The real AI bottleneck is data ownership, not model choice

Many AI discussions start with the model.

Which model should we use? Should it be OpenAI, Anthropic, Microsoft, open source, local, fine-tuned, embedded in a platform, or part of an agent framework?

Those questions matter. But they are rarely the first bottleneck.

For most operational companies, AI value is limited less by model choice than by whether the organization knows who owns the data, where it comes from, how quality is controlled, and which decisions or workflows it should support.

Short answer: AI becomes useful when it can work with trusted data, clear ownership, controlled access, defined business rules and a real workflow. Model choice matters, but it comes after the organization understands which data is reliable enough to support decisions, automation or AI agents.

This is especially true for companies where daily work depends on ERP data, master data, planning data, customer information, supplier records, product data, documents and business rules.

In that environment, the model is only one part of the system. The real question is whether the business can trust the context the model is asked to use.

Why model choice gets too much attention

Model choice is visible. It is easy to compare vendors, features, benchmarks and licenses.

Data ownership is less visible. It sits in process maps, ERP structures, field definitions, approval routines, data-quality exceptions, access rights and unclear handovers between teams.

That makes it easier to talk about models than to answer harder questions:

• Who owns item data from a business perspective?
• Which customer record is the trusted one?
• What happens when supplier data is incomplete?
• Which fields must be validated before an order, invoice or item update can move forward?
• Who decides whether an AI output is good enough to use?
• Which data can an agent read, and which data can it write back to a system?

These questions are not side issues. They decide whether AI can safely support real work.

AI Sweden's material on the strategic data journey is useful here because it frames data governance around roles, quality and responsibility, not only infrastructure. Its complementary material on strategic data work defines concepts such as master data, data lineage, data quality dimensions, data catalogs, data owners and data stewards. The practical message is clear: organizations need to know what data means, where it comes from, who is responsible for it and how it should be maintained.

That is the work AI depends on.

Data ownership is a business responsibility

Data ownership is sometimes treated as an IT topic. That is a mistake.

IT can manage systems, pipelines, integrations, access controls and platforms. But the business usually owns the meaning of the data.

The business knows whether a product record is complete enough for sales, production, purchasing, finance, logistics or compliance. The business knows which customer or supplier fields create risk when they are wrong. The business knows which exceptions matter and which ones are acceptable.

That is why data ownership needs both mandate and operational understanding.

A data owner should be able to answer:

• What does this dataset represent?
• Which decisions or workflows depend on it?
• What does good quality mean in this context?
• Which rules control how the data is created, changed and approved?
• Which users, systems or agents should have access?
• What should happen when the data is wrong?

Without clear ownership, data quality becomes everyone's responsibility. In practice, that often means no one has the mandate to fix it.

For AI, this becomes a scaling problem. A pilot can survive with manual fixes and local knowledge. Production automation cannot.

The data problem becomes operational when AI acts

Poor data has always created problems. It creates reporting errors, manual work, duplicate checks, missed follow-ups and low trust in analysis.

AI raises the stakes because it can move closer to action.

An AI assistant may help a user find information or summarize a document. An AI agent may gather data, check rules, prepare a recommendation, create a draft update, call an API, or support a workflow in an ERP or connected business system.

That does not mean every process should become autonomous. It means the data foundation needs to match the level of risk.

AI Sweden's AI Implementation Spectrum makes a useful distinction between building, adapting, reusing and integrating AI. The report argues for a practical spectrum of approaches, where many organizations should reuse or integrate existing capabilities before building from scratch. It also points out that approaches such as RAG and agentic AI shift complexity into integration, data quality and governance.

That is an important point for business leaders.

If the right approach is to integrate existing models with company data and workflows, then model choice is not the central capability. The central capability is knowing which data, rules, systems and review points the AI can rely on.

ERP data is often the critical context

For many operational companies, ERP data is one of the most important foundations for AI.

Systems such as Infor CloudSuite M3 contain core business context: items, customers, suppliers, orders, inventory, finance, planning, prices, statuses, terms and process rules.

That data often decides whether an AI-enabled workflow can create real value.

If the item master is inconsistent, an agent can help identify issues, but it should not be trusted to automate changes without clear rules and review. If customer records are duplicated, AI may retrieve context, but the organization still needs a source-of-truth decision. If order data is spread across ERP, emails, PDFs and transport systems, the workflow needs integration and validation before automation can be trusted.

This is why AI readiness is closely connected to modern data platforms and ERP data governance.

Our article on what a modern data platform is describes the platform as a governed foundation for analytics, automation and AI. The practical value is not only better reporting. It is making operational data trusted, reusable and available to the right people, applications and workflows.

The same logic applies when comparing modern data platforms in 2026. The platform decision matters, but it should follow the business workload, governance model, integration needs and data ownership structure.

AI maturity is moving from experiments to operations

RISE's State of AI report describes a shift from AI experimentation to deployment, with applied AI moving into areas such as infrastructure, industrial processes, healthcare, governance, testing and long-term operations.

