AI Implementation for Mid-Sized Organizations: Why Your Pilots Fail Without Cross-Functional Data Governance
- VCM Management
- Aug 19
- 7 min read
If your AI pilot looked promising in a controlled demonstration but struggled as soon as you connected it to real business operations, you are not alone.
You may have seen accurate outputs in a small test environment. Your team may have built a working prototype, secured enthusiastic leadership support and even identified a compelling business case. Then the questions started:
Which customer definition should the model use?
Why do Finance and Sales report different revenue figures?
Who owns the product data?
Can the AI system access information held in another department?
What happens when the source spreadsheet changes?
Who is accountable when the model produces an incorrect recommendation?
This is where many AI initiatives enter what business leaders increasingly recognise as pilot purgatory: too valuable to abandon, but too fragile, fragmented or risky to scale.
The technology is rarely the only problem. In many cases, the real barrier is the absence of cross-functional data governance.
The scaling gap is widening, and poor data is driving it
AI adoption is moving quickly. McKinsey’s State of AI 2025 report found that approximately 88% of organisations regularly use AI in at least one business function, while around 71% use generative AI.
But adoption is not the same as value.
Nearly two-thirds of organisations remain in experimentation or pilot mode, and only a small proportion report that AI is fully scaled across the enterprise. McKinsey also found that approximately 70% of organisations experience data-related difficulties, including governance, integration and insufficient training data.
Here’s where it gets interesting: your organisation does not need more AI experimentation if the data foundation cannot support the next stage. You need the operating model that allows a successful pilot to become a reliable business capability.
That means connecting strategy, data, technology, processes and people before you attempt to scale.
Your AI model may be working, but your organisation is not ready for it
An AI pilot often starts inside one department.
Operations develops a demand forecast. Customer Service tests an automated response assistant. Finance experiments with anomaly detection. Sales explores lead scoring. Each team has a legitimate business problem and may even have enough data to produce an encouraging initial result.
The difficulty appears when you try to connect the pilot to the wider value chain.
The Operations team may define an “order” as a confirmed purchase. Finance may define it as an invoiced transaction. Customer Service may include cancelled orders in its reporting because those cases still generate support activity.
The model cannot resolve these contradictions by itself.
AI is not a highly capable assistant if you give it conflicting instructions from six departments. It becomes more like a new employee handed six different process manuals, none of which explains which one takes priority.
The output may look intelligent. The underlying decision may still be unreliable.
Siloed datasets create invisible failure points
Siloed data is not always obvious. You may have integrated systems and dashboards but still lack shared meaning.
Common examples include:
Separate customer records across CRM, finance and service platforms
Product codes that vary between procurement, operations and sales
Different definitions of revenue, margin, delivery performance or churn
Critical spreadsheets managed by one individual or team
Data held in legacy systems that cannot easily connect to cloud platforms
Supplier information split between procurement, contract management and finance
Inconsistent historical data that makes model training unreliable
The issue is not simply that the data sits in different systems. The issue is that each function may have developed its own assumptions, controls and priorities.
When an AI pilot uses one department’s version of the truth, it may work locally. When you place it into a cross-functional workflow, its assumptions are exposed.
Cross-functional governance gives your data a shared operating model
Cross-functional data governance does not mean creating a large committee that delays every technology decision. For a mid-sized organisation, that approach would be expensive, slow and difficult to sustain.
Instead, you need a practical governance model with four characteristics:
Clear ownership
Shared definitions
Controlled access
Ongoing monitoring
Start by assigning named owners to your most important datasets. If customer, product, supplier and transaction data are essential to your AI use case, someone must be accountable for their quality, definition, access and change management.
That person does not need to personally maintain every record. Their role is to ensure that the right standards exist and that issues have somewhere to go.
Then establish data stewards within each relevant function. A Finance data steward understands financial records. A Sales steward understands customer and opportunity data. An Operations steward understands inventory, fulfilment and production information.
This is a federated model: common standards are established centrally, while responsibility remains close to the business context.
The definitions you do not agree on today will become tomorrow’s model risk
Before you build or scale an AI system, agree on the meaning of the data it will use.
Create a concise business data dictionary for the terms that matter most to your use case. Depending on the pilot, this may include:
Customer
Order
Product
Revenue
Cost
Margin
Delivery date
Churn
Supplier performance
Service resolution
You do not need to document every field in every system before starting. That can become an excuse for inaction.
Focus first on the data that drives the decision your AI pilot is intended to improve. Define the term, identify the source system, assign an owner and document acceptable quality thresholds.
For example, if your AI system predicts late deliveries, you need to agree on what “late” means, which delivery date is authoritative and how missing or amended dates are handled.
