← Insights
AI · · 11 min read
AI Strategy Beyond the Pilot: What Boards Actually Need to Know
By Joseph Antoun
AI creates value when technical promise is matched by operating readiness. Boards need a sharper way to judge when that condition is in place.
Enterprise interest in AI continues to rise, and investment keeps pace. Large-scale value, meanwhile, still sits with a relatively small group. MIT NANDA's 2025 research found that only a small share of enterprise AI initiatives reach production and deliver material value, while BCG's 2025 global study found that only 5% of firms are achieving AI value at scale and 60% still report minimal revenue and cost gains despite substantial investment. Gartner's January 2026 analysis adds a practical explanation for that gap: weak business value, fragile data foundations, rising cost of ownership, immature responsible AI controls, and limited change management continue to undermine progress after proof of concept.
Taken together, these findings reveal a familiar pattern in which technical activity is rising quickly, while durable value is emerging more selectively. The distance between those two realities has become one of the defining leadership questions in AI.
That distinction matters more than many board packs currently reflect. A pilot can show that a model works. It can show that a workflow can be augmented. It can even produce a convincing demonstration. What it still leaves open is a larger question: is the organization ready for the consequences of using that capability at scale? Those consequences include operational dependence, data pressure, control requirements, new accountability lines, process redesign, and real economic exposure.
For boards, that changes the center of gravity. The first question often focuses on technical performance, which is understandable, since a pilot must prove that the capability can do something useful. The more useful question, however, focuses on readiness: are we prepared for what successful use creates next? That is where oversight becomes more valuable, and where AI governance starts to mature.
Technical success is the entry ticket. Organizational readiness turns it into value.
The Pilot Trap
This is where many organizations meet the real challenge, because a technically successful pilot shows that the technology can perform, while the next phase asks something broader and more demanding: can the organization carry that capability inside day-to-day operations, where speed, complexity, accountability, and consequence all become real at once?
At first, the story usually looks encouraging. The demonstration is strong, the sponsor is enthusiastic, and the team can point to faster outputs, better summaries, sharper recommendations, or smoother handling of a task. Then the capability meets real life, and the shape of the problem changes. Ownership becomes diffuse. Integration grows heavier than expected. Data quality surfaces quickly. Exceptions multiply. Users create workarounds. Legal and risk teams join later in the process, and costs rise faster than planned. What felt like momentum starts to feel more fragile because the pilot has moved from a controlled setting into the texture of real business life, where systems, people, rules, and incentives all press on the same workflow.
MIT CISR offers a useful lens for this moment by separating two very different modes of generative AI use. One centers on widely available tools that improve individual productivity. The other centers on tailored capabilities embedded in processes, systems, and offerings in ways that create business value at scale. The first can generate visible local gains quite quickly. The second demands far more of the organization, because enterprise value depends on integration, ownership, controls, and operating fit. That distinction helps explain why early enthusiasm and enterprise value so often diverge.
At board level, this difference sharpens the decision. A successful pilot deserves attention because it proves potential, yet readiness for exposure determines whether that attention should translate into scale, sustained investment, and operating dependence.
The Real Question Boards Should Be Asking
Once that difference becomes clear, AI strategy becomes easier to place in the right frame, because the issue is less about maintaining a long list of experiments and more about making explicit business choices about where AI matters, where automation creates the stronger answer, and where process redesign should come first. Those choices shape value far more than a broad portfolio of loosely connected pilots, especially when those pilots are evaluated in isolation rather than as part of a larger operating design.
AI therefore belongs inside the normal disciplines of capital allocation, operating design, control, and accountability. It is an enabling capability, powerful and often transformative, yet still subject to the same core question as any serious investment: where will it create durable value, and what conditions will allow that value to hold?
BCG's 10-20-70 principle helps restore proportion here. Leading organizations devote roughly 10% of effort to algorithms, 20% to data and technology, and 70% to people, processes, and cultural transformation. That distribution carries an important implication for boards, since a large share of value sits in the operating environment around the model: in roles, workflows, incentives, controls, adoption, and execution quality. Seen from that perspective, board oversight becomes clearer, because the central issue is whether the organization can convert promising technical capability into controlled, repeatable operating reality.
