A company can run a modern ERP, have years of lean improvement behind it and employ an excellent technology team—and still fail to make artificial intelligence part of the business.
That apparent contradiction is now common. The tools are available. The organisation has data. A demonstration has probably impressed a room full of managers. Yet months later, purchasing, planning, quality or maintenance still works in essentially the same way.
The remaining obstacle is often described as technical readiness. That explanation is comforting because it postpones the difficult decision. Leadership can wait for a better model, cleaner data or another vendor presentation. But for many bounded operational use cases, the technology is already mature enough. The unresolved question is whether the company is willing to change a real process, assign responsibility for the result and carry the change into daily work.
That is why enterprise AI transformation is a business decision before it is an IT project.
The technology question is no longer the whole question
“Is AI mature?” is too broad to be useful. No technology is mature for every task, environment or risk level. A general-purpose model should not be given unchecked control of a safety-critical line, and a forecasting system cannot compensate for missing operational history.
But this nuance should not become a universal excuse for inaction. Machine learning, computer vision, document extraction, forecasting and language-based assistants are already used in production in well-defined settings. The practical question is narrower:
Is a particular capability reliable enough, with the right controls, to improve a particular workflow?
That question can be tested. It has users, inputs, exceptions, risks and measurable outcomes. “Is AI ready?” does not.
The transformation data reveals the gap. In McKinsey’s 2025 global survey, 78% of respondents said their organisations used AI in at least one business function. Yet only 21% said their organisations had fundamentally redesigned at least some workflows. Among the 25 organisational attributes tested, workflow redesign had the strongest relationship with reported EBIT impact from generative AI (McKinsey).
Using a tool and changing the business are not the same milestone.
Why ERP and lean maturity do not automatically lead to AI transformation
ERP programmes standardise records and transactions. Lean methods expose waste, stabilise work and create a discipline of continuous improvement. Those investments are valuable foundations. They do not, by themselves, make the next decision.
AI changes the range of problems a company can address. It can interpret information that never fitted neatly into the ERP: drawings, emails, maintenance notes, images, old spreadsheets and patterns across thousands of variables. It can also change when a decision is made and who receives support at that moment.
That crosses organisational boundaries. A production-planning use case may touch sales forecasts, inventory, supplier performance and machine availability. A quality assistant may change an operator’s task, the escalation path and the evidence kept for an audit. IT can integrate and secure the system, but it cannot decide alone which operational trade-off the company should make.
This is where mature companies can become unexpectedly cautious. Their current systems work. Their responsibilities are established. Their KPIs reward local performance. A technically successful AI deployment may require changing all three.
The obstacle is not ignorance. It is the cost of making a consequential choice.
The pilot trap: proving a model while avoiding a decision
A pilot can be useful. It can test data availability, model behaviour, user interaction and integration assumptions before a company commits more capital. But “pilot” is also a convenient label for activity that has no route to production.
The warning signs are familiar:
- the use case was chosen because the data was easy, not because the business problem mattered;
- the sponsor wants innovation visibility but does not own the affected process;
- success means an accurate demo rather than a business result;
- frontline users see the solution only after it is built;
- production integration, monitoring, support and security are deferred;
- nobody has authority to change the workflow if the test succeeds.
Deloitte’s 2025 survey of 2,773 director-to-C-suite respondents across 14 countries found that more than two-thirds expected 30% or fewer of their generative-AI experiments to be fully scaled in the following three to six months (Deloitte). That does not mean pilots are pointless. It means experimentation is much easier to authorise than operating change.
RAND reached a similar conclusion from a different method. Its researchers interviewed 65 experienced data scientists and engineers about AI-project failure. Leading causes included misunderstanding the problem, building models around the wrong metrics or workflow context, insufficient data and focusing on the newest technology instead of a real user problem (RAND).
A model can pass its test while the project fails its company.
What leadership must decide
Executive sponsorship is sometimes reduced to approving a budget or opening a kickoff meeting. Real ownership is more specific. Leadership has to make at least six decisions that no technology team can make on its behalf.
1. Which business outcome matters enough to change work?
Start with a recurring decision or constraint: schedule instability, avoidable scrap, slow quotation, missed maintenance, excess working capital or claims leakage. Define the economic and operational consequence in terms the process owner already uses.
“Deploy AI in operations” is not an outcome. “Reduce the time planners spend reconciling order changes while maintaining service level” is.
2. Who owns the result?
The accountable owner should lead the function whose performance is expected to change. A plant manager, operations director, head of claims or procurement leader may sponsor the workflow. IT, data, security and legal remain essential partners, but ownership stays close to the value.
This prevents a common failure: the business supplies requirements, IT delivers a tool, and neither side owns transformation after launch.
3. What will change in the workflow?
Map the current sequence from signal to decision to action. Then specify where AI enters, what information it uses, what it recommends or performs, when a person reviews it and how exceptions return to the system.
If the answer is merely “employees can consult a chatbot,” the design is probably still at tool level. A production design explains what happens on Monday morning when a forecast changes or an operator rejects a recommendation.
4. What risk will the company accept?
AI produces uncertainty, and the acceptable level depends on the decision. A low-risk document classifier can operate with sampling and review. A quality release or safety action may require explicit human approval. Leadership must decide the boundary, with technical and risk specialists translating it into permissions, thresholds, logs and escalation rules.
Waiting for zero risk means waiting forever. Ignoring risk makes scale impossible. Governance is how a company moves between those extremes.