This is not administrative detail. It is part of the model design.
Governance must begin before the pilot, not after the failure
Here’s where most business leaders get confused: governance is often treated as a compliance activity that happens once an AI system is nearly ready for deployment.
By then, important decisions have already been made.
A stronger approach is to include governance in the use-case selection process. Before approving a pilot, ask:
What business decision will this system support?
Which functions are affected?
What data does it require?
Who owns each critical dataset?
Are the definitions consistent across departments?
What information is sensitive or restricted?
How will a human review the output?
What measurable outcome will determine success?
What would be required to move from pilot to production?
These questions help you distinguish between a useful experiment and an initiative with a realistic path to scale.
The NIST AI Risk Management Framework offers a useful structure through its four functions: Govern, Map, Measure and Manage. You can adapt these principles without adopting an unnecessarily complex framework.
A cross-functional governance team should make decisions, not just recommendations
Your governance group should include the people who understand the full impact of the AI use case.
Depending on the application, that may include:
A senior business sponsor
Data and technology leadership
Representatives from affected departments
Information security
Legal, privacy or compliance
Risk management
Operational users of the system
The group should have defined decision rights. It should be able to approve, pause or reject an AI use case based on business value, data readiness and risk.
It should also maintain a simple AI system inventory. Record the purpose of each system, its owner, the data it uses, the vendor involved, the risk level and the stage of deployment.
This matters because shadow AI is already appearing across many organisations through software with embedded AI features. If you do not know which tools are being used, what information they receive or who relies on their outputs, your governance model has a significant blind spot.
Your 90-day roadmap should create the conditions for scale
You can establish a credible foundation for AI implementation for mid-sized organizations in three focused phases.
Days 1–30: Identify the data and assign responsibility
Begin with your highest-value AI use case. Map the systems, spreadsheets, APIs and manual processes that support it.
Document:
Where the data originates
How it moves between teams
Who owns it
Where definitions conflict
Which fields are incomplete or unreliable
What access restrictions apply
At the same time, create your initial AI inventory and nominate an executive sponsor.
Days 31–60: Establish shared standards
Create the minimum policies needed to operate safely and consistently.
These may include:
Data classification rules
AI acceptable-use guidance
Dataset ownership and stewardship responsibilities
Common definitions for critical data elements
Vendor assessment requirements
Human review and escalation procedures
Keep the documents concise and usable. A policy that employees understand and follow is more valuable than a comprehensive document that sits unread in a shared folder.
Days 61–90: Build governed data into the pilot
Now test the pilot using curated, documented and governed data rather than raw extracts from individual systems.
Introduce basic controls for:
Data completeness
Accuracy
Consistency
Timeliness
Access
Lineage
Model performance
Output quality
Data and model drift
Define the conditions under which the pilot can move into production. If those conditions cannot be measured, the pilot is not yet ready to scale.
The business case is stronger when governance is connected to value
Let’s talk money.
A pilot that produces an impressive demonstration but cannot be integrated into daily workflows is not delivering a return. It is creating another technology asset for your organisation to maintain.
Cross-functional governance improves the business case by helping you:
Reduce duplicated data preparation
Avoid building separate solutions for the same problem
Improve confidence in AI-generated recommendations
Reduce rework caused by inconsistent information
Identify integration requirements earlier
Assign accountability for measurable outcomes
Create a repeatable path from pilot to production
It also helps you connect AI to the wider value chain. A forecasting model affects procurement, inventory, production, fulfilment and customer experience. A customer service assistant affects data quality, employee capability, privacy and brand reputation.
The value does not sit inside the model. It emerges from the connected system around it.
Your next step is not another pilot
If your current AI initiatives are stuck, do not immediately assume that you need a different model or a more advanced platform.
Ask whether your organisation has:
A clearly defined business outcome
An accountable executive sponsor
Named owners for critical datasets
Shared definitions across affected functions
A documented data flow and access model
A cross-functional approval process
A measurable path from pilot to production
Ongoing monitoring for data and model performance
If several answers are “no”, your next investment should be in the foundation.
Value Chain Management helps organisations connect strategy, data and digital innovation with practical business outcomes. Our holistic approach to business transformation is designed to help you move from isolated initiatives to coordinated, resilient capability across your value chain.
The organisations that gain lasting value from AI will not necessarily be those with the biggest budgets or the most experimental pilots. They will be the organisations that make data a shared responsibility: and give every AI initiative a clear route into the business.
When you are ready to assess your data foundations and AI roadmap, contact Value Chain Management to discuss your next step.
Further reading
Comments