The Intervention Order
That question becomes sharper when the sequence is right, because before applying AI to a problem, organizations need an intervention order that protects both value and execution quality.
The intervention order
-
1
Clarify the business problem. Understand what is slow, expensive, inconsistent, or exposed, and what business outcome is at stake.
-
2
Simplify the process. Many organizations apply AI to a process that still carries avoidable complexity, which loads friction into the model from day one.
-
3
Automate what is stable. When the task is repetitive, rule-based, and predictable, conventional automation may create the stronger return.
-
4
Apply AI where judgment, language, prediction, variability, or pattern recognition justify it, which is where the technology can create leverage that simpler tools cannot.
This sequence matters because it keeps the organization focused on leverage rather than novelty. In some cases, the most valuable early use of AI is diagnostic: documenting how work actually happens, surfacing variation, revealing broken handoffs, and creating the basis for redesign before heavier deployment begins. In that sense, AI can improve organizational understanding before it expands operational exposure, which is often a far better starting point than rushing straight toward scale.
This is also where AI strategy reconnects directly to operating model and data governance. Gartner reported in February 2025 that 63% of organizations either lack, or remain unsure about, the right data management practices for AI, and predicted that through 2026 organizations will abandon 60% of AI projects unsupported by AI-ready data. That matters because weak foundations become visible very quickly once AI enters live operations, and what looked manageable during a pilot can become costly once real decisions, real users, and real dependencies start to gather around the system.
60%
of AI projects unsupported by AI-ready data will be abandoned through 2026
Source: Gartner, Lack of AI-Ready Data Puts AI Projects at Risk
From Experiment to Exposure
Once sequence is clear, the next question is exposure. The more useful transition is experiment to exposure, because that language reflects what actually changes when AI moves closer to scale.
A pilot proves that something can work under limited conditions, yet a board still needs to know whether the organization is ready for the exposure that comes with using it in real operations. That exposure includes workflow dependence, control obligations, data risks, accountability shifts, vendor concentration, regulatory implications, and investment consequences. In other words, success increases consequence, and the board's role is to judge whether the organization is ready to carry that consequence with discipline.
That is why board-level AI review works best when it follows five tests.
Five board-level tests
-
1
Technical success. The capability has to work with credible quality. This is the opening condition for every subsequent decision because performance creates the basis for further judgment.
-
2
Business relevance and expected value. The next question is strategic and economic. What priority does the use case support, and what value is expected in concrete terms: cost, speed, quality, revenue, resilience, risk reduction, or customer impact? A use case with strong technical performance and weak economic logic belongs in a different category from a true scaling candidate because value discipline shapes where leadership attention should go.
-
3
Operational fit. From there, the focus moves into the workflow itself. Can this work inside real processes, with real users, real exceptions, real controls, and real consequences? This is where many pilots lose force, because contained success meets day-to-day complexity.
-
4
Scale path. Even when operational fit is promising, other questions remain. Is there a credible route from local success to operating reality? Who will support it, maintain it, govern it, and fund it beyond the pilot? What integration work, data work, and role redesign still sit between "promising" and "usable"? Scale becomes real when those questions have credible answers.
-
5
Accountability and control readiness. Finally, boards need clarity on ownership and consequence. Who owns the outcome? Who handles exceptions? What remains human by design? Where are the boundaries of use? These are operating questions with direct board relevance, because accountability is what turns capability into governable practice.
The more useful transition is experiment to exposure.
That framing improves the quality of board oversight because it moves the conversation from technical optimism to business readiness, and from enthusiasm to judgment.
What Boards Should Actually Be Receiving
Once management adopts that exposure lens, board reporting changes as well, because the board needs more than a catalogue of pilots and a general message of progress. What it should receive is a disciplined view of where AI is tied to business priorities, where readiness gaps sit, what value is expected and when, what operating changes are required, who owns the outcome, and what management recommends for each priority use case: stop, redesign, or scale.
At a practical level, that means six elements in every serious board pack:
- A set of priority AI use cases mapped directly to business priorities, so the discussion starts with strategic relevance rather than technical curiosity.