5. What evidence will justify scale?
Measure the business outcome, not just model accuracy or logins. Depending on the workflow, useful measures include cycle time, first-pass yield, forecast error, downtime, manual touches, exception rate, service level and working capital.
Transformation is also behavioural. Are people using the recommendation at the decision point? When do they override it? Are overrides improving the system, or being lost in email? A high login count can coexist with no operational impact.
6. What resources will continue after the demonstration?
Production AI needs monitoring, feedback, data stewardship, support, training and an improvement cadence. These are operating costs, not flaws in the technology. A sponsor who funds only the build has funded a demonstration.
A better route from AI pilot to production
The alternative is not a giant multi-year programme. It is a deliberately bounded production path.
- Diagnose the workflow. Observe how the decision is actually made, including spreadsheets, informal workarounds and tacit knowledge.
- Select one valuable boundary. Choose a problem meaningful enough to matter but contained enough to govern.
- Name the business owner and users. Give them authority to shape the workflow and the obligation to measure the result.
- Assess data and controls. Identify the minimum sources, permissions, quality checks and human-review points required.
- Build into the real environment. Use representative complexity, not a cleaned dataset that hides daily exceptions.
- Run with users and capture exceptions. Treat disagreement as information. It reveals missing context, weak rules and training needs.
- Decide explicitly: stop, improve or scale. Compare the evidence with the pre-agreed threshold. Do not let a successful test drift into indefinite limbo.
This is the logic behind our method: diagnose the operational constraint first, then design technology, data and change around the result the company needs.
You probably do not need to replace your ERP
Leadership commitment does not mean authorising another costly migration by default. In many companies, the ERP should remain the system of record. The AI layer can connect approved data from the ERP, spreadsheets, technical documents, email and machines, structure what is relevant and return intelligence to the point of work.
That architectural choice matters for established industrial firms. It preserves investments that already work and narrows the transformation to the workflow where value is expected. Data Core is Embiggen X’s approach to creating that governed, usable data layer across existing sources.
The leadership decision is not “replace everything.” It is “make this process work differently, and integrate the existing systems required to do it.”
The operating model behind enterprise AI transformation
Once several workflows move into production, isolated project governance becomes inadequate. The company needs a small but explicit AI operating model:
- a senior owner for the portfolio and value targets;
- business owners for individual workflows;
- shared standards for security, data, evaluation and human oversight;
- technical teams that can reuse integrations and monitoring;
- a route for employees to propose opportunities and report problems;
- regular decisions to stop, improve or scale initiatives.
This is not bureaucracy for its own sake. It reduces repeated debate and makes responsible deployment faster. McKinsey’s 2025 survey found that nearly two-thirds of respondents’ organisations had not started scaling AI across the enterprise, while only 39% reported enterprise-level EBIT impact. The small group classified as AI high performers was more likely to pursue transformative change and redesign workflows (McKinsey).
For SMEs, the constraints are sharper. The OECD identifies skills, finance, resources and digital infrastructure as persistent barriers to AI transformation, particularly among smaller firms (OECD). That makes prioritisation more important, not less. A mid-market company cannot afford twenty disconnected experiments. It can choose one consequential workflow, build reusable foundations and expand from evidence.
The decision that separates activity from transformation
The dividing line is not whether a company has purchased an AI licence or completed a proof of concept. It is whether leadership has authorised a new way of operating and accepted accountability for making it work.
That does not require blind confidence in AI. It requires a disciplined commitment:
- this problem matters;
- this leader owns the result;
- these people will shape and use the workflow;
- these controls define the safe boundary;
- these measures will tell us whether it works;
- this is how we will support it in production.
Once those decisions exist, technology teams can do their best work. Without them, even a mature technology remains a demonstration in the corner.
FAQ
Who should own enterprise AI transformation?
A senior business leader should own the portfolio and each deployed workflow should have an accountable process owner. IT, data, security and legal should co-design the solution and its controls, but the function receiving the business value should own the outcome and transformation.
Is AI transformation an IT project or a business transformation?
It requires both. The technical build is an IT and data discipline; selecting the outcome, redesigning work, changing roles, setting risk tolerance and measuring value are business-transformation responsibilities. Treating either side as optional creates a predictable gap.
How do you move an AI pilot into production?
Choose a real workflow, name its owner, agree on outcome metrics and risk limits before building, test with representative data and actual users, include production integration and monitoring in scope, and make an explicit stop/improve/scale decision at the end.
Does a company need to replace its ERP before adopting AI?
Usually not. A functioning ERP can remain the system of record while a governed data and intelligence layer connects the additional sources required by the use case. Replacement is justified only when the current system creates a specific constraint that cannot reasonably be integrated around.
How should leaders choose the first AI workflow?
Look for a recurring decision with measurable cost or delay, enough representative data, a willing process owner and a risk level that can be bounded. Avoid choosing solely because the demonstration will be easy.
How do you measure AI transformation beyond logins?
Measure behaviour and operating results: whether recommendations are used at the decision point, override patterns, exception resolution, cycle time, quality, downtime, service level or another metric tied to the original constraint.
Sources
- McKinsey, The state of AI: How organizations are rewiring to capture value, March 2025.
- McKinsey, The state of AI in 2025: Agents, innovation, and transformation, November 2025.
- RAND Corporation, The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed, 2024.
- Deloitte, The State of Generative AI in the Enterprise, January 2025.
- OECD, AI adoption by small and medium-sized enterprises, 2025.
If your AI activity is stuck between demonstration and deployment, book a free AI assessment to identify the first workflow worth changing.
