, which records are current, or which context matters for a decision.

This aligns with IBM’s 2025 CEO research, where 68% of CEOs identified integrated enterprise-wide data architecture as critical for cross-functional collaboration, and 72% viewed proprietary data as key to unlocking generative AI value. That is not just a technology point. It is an operating model point. If the enterprise cannot connect and contextualize its own data, AI has a very small foundation to stand on.

I think the people who selected this option have probably lived through the painful stage where the AI conversation becomes a data-definition conversation. Which number is right? Which system is authoritative? Why does finance define it differently than operations? Why does the plant trust a spreadsheet more than the enterprise dashboard? Why does the model produce a reasonable answer that nobody believes? In manufacturing, this becomes especially visible. A machine temperature, downtime event, defect code, or cycle-time variance means very little in isolation. It only becomes useful when connected to asset state, product, recipe, material, shift, operator, process step, quality outcome, and operating conditions. That is the difference between clean data and contextualized data.

If this was someone’s top choice, it probably means their organization is still fighting foundational data maturity issues. My advice would be to avoid launching a broad, abstract “fix the data” program. Those often become expensive, endless, and disconnected from value. Instead, start with the highest-priority AI use cases and work backward. Define the decisions AI is meant to improve, then identify the minimum trusted, contextualized data foundation needed to make those decisions better. Data readiness should be driven by business value, not by a theoretical desire to clean everything.

Clear Business Value Owners: 22.5%

I see this answer as the “we have seen the movie before” response. The model works. The demo is impressive. The executive team is intrigued. The AI team is proud. The vendor is smiling. Everyone agrees it has potential. Then nothing really changes.

People who picked clear business value owners have probably watched AI become a technology-led science project. The use case may be technically valid, but no business leader truly owns the outcome. No one funds the operational change. No one commits to the KPI. No one manages adoption. No one decides what process changes, what behavior changes, or what old way of working should stop. The condition behind this answer is usually not technical immaturity. It is accountability immaturity. The organization may have smart people, good tools, and legitimate ideas, but AI is being managed as a capability instead of as a business transformation mechanism. This connects directly to BCG’s work on AI value realization. BCG reported that 74% of companies struggle to achieve and scale AI value, and highlighted that many of the challenges are tied to people, process, change management, workflow optimization, governance, and business capability rather than the algorithm itself. The practical implication is simple: if nobody owns the value, nobody owns the scale.

If this is the most important factor in your company, I would make one rule immediately. No AI initiative moves forward without a named business value owner. Not a steering committee. Not a passive sponsor. Not someone who likes the idea. A real owner who is accountable for the business result, the adoption path, the workflow impact, and the value realization. A useful test is this: if the AI works, whose metric improves? If that person is not directly involved, the initiative is probably not ready to scale.

Trust, Governance, and Risk: 20.8%

Why would someone select trust, governance, and risk as the most important factor? I believe the answer depends heavily on industry context. In regulated, safety-critical, cyber-sensitive, or reputationally exposed environments, AI does not scale because people are excited. It scales when people believe it can be used responsibly. The question in these environments is not just, “Can AI generate a useful answer?” It is, “Can we act on this answer with confidence?” That confidence depends on several conditions:

  • The organization must know what data the system can access.

  • The output must be explainable enough for the use case.

  • Human oversight must be clear where risk demands it.

  • Legal, cyber, compliance, and operational stakeholders must know their role.

  • The business must understand what happens when the AI is wrong.

  • The governance model must distinguish low-risk experimentation from high-risk deployment.

This also shows up in McKinsey’s 2025 State of AI research. McKinsey found that AI use is broadening, but most organizations are still in the experimentation or piloting phase, with nearly two-thirds not yet scaling AI across the enterprise. They also reported that 51% of organizations using AI had experienced at least one negative consequence, with AI inaccuracy among the most commonly reported issues. That matters because trust is not a soft issue. It is a scaling condition. If people do not trust the system, they do not change their behavior. They double-check everything. They keep the old process alive. They route around the tool. They may even praise the technology while quietly refusing to rely on it. This is why I do not view governance as the enemy of AI scale. Bad governance slows everything down. No governance creates chaos. Good governance creates confidence.

People who selected this option may work in manufacturing, healthcare, finance, energy, infrastructure, transportation, or any environment where a flawed recommendation can create safety, regulatory, operational, financial, or brand consequences. They may have seen AI ideas stall because legal or cybersecurity had no clear review path. Or they may have seen the opposite problem, where experimentation spread faster than the organization’s ability to control it.