That shift changes what companies need to get right.

In an experiment, it may be enough to show that AI can classify, summarize, recommend or generate something useful.

In operations, the requirements are different:

• Can the system access the right data reliably?
• Can the output be explained or reviewed?
• Are permissions and sensitive data handled correctly?
• Who monitors quality and exceptions after launch?
• How will changes in source systems affect the workflow?
• Who owns improvement when the process changes?

This is where many AI initiatives slow down. Not because the model cannot produce an answer, but because the organization is not ready to put that answer into a controlled business process.

What data owners should check before AI agents or automation

Before starting AI agents or broader automation, data owners and operational teams should answer a practical checklist.

1. Workflow

Which workflow should AI improve?

Be specific. "Improve operations with AI" is too broad. "Help the master data team identify incomplete supplier records before they affect purchasing and finance" is much easier to govern.

Define the process, owner, users, pain point, expected improvement and measurable outcome.

2. Source data

Which systems, documents and data fields does the workflow depend on?

Include ERP, CRM, planning, warehouse, finance, supplier portals, emails, PDFs, spreadsheets and knowledge bases where relevant. AI often needs both structured and unstructured data.

3. Source of truth

Where is the trusted version of each important record?

If the same customer, supplier, product or item exists in several places, define which source wins or how conflicts are resolved.

4. Ownership

Who owns the data from a business perspective?

This should include meaning, quality expectations, approval rights, acceptable use and accountability when something changes.

5. Quality rules

What does "good enough" mean for this workflow?

AI Sweden's strategic data material describes data quality through dimensions such as accuracy, completeness, consistency, uniqueness, timeliness and validity. Translate those dimensions into the fields and records that matter for the workflow.

6. Access and permissions

Who or what is allowed to read, use, combine or update the data?

An agent should not give users access to information they would not otherwise be allowed to see. Access should follow role, purpose, sensitivity and workflow risk.

7. Business rules

Which rules should the AI use when it prepares or recommends an action?

Rules may come from ERP configuration, process documentation, finance policies, customer agreements, product structures, compliance requirements or local operating practice.

8. Human review

Where should a person stay in the loop?

Many AI workflows should start as decision support or human-in-the-loop execution. The agent gathers context, validates information and prepares the action. A person reviews before anything is written back to a business system.

9. Monitoring

How will the workflow be monitored after launch?

Track output quality, errors, exceptions, usage, time saved, manual overrides, data-quality issues and improvement needs. AI in operations needs ongoing ownership, not only a launch project.

10. Reuse

Will this work create reusable data capability?

A good AI pilot should improve more than one isolated use case. It should create better definitions, pipelines, data products, access rules or governance routines that future workflows can reuse.

Where Elvenite fits

Elvenite works with Data Intelligence as a business discipline, not as a dashboard factory.

That means connecting data strategy, ERP data, data platforms, master data, governance, automation and AI-enabled workflows to practical business value.

For companies running operationally complex processes, the starting point is often close to ERP and daily work. Infor CloudSuite M3 may hold the core transaction and master data. Documents, emails, customer inputs, supplier records, planning signals and other systems may hold the surrounding context. AI becomes useful when those sources can be connected, governed and used in a controlled workflow.

Our existing guide to AI agents explains the broader concept. The article on AI agents in Infor M3 for order management, item data and customer setup shows the same point in a more concrete ERP context: agents need process logic, the right data, business rules and the right level of human oversight.

That is why the first AI question should often be a data ownership question:

Can we trust the data enough for this workflow, and does someone own the responsibility for keeping it trustworthy?

If the answer is unclear, the next step is not another model comparison. It is to strengthen the data foundation, ownership model and workflow governance around the use case that matters.

FAQ

Why is data ownership important for AI?

Data ownership makes responsibility visible. AI needs clear rules for what data means, where it comes from, how quality is controlled, who can access it and who approves changes. Without ownership, AI outputs may look useful but be hard to trust or act on.

Is model choice still important?

Yes, but it is usually not the first bottleneck. Model choice should follow the use case, data sensitivity, integration needs, performance requirements, cost and governance model. For many business workflows, the bigger issue is whether the model can access reliable, well-governed business context.

What data is most important before starting AI agents?

The most important data depends on the workflow. For operational companies, it often includes ERP data, master data, customer and supplier records, item and product data, orders, invoices, planning inputs, documents, emails, business rules and internal knowledge.

How do companies make ERP data AI-ready?

Start by mapping the workflow, identifying required ERP fields and surrounding data sources, assigning business ownership, defining quality rules, documenting source-of-truth decisions, setting access permissions and deciding where human review is required before system updates.

Should AI agents be fully autonomous?

Not by default. Many workflows should start with decision support or human-in-the-loop execution. Selected autonomous execution only makes sense when data quality, permissions, business rules, monitoring and risk level support it.

What is the best first step for leaders?

Choose one important workflow where better data and automation could create visible value. Then map the data, ownership, quality rules, access needs, review points and success metrics before choosing the AI implementation approach.