Artificial intelligence has no shortage of pilots.
Across the mid-market, employees are using copilots to prepare reports, summarize meetings and draft customer communications. Functional teams are testing assistants against policies and operating procedures. Technology groups are exploring agents that can retrieve information, interact with systems and initiate work. Some of this activity is sanctioned. Some are happening outside approved channels on personal accounts and corporate devices.
The volume of activity can make an organization appear further ahead than it is. In many cases, it demonstrates how easy modern AI tools are to access, not whether the business is ready to depend on them.
That distinction becomes clear when an experiment moves into everyday operations. It must connect to the systems and information on which the business runs, produce reliable outputs, and operate within defined controls. To deliver more than an isolated win, the business must repeat that process across functions, locations, and business units.
The technology may be new. The underlying transformation challenge is familiar.
What ERP transformations can teach us about scaling AI
ERP programs have often been described and managed as technology projects: select a system, configure it, migrate the data, train users, and go live. That sequence may install the software, but it does not, by itself, improve how the enterprise operates.
Companies that generate lasting value from ERP use the program to make broader operating decisions. They define how work should flow, who owns processes and data, which decisions belong at the enterprise level and where local variation is justified. Roles change, governance is established and the platform is integrated with the rest of the technology landscape. The software matters, but its value depends on the organization built around it.
AI requires the same breadth of thinking, although its delivery model should be different. A traditional ERP implementation may organize change around a major platform release. AI will develop through use cases introduced at different speeds and levels of risk while models, vendors and capabilities continue to change. A rigid, multiyear implementation plan would be a poor fit.
What carries over from ERP is the need to align organizational design, change management, governance, business process, data, and solutions architecture. Those disciplines determine whether a promising tool becomes a dependable business capability. Leaders cannot predict exactly where AI will be in three years. They can build an enterprise that is equipped to evaluate new capabilities, adopt the right ones and absorb the resulting change.
Why AI pilots rarely prove organizations are ready to scale AI
A narrowly defined pilot can succeed under unusually favorable conditions: a motivated team, selected data, manual oversight and temporary workarounds. Errors remain manageable because the output is not yet driving consequential decisions.
Those protections fall away in normal operations. Access and performance must be managed, source information must be current and exceptions need a defined route. Someone remains accountable when the system produces a poor recommendation or takes the wrong action.
The use case also begins to affect work beyond its original boundary. An AI tool that accelerates order intake may increase pressure on credit review, planning or fulfillment. An assistant that improves maintenance triage may expose weaknesses in asset data, parts availability or scheduling. Local gains can create downstream congestion if the surrounding process is not ready.
ERP programs have taught this lesson repeatedly: improving one function does not necessarily improve enterprise performance. AI raises the stakes because it can accelerate decisions and activity beyond the speed at which people can coordinate manually. A successful pilot provides evidence about a use case. It says much less about the organization’s ability to deploy AI repeatedly and safely.
The five foundations for scaling AI
Moving beyond the pilot requires five connected foundations. A well-designed use case can still fail because its data is weak, its owner is unclear, or the affected employees do not trust it. Strong governance without viable use cases, meanwhile, creates control without value.
Organizational change management
Public discussion about AI and work tends to focus on job loss. For many organizations, job redesign is the more immediate issue.
As AI assumes parts of research, analysis, monitoring and transaction processing, employees will spend less time assembling work and more time reviewing it. They will validate recommendations, resolve exceptions, judge ambiguous cases, and intervene when automated activity creates disruption. Managers may find themselves supervising a combination of people, systems, and AI-enabled workflows.
Each use case should assign who initiates the work, reviews the output, approves the action and handles exceptions, along with the actions the system cannot take autonomously.
Training should be based on the redesigned work. Prompting is a limited part of the requirement. Employees need to test answers, recognize weak recommendations, verify sources, document judgment, and escalate concerns. AI increases the value of sound judgment because it places more information and more potential action in the hands of each employee.