If this is the primary constraint, the answer is not to create one massive approval process for every AI use case. That will only force people into workarounds. The better approach is risk-tiered governance. Low-risk productivity use cases should move quickly. High-risk use cases involving safety, customers, regulated decisions, sensitive data, or operational automation should receive deeper review. Governance should be designed as a scaling system, not a permission trap.

Workflows Ready for AI: 8.4%

This answer received fewer votes, but I think it may reflect a more mature stage of the AI journey. Early in the journey, companies ask whether AI can produce something useful. Later, they ask whether the organization can actually absorb that usefulness into the way work gets done. The recommendation is reasonable. The prediction is useful. The summary is accurate. The alert is timely. But it exists outside the workflow. It becomes another dashboard, another tab, another email, another thing someone has to remember to check.

The people who selected workflow readiness have probably seen AI create insight without changing behavior. The planner still plans the old way. The operator still follows the old process. The salesperson still manages the opportunity the same way. The quality engineer still investigates through the same routine. AI is present, but the work remains unchanged. If so many companies are struggling to scale value, and so much of the challenge is tied to people and process, then the workflow question cannot be treated as secondary. AI value is not captured when the model produces output. It is captured when that output changes a decision, action, or process.

If this is your most important factor, I would stop asking, “What can AI do?” and start asking, “Where does the decision actually happen?” Map the workflow. Identify the decision point. Understand who makes the decision, when they make it, what information they use, what action they can take, and what changes if AI is introduced. If the workflow does not change, the value probably will not either.

Reusable AI Architecture: 4.7%

Reusable AI architecture was selected least often as the single most important factor, but I would be very careful with that interpretation. I do not think this means architecture is unimportant. I think it may mean architecture is often not felt as the first constraint. It tends to become obvious later, after early experimentation creates enough success to expose the limits of custom work. A company starts with a promising use case. The team moves quickly. They build the pipeline, connect the data, choose the tool, configure the model, create the interface, and prove something valuable. Then another team does the same thing differently. Then another. Then another. Each use case has its own integration pattern, security model, data approach, monitoring process, vendor dependency, support model, and lifecycle management problem.

At first, this feels like speed. Eventually, it becomes fragmentation. That is why I think people who selected reusable AI architecture may be further along in their AI maturity. They are no longer asking whether one use case can work. They are asking whether the tenth, twentieth, or fiftieth use case can scale without rebuilding the foundation every time. This is where the IBM research becomes relevant again. If CEOs see integrated enterprise-wide data architecture as critical, then architecture is not merely a technical concern. It is part of the enterprise’s ability to collaborate, reuse capabilities, and turn proprietary data into scalable AI value.

When architecture becomes the bottleneck, the company starts to experience predictable symptoms. Costs rise. Security reviews repeat. Integration work gets duplicated. Model monitoring is inconsistent. Governance becomes harder. Teams build overlapping capabilities. Successful pilots become difficult to replicate. The organization realizes it has not built an AI capability. It has built a collection of AI exceptions.

If this is your most important factor, the advice is to build for reuse without suffocating experimentation. I would not recommend over-architecting too early because that can create a platform in search of a problem. But I also would not allow every use case to become bespoke. Companies need reusable patterns for data access, identity, security, model deployment, monitoring, integration, governance, and lifecycle management. The objective is not architectural purity. The objective is repeatability at a lower marginal cost.

The Strategic Interpretation

The most important lesson from this poll is not that clean, contextualized data is universally the most important factor in every company. It is that data was the most common answer when people were forced to choose one capability as the highest-leverage constraint. That matters, but it is not the same as saying the other four are secondary. My broader interpretation is that AI scaling is situational. The most important capability depends on the company’s maturity, operating model, industry risk profile, organizational structure, and prior failure patterns. If the data landscape is fragmented, data wins. If nobody owns the business outcome, ownership wins. If people will not trust or govern the system, governance wins. If work does not change, workflow readiness wins. If every use case is custom-built from scratch, architecture wins.

This is why I view AI scaling as a strategy problem before it is a technology problem. Strategy is about diagnosis, sequencing, and tradeoffs. It is about understanding which constraint is limiting the system and which intervention would create the greatest leverage.

AI scales only as far as the organization is ready to carry it.

References:

Jeff Winter
Next

Why AI Strategy Feels Like the World’s Hardest Escape Room