- A view of readiness gaps by use case, covering process, data, controls, skills, operating model, and vendor dependency.
- Expected value and timing, stated in defensible business terms.
- A clear ownership and accountability map.
- A decision recommendation for each priority use case: stop, redesign, or scale.
- A standing review of what AI now makes newly possible for customers, operations, and decision-making.
That final element deserves special attention because it expands the board's field of view. Many organizations apply a risk filter to AI and keep the opportunity filter underdeveloped. Both matter. The risk filter asks whether the organization is ready for the exposure. The opportunity filter asks where the economics of value creation have shifted enough to justify a fresh move. Strong board oversight needs both, because protection and ambition belong in the same conversation.
BCG found that more than 60% of future-built firms rigorously track AI value, compared with 17% of stagnating companies. That gap points to governance discipline more than talent alone, which is precisely why board quality matters so much in this phase of AI adoption.
5%
of companies generate substantial AI value at scale
Source: BCG, Are You Generating Value from AI? The Widening Gap
The Board's Real Job
Taken together, these elements clarify the board's role. The board governs the business consequences of using AI, and that responsibility includes capital allocation, operating readiness, accountability, risk posture, value realization, and exposure management.
Article 4 of the EU AI Act requires providers and deployers of AI systems to ensure a sufficient level of AI literacy for staff and others dealing with the operation and use of AI systems on their behalf, and the European Commission's guidance confirms that this obligation already applies. AI oversight therefore sits firmly within mainstream management discipline, which means boards are not looking at a side topic or a temporary innovation issue, but at a core part of how the organization is governed.
BCG's 2025 research also shows what strong discipline looks like in economic terms. The top 5% of AI-future-built firms are pulling away materially from laggards, and BCG ties that advantage to clearer business choices, stronger workflow redesign, leadership engagement, and a stronger tech and data foundation. In other words, value leadership comes from management quality expressed through AI, rather than from AI floating above management quality like some magical object executives can admire from a safe distance.
Technical success marks the start. The board's real job is to judge whether the organization is ready for what success creates next, because that is the threshold that matters most once a pilot begins to move toward scale.
50%
at least half of GenAI projects ended after proof of concept by the end of 2025
Source: Gartner, Why 50% of GenAI Projects Fail
Closing Note
This article extends the same idea explored in Data Governance Is an Operating Model Problem, Not a Technology One: durable value depends less on the promise of the tool than on the conditions around it. In data governance, that means decision rights, accountability, workflow integration, and operating discipline. In AI, it means the same logic applied to a new source of speed, judgment, and exposure. The technology matters, and the operating model decides whether that value holds under real business pressure.
← Insights
Data · · 10 min read
Data Governance is an Operating Model Problem, not a Technology One
By Joseph Antoun
Governance frameworks fail when they are designed in isolation from how the organisation actually works. The real challenge is alignment, accountability, and decision rights, not tooling.
Adoption is rising. Research by Precisely and Drexel University's LeBow College of Business found that 71% of organisations now have a data governance program, up from 60% in 2023. Yet Precisely also reports that 67% of organisations do not completely trust the data they use for decision-making.
Source: Precisely + Drexel, 2025 Outlook (71%) · Precisely, Data Quality Challenges (67%)
The gap is revealing. Many organisations have launched, funded, and formalised governance. What they have not done is design an operating model people will actually use.
The result is familiar: policies exist, councils meet, standards are documented, tools are in place, and the business still struggles with inconsistent definitions, unresolved ownership, and slow decisions.
Governance works when it becomes part of how the company runs. It stalls when it arrives as a structural exercise before it solves a concrete business problem.
71%
of organisations report having a data governance program, up from 60% in 2023
67%
do not completely trust the data used for decision-making
Sources: Precisely + Drexel, 2025 Outlook (71%) · Precisely, Data Quality Challenges (67%)
Why governance stalls
Most programs lose momentum for a simple reason: they add process before they remove friction.
Teams are asked to classify, define, document, validate, escalate, and comply, often before they see any improvement in delivery, decision quality, or risk reduction. That is when governance starts to feel like overhead.
Standards exist, but delivery teams bypass them through exceptions. Councils exist, but decisions do not stick because authority is unclear.