Governance and decision rights
Governance is often presented as a choice between unrestricted experimentation and a large compliance apparatus. Most mid-market organizations need neither. They need clear ownership, proportionate controls and timely decisions.
An executive sponsor sets the ambition and risk appetite. Business leaders own the operating outcomes. Process and data owners protect the integrity of their domains. Technology and security leaders establish platform, access and monitoring standards. Legal, risk and change leaders participate according to the use case.
This structure only works when its decision rights are explicit. Leaders need criteria for approving pilots, moving them into production, requiring human review and suspending systems that behave unexpectedly. Redundant, unsafe or unsupported tools also need a clear retirement path.
The governance model can begin with a small group and a simple intake process. What matters is that ownership does not become more ambiguous as adoption spreads. Otherwise, shadow use grows and accountability disappears when a problem crosses functional boundaries.
Business process
The search for AI opportunities often begins with tasks that are easy to automate. A more useful starting point is the process: where work stalls, which decisions consume scarce expertise, and where errors, exceptions and handoffs create cost or delay. This connects the use case to outcomes such as service, throughput, margin and growth.
It also tests whether the underlying work is stable enough to automate. If policies vary by person, ownership is unresolved or exceptions are poorly understood, AI may accelerate the inconsistency and make it harder to detect. The use case therefore has to be evaluated within the end-to-end flow of work, including its upstream inputs and downstream consequences. That broader view is how isolated automation develops into enterprise orchestration.
This is particularly important where AI sits on top of ERP. Consider an agent that monitors demand, inventory and production schedules, then recommends or places material purchase orders. It cannot compensate for a poorly controlled ERP process. If inventory transactions are late, production orders remain open, bills of material are unreliable or planners work outside the system, the agent sees a distorted version of operations. It may buy too early, miss genuine shortages or amplify excess inventory. The underlying process must be accurate, stable and used consistently before AI is given a role in material planning.
Data
Generative AI has expanded what organizations can do with documents, messages, images and other unstructured information. The basic requirements of data management still apply.
Leaders need to know which sources are authoritative, who owns them and whether they are reliable enough for the intended purpose. Access must reflect privacy, confidentiality and contractual restrictions. Outputs should be traceable when the decision warrants it, with feedback and results captured for evaluation and improvement.
The material-planning example illustrates how process and data readiness meet. Even if inventory balances and production demand are current, the recommendation will still be wrong if item purchase lead times, minimum order quantities, order multiples, safety-stock settings or supplier constraints are inaccurate. An agent may calculate precisely and still produce a poor purchasing plan because the master data describes conditions that no longer exist. Giving AI authority over a transaction therefore raises the standard for the data feeding it and for the controls used to maintain that data.
Perfect data across the enterprise is not a realistic prerequisite. Data should be fit for the decision being supported, with controls that reflect the cost of an error. An assistant that helps an employee search internal policies does not require the same level of validation, testing and oversight as an agent that releases a purchase order. Applying one standard to both would either expose the business to unnecessary risk or prevent useful low-risk applications from moving forward.
Solutions architecture
AI is entering through ERP, CRM and productivity platforms, specialist applications, stand-alone tools and internally developed assistants. Some capabilities belong inside ERP, close to the transactions, permissions and business rules they use. Others must work across ERP, CRM, maintenance, planning and external data, making a shared AI services or orchestration layer more appropriate.
These placement decisions belong within an enterprise architecture. Without that discipline, each use case can add another tool, data copy and direct connection. Many companies have already seen the result in older integration environments. Before governed integration platforms and APIs, point-to-point interfaces proliferated until dependencies became difficult to map. Routine ERP upgrades grew risky because changing one interface could break several others. An unmanaged AI estate will recreate that spaghetti architecture at greater speed, adding models, agents and knowledge sources to the web of dependencies.
Architecture must establish where AI capabilities should reside, how they connect to core systems and knowledge sources, and which services should be shared. It should also guide the choice among buying a capability, configuring one within an existing platform and developing something for a genuinely differentiating need.
