The first process for AI should not be chosen because it is highly visible or popular, but because it can create clear business value with controlled risk. A strong candidate has a repeatable workflow, sufficiently reliable data, a person accountable for the decision, and an agreed success criterion. That is the purpose of AI screening: to establish where AI is justified before building a solution, and where it would merely add complexity.
A company starting with AI often sees many possible applications: handling enquiries, demand forecasting, document classification, customer-support assistance, or analysis of production data. All may be useful applications, but none is automatically the right first step.
The first project shapes how the organisation will later assess AI implementation: how it prepares data, who validates outputs, how errors are handled, and which indicators determine whether value has been created. It is therefore sensible to begin with a process important enough to make measurement worthwhile, but not so sensitive that one error creates an unacceptable business, financial, or operational consequence.
AI screening is not a catalogue of tools or a promise to automate every decision. It is a structured AI readiness assessment focused on a specific process, its inputs, outputs, exceptions, and responsibilities.
It is useful to assess each potential process in the same order. This prevents comparison from being driven by one department's impression or by the current availability of a technology.
Start with a concrete description of the work, not a technology label. Instead of asking, “Do we need a chatbot?”, ask:
For example, an incoming-document process may include retrieving a document, checking mandatory information, matching it with a received order, and preparing it for further processing. AI may support recognition, classification, or exception flagging, while accounting or business validation remains a human responsibility.
Business value is not the same as technical interest. Value may come from shorter processing time, less manual information searching, earlier exception detection, more consistent work, or better availability of information for a decision. The expected change and the party for whom it matters should be stated clearly.
AI does not automatically repair an unclear process or disconnected data sources. Before selecting a process, establish whether relevant data exists, where it is stored, and what it actually represents.
Check at least the following:
The absence of perfect data does not necessarily mean the process should be rejected. It may mean the first step needs to be narrower: for example, AI only suggests a category for a clearly defined document type rather than making a final decision in every case. In some situations, it is more useful to first improve master data, input rules, or integrations between the ERP and other systems.
ERP and process screening can help distinguish a data or process issue from a problem that is genuinely suitable for AI.
Every AI output can be inaccurate, incomplete, or unsuitable for an exceptional case. Assessment must therefore go beyond whether a model can generate a useful response. It must also cover what happens when that response is not good enough.
A low-error-risk use case may be a suggested internal summary that an employee reviews before use. Higher-risk outputs affect payments, accounting, contractual obligations, production planning, access to sensitive data, or customer communication without review.
For each candidate, define:
This approach aligns with the idea of governing AI risk through risk identification, assessment, and management described in the NIST AI Risk Management Framework 1.0 . For organisations operating in the EU, it is also relevant to follow European Commission information on the AI Act and assess the applicability of obligations to the specific use case.
AI can support work, but it does not remove the need for a business owner of the process. Before implementation, define the person or role that accepts, rejects, or further reviews an AI output and manages exceptions.
Human oversight is not merely a final click on a “confirm” button. It must be operationally workable. If an employee has neither the time, context, nor authority to review a suggestion, a formal control will not reduce actual risk. The workflow should therefore make clear:
A strong first process is often one in which AI prepares an action or ranks cases, while an experienced person retains the final business decision.
For an AI readiness assessment, bring together the process owner, process users, the person responsible for data, and, where relevant, IT, accounting, production, or management. Give each candidate a qualitative rating: low, medium, or high.
The process with the highest business value does not necessarily win if its data is poor and an error is costly. A better initial candidate is often a process with medium value, good data, a limited scope, and a clear review method. Such a project can establish a reliable way of working before broader use.
Without a baseline, it is impossible to know whether a change is an improvement. The success criterion should connect to the process, not to a general statement that AI is “useful”.
Depending on the process, this may be:
Alongside the main indicator, define a safeguard criterion. If faster processing is the goal, the safeguard may be that the number of errors detected in later control does not increase. If improved enquiry routing is the goal, the safeguard may be mandatory review for predefined sensitive categories.
Measurement should also include the comparison method: which period or case set serves as the baseline, who validates the result, and how exceptions are recorded. This avoids judging success on only a few impressive examples.
Some processes are not ready for a first AI project. This is not a screening failure; it is a useful conclusion. A process may first need clearer rules, cleaner data, better ERP records, defined authorities, or more stable integration.
Nor is every automation task an AI task. For strictly defined, stable steps, a conventional rule, form, or integration may be simpler to maintain and easier to verify. AI is more relevant when work involves unstructured content, variation in documents, or a high number of cases where recognising a pattern, priority, or exception is useful.
An AI solution should be considered as part of a process, its data, and its responsibilities, rather than as an isolated function. An AI upgrade for business processes makes sense only when it is known what is being upgraded, who governs the output, and how the effect is checked.
The practical next step is not buying a tool. It is a short workshop with the owners of two to five potential processes. For each process, record the business objective, work steps, data sources, exception types, cost of error, person accountable for oversight, and one primary success indicator.
If one candidate has a clear scope, available data, a controlled consequence of error, and a measurable objective after this assessment, it is ready for the next phase. If no such candidate exists, the outcome is equally valuable: the organisation now knows whether to improve the process, data, or integrations first. For a structured assessment of processes and potential AI roles, talk to the ORKA team .