Quality appears in dashboards, but no one owns the root cause. The formal model says one thing. The real operating model, revealed by workarounds and unresolved escalations, says another.
Gartner predicts that by 2027, 80% of data and analytics governance initiatives will fail due to a lack of a real or manufactured crisis, arguing that a governance program not tied to prioritised business outcomes will not sustain itself. Separately, Gartner says that 59% of organisations do not measure data quality.
Source: Gartner, 2024 D&A Governance press release · Gartner, Data Quality
This is where many programs go wrong. They start large, become abstract, and drift away from the problem they were meant to solve.
Leaders also underestimate the adoption burden. People adopt governance when they can see why it matters and how it helps them work with more clarity.
80%
of data and analytics governance initiatives will fail by 2027
59%
of organisations do not measure data quality
Sources: Gartner, 2024 D&A Governance · Gartner, Data Quality
Governance as an operating model
Governance becomes durable when leaders treat it as a question of operating design. The central issue is whether the organisation has made the hard decisions: who decides, who owns, where governance happens, how exceptions are handled, and how performance is reviewed.
At Knowledge Sphere, we use a practical operating-model lens built around six components. It keeps the conversation focused on how governance works in practice, not just how it is documented.
The strongest models are usually hybrid. They combine central standards with federated execution, domain ownership with enterprise guardrails, and stronger central control only where data is critical or business risk is high.
That balance matters. Governance has to support both standardisation and autonomy, both speed and control.
Design should fit the organisation, not follow a generic maturity model. A smaller company may need clear ownership, a few critical controls, and lightweight review rhythms. A larger or more regulated organisation may require formal decision rights, escalation paths, and greater control over critical domains.
The principle is the same. The design is not.
Deloitte's 2024 Chief Data Officer survey found that 72% of CDOs report into the C-suite. 66% say they have improved business process efficiency and compliance, and 63% say they have improved strategic decision-making.
Source: Deloitte, Chief Data Officer Survey 2024
Data governance is agile, proportionate, and risk-based. It should scale with the company's size, complexity, regulatory exposure, data landscape, and business priorities.
The six components of a governance operating model
At Knowledge Sphere, we structure governance through six operating-model components.
| Component |
Definition |
| Decision rights |
Who decides definitions, priorities, thresholds, policies, and exceptions. |
| Ownership & accountability |
Who owns the data issue, the remediation, and the business consequence, not just the reporting. |
| Workflow integration |
Where governance is embedded into delivery, operations, and change, making it part of normal work rather than a parallel exercise. |
| Controls & exceptions |
Which rules are mandatory, which are flexible, and how deviations are reviewed and resolved. |
| Enablers |
Technology, metadata, catalog, lineage, automation, and supporting mechanisms that make governance easier to execute. |
| Metrics & cadence |
How progress, adoption, trust, exception volume, and time-to-resolution are measured and reviewed. |
Many programs jump straight to enablers, buying the platform, defining the taxonomy, and assuming execution will follow.
The enabling layer only performs when the operating model is clear enough to support it. Tools strengthen structure, visibility, and automation. They do not create sponsorship, resolve political ambiguity, or make accountability appear.
Why a technology-first approach underdelivers
A technology-first approach underdelivers because it mistakes visibility for resolution.
A catalog can show that definitions vary. A lineage tool can show where data flows. A quality platform can flag failing records. A policy tool can track approvals. All valuable, creating structure and scalability.
What these tools cannot do is settle cross-functional authority. They cannot decide whether HR, Finance, or Operations owns the authoritative definition of a workforce metric.
They cannot resolve disagreements about whether a rule is mandatory or situational. They cannot persuade busy teams to invest effort when the value is unclear.
Organisations that get value use technology as an enabler of a broader operating model.
Precisely and Drexel University found that the most common gains from governance are improved analytics, better data quality, and increased collaboration. Those are operating outcomes. They emerge when governance changes how people work together.
Source: Precisely + Drexel, 2025 Outlook
FTE across HR, Finance, and Operations
Take something as ordinary as FTE (Full Time Equivalent).
HR may define it around contracted headcount. Finance may use a cost-allocation lens. Operations may care about deployed capacity and productive availability. Each view can be valid in context.