The architecture also has to preserve room to change. Tight dependence on today’s model or vendor is a poor bet in a fast-moving market. A sound architecture allows components to evolve without forcing the business to rebuild its controls, integrations and operating model each time.
How to scale AI responsibly
The answer is to match ambition to maturity. The five foundations do not have to be complete before the organization begins. They do have to keep pace with the authority and reach given to AI.
Early opportunities may focus on individual productivity and bounded decision support. As governance and operating capabilities improve, AI can be embedded within functional workflows and then connected across systems and departments. At higher levels of maturity, people, enterprise applications, machines and AI agents can be coordinated around shared outcomes.
Risk changes at each stage. Drafting an answer, recommending a decision and initiating an action carry different consequences. Data standards, testing, monitoring and human supervision should become more rigorous as the cost of error and the degree of autonomy increase.
This makes the most impressive use case a poor default choice for the first pilot. An opportunity with clear ownership, usable data and measurable operating value will usually teach the organization more. It tests the technology while also revealing whether the business can govern the use case, redesign the work and support adoption.
The right level of ambition is the one the organization can execute reliably and build upon.
How to build the capability to scale AI
The practical starting point is an honest inventory of current activity: where AI is being used, what information is being shared and which experiments have produced measurable value. This baseline will usually reveal more adoption—and more unmanaged risk—than leadership expects.
The organization can then develop a portfolio tied to operating priorities. Each use case should be assessed for business value, process maturity, data readiness, risk, human supervision, change impact and implementation effort. Technical feasibility is one consideration among several, not the investment case itself.
A controlled pilot can test a bounded use case with a named business owner, defined measures, human oversight and clear criteria for scaling or stopping. Its lessons should feed back into governance, architecture, training and the rest of the portfolio. Reusing those lessons is what turns a sequence of experiments into an enterprise capability.
Individual tools will improve, converge and disappear. The durable advantage lies in the ability to decide where AI belongs, prepare the conditions for success and introduce it without destabilizing the business.
Most companies have already begun experimenting with AI, whether leadership initiated the activity or discovered it after the fact. The companies that scale it successfully will be those that strengthen the enterprise beneath it—one governed use case at a time.
Frequently Asked Questions
AI pilots often succeed because they’re conducted under controlled conditions with dedicated teams, curated data, and manual oversight. When organizations try to deploy AI across business units, they encounter governance, process, data, and organizational challenges that the pilot never addressed.
Organizations need clear governance, redesigned business processes, high-quality data, solution architecture that supports enterprise integration, and change management that prepares employees for new ways of working.
AI governance establishes clear ownership, decision rights, risk controls, and accountability. It helps organizations evaluate AI use cases, manage risk, and ensure AI systems operate safely and consistently as adoption expands across the business.
About the Author
Jonathan Gross, LL.B., MBA, is Pemeco’s Managing Director and head of its technology contracts practice. As a former litigator turned consultant and commercial lawyer, Jon’s clients benefit from his unique practice that includes technology law, technology strategy, enterprise software selection, and implementation management. By bridging the gap between legal and business, Jon’s clients benefit from his holistic approach to negotiating deals that drive commercial interests, manage risk, rebalance contractual equities, and promote successful implementations and long-term business partnerships. From high-growth start-ups to multi-national enterprises, Jon works with a cross-sector client base in private equity, manufacturing, distribution, property management, technology, professional services, and construction and engineering industries. Â
About Pemeco Consulting
Pemeco Consulting helps organizations succeed where most ERP projects fail. With a 100% success rate across 800+ projects, Pemeco guides clients through ERP strategy, selection, implementation, and transformation. Its globally recognized Milestone Deliverables methodology brings structure and clarity to complex programs. Independent and vendor-neutral, Pemeco serves private equity firms, manufacturers, and public sector clients. From strategy to execution, Pemeco delivers the insight, tools, and leadership needed to achieve ERP success—on time and in scope.
