An organisation is ready for AI when it can clearly describe a decision or task, provide sufficiently reliable data, appoint an accountable person, and define human oversight and stop conditions. Without these elements, AI implementation in a company remains a technology experiment with an unclear business effect.
AI readiness is not about holding a large volume of data or already using an AI tool. It is the ability to manage a specific use case, from input data and business decision to result review, exception records, and a response when the system is wrong.
The first question is not which AI model to choose. A more useful question is: which process do we want to improve, and which decision within it currently consumes time, creates risk, or relies heavily on manual work?
A strong initial use case has several characteristics:
A manufacturing example could be preparing proposed production order priorities for a planner. The process needs a direct description: which orders are considered, which data affects priority, who confirms the schedule, when the schedule changes, and what follows from an incorrect proposal. A broad wish for smarter planning is not a sufficient specification for responsible use.
The same applies to accounting, procurement, customer support, and sales. If a team cannot describe the current workflow, its exceptions, and the person making the final decision, AI will not remove that ambiguity. It may only transfer it more quickly into a new system.
Before developing or procuring a solution, an ERP and process screening can be useful. This type of review can connect actual user work, ERP data, documents, integrations, and process control points.
Data availability is not the same as data fitness for use. AI readiness requires a review of the origin, completeness, accuracy, timeliness, access, and meaning of every important field.
For proposed production order priorities, relevant inputs may include due dates, material availability, capacity, order status, and priority rules. If some status information is managed outside the ERP, due dates are changed manually without a record, or item codes are inconsistent across systems, an AI output can appear convincing while resting on unreliable inputs.
Useful review questions include:
It is not necessary to wait for perfect data. It is necessary to understand the limits of data quality and align the use case with those limits. A proposal reviewed by a subject matter expert before execution may tolerate less completeness than an automatic order change, posting, or decision with a greater business consequence.
An AI system can prepare a summary, classify documents, propose a response, or recommend a priority. It cannot assume an organisation's business accountability. Each use case therefore needs an owner of the business decision, a process owner, and accountability for the technical operation of the solution.
In smaller projects, the same people may cover several roles, but the responsibilities still need to be explicit. The business decision owner defines the purpose, acceptable outcome, and limits of use. The process owner confirms operating rules, exceptions, and changes to the operational workflow. Technical accountability covers data access, integrations, performance monitoring, and incident handling.
It is also important to identify the user receiving the system output. The organisation should know whether that user is expected to confirm, modify, or reject a recommendation, and where that action is recorded. Without this step, oversight remains an informal assumption.
The question is not whether AI will make an error. Any operational system can produce an inaccurate, incomplete, or untimely result. The relevant question is: which error is acceptable, which requires human review, and which requires an immediate stop?
Error tolerance depends on the decision. A system proposing a draft internal summary can allow more room for editing. A system affecting an accounting posting, price, production sequence, contractual communication, or access to sensitive data requires tighter controls.
Before launch, document:
This record does not need to be a complex document. Its purpose is to align the business team, IT, and users before recommendations begin to affect daily work.
Human oversight is not merely the option for someone to look at output from time to time. It is a defined method for review, confirmation, rejection, and escalation.
For initial operation, it is often sensible to limit AI to preparing proposals. A user reviews the result before sending, posting, changing a priority, or performing another action. During this period, the team can observe where recommendations are rejected, what data users add, and whether new exception types appear.
A decision trail is useful: retained inputs, the version of rules or solution, the recommendation, the user's decision, and the reason for an exception when relevant. This record supports process correction, incident review, and explanations to users about how a particular output arose.
The NIST AI Risk Management Framework 1.0 provides a framework for considering AI risk management through governing, mapping, measuring, and managing risk. For organisations operating in the European Union, the European Commission's overview of the AI Act is also relevant. The application of obligations depends on the specific system, the organisation's role, and the context of use, so legal assessment should reflect the organisation's own circumstances.
Every AI use case needs a clear criterion for stopping operation or returning to a manual process. This does not dismiss the value of the solution. It protects the process when input data, system behaviour, or business conditions move outside agreed limits.
A criterion may include:
Alongside the criterion, define who may stop operation, how users are informed, which manual procedure continues the work, and who approves a restart. The operational team then does not need to improvise during a problem.
Rather than applying an artificial scoring model, management can work through six direct questions.
A negative answer does not automatically mean abandoning AI. It shows what needs attention first. The appropriate next step may be cleaning master data, standardising a process, completing an ERP integration, or clarifying responsibilities. These are useful outcomes of an AI readiness assessment because they reduce uncertainty before a larger investment.
When process, data, and responsibilities are sufficiently clear, an AI upgrade for business processes can be considered through a limited scope, a measurable workflow, and controlled introduction. For an initial assessment of a specific process, management can talk to the ORKA team with existing process documentation, a sample of input data, and a description of the decision it wants to improve.