The problem starts when reporting, planning, or AI use cases assume one universal definition, without any explicit decision about which one applies where. That is a decision-rights problem.
A catalog can surface the inconsistency and document the variants. Progress begins when the organisation decides which definition is authoritative for each use case, who owns the decision, and how disagreements are resolved.
The same pattern appears with replication issues. Data changes in one system, and downstream copies update late or inconsistently. The symptom looks technical. The root cause is often weak ownership, unclear controls, and no process for resolving breaks at source.
What leaders should fix first
At Knowledge Sphere, we start from business problems and use cases, then design the governance needed to support execution.
The Knowledge Sphere sequence
-
1
Identify business problems
-
2
Define use cases
-
3
Measure time, cost, risk, and ROI
-
4
Prioritise
-
5
Build governance
Start by identifying the three to five business problems or decision failures that matter most. Define the use cases connected to them. Measure in terms of time, cost, risk, and ROI. Prioritise. Build governance around those use cases first.
McKinsey found that transformations involving at least 7% of employees in initiatives or milestones are twice as likely to generate positive total shareholder returns. Most organisations involve only 2%.
Source: McKinsey
Prosci reports that projects with effective sponsors are 79% likely to meet objectives, versus 27% with ineffective sponsors. Separately, Prosci, citing Gartner research, notes that 74% of leaders say employees were involved in change initiatives, while only 42% of employees say they felt included.
Source: Prosci, Change Management Success · Prosci, Why Change Management Fails (citing Gartner)
For each priority use case, governance design should cover:
- Mandatory versus negotiable controls
- Decision rights
- Domain accountability
- Escalation path
- Integration into real workflows
- Metrics: adoption, exception volume, time-to-resolution, and agreed quality indicators
This sequencing matters. Workflow integration belongs in the operating model. Workflow tools belong in the enabling layer. One defines how work gets done. The other helps it scale.
AI raises the cost of weak governance
AI has made this more urgent.
More than three-quarters of organisations now use AI in at least one business function, according to McKinsey's 2025 State of AI survey. Yet only 21% have redesigned workflows, and only 28% say the CEO oversees AI governance.
Source: McKinsey, State of AI 2025
60%
of AI projects unsupported by AI-ready data will be abandoned through 2026
Source: Gartner
AI amplifies whatever quality, consistency, and control exist in the data environment. Weak governance now produces faster, more scalable, and more persuasive errors.
Deloitte reports that 91% of organisations still have only basic or in-progress AI governance structures.
Source: Deloitte
IBM reports that 43% of COOs identify data quality as their most significant data priority. The same IBM article notes that more than a quarter of organisations estimate losses above USD 5 million annually from poor data quality, with 7% estimating losses above USD 25 million.
Source: IBM, The True Cost of Poor Data Quality
Governance has to be operational, not ceremonial. Definitions need to be clearer. Ownership needs to be real. Lineage, control, and escalation need to function in practice.
Conclusion
Data governance gains traction when it is tied to business problems, embedded into work, and backed by real authority.
The organisations that get value treat it as operating discipline: useful, proportionate, and enforceable. They start where the business feels pain, not where the architecture diagram looks impressive.
The hard part is not defining governance in principle. It is designing an operating model that people will use, leaders will sponsor, and the business can defend in practice.
Sources
- Precisely + Drexel University, 2025 Outlook: Data Governance Adoption Has Risen Dramatically
- Precisely, 2025 Planning Insights: Data Quality Challenges
- Deloitte, Chief Data Officer Survey 2024
- McKinsey, Going all in: Why employee "will" can make or break transformations
- Prosci, Change Management Success
- Prosci, Why Change Management Fails
- McKinsey, The State of AI: How Organizations Are Rewiring to Capture Value, 2025
- Gartner, Predicts 80% of D&A Governance Initiatives Will Fail by 2027
- Gartner, Data Quality: Why It Matters and How to Achieve It
- Gartner, Lack of AI-Ready Data Puts AI Projects at Risk
- IBM, The True Cost of Poor Data Quality
- Deloitte Asia Pacific, AI at a Crossroads