From business need to use case: when AI makes sense

October 5, 2026

In many companies, AI adoption starts with an apparently logical question: “Where can we apply it?”. If the problem is not clearly defined, the result may be a use case that looks attractive on paper but is difficult to integrate into day-to-day operations, justify in terms of the value it delivers or sustain over time.

The first step is to understand how the process works, where friction occurs and what outcome the business expects to achieve. This assessment makes it possible to identify and qualify business needs that may be addressed through data-driven initiatives, and then determine whether the right response requires AI, advanced analytics, visualisation, data engineering, automation, systems integration or process redesign.

Before choosing a technology, it is important to understand which process needs improving and what outcome is expected.

Start with the process, not the solution

A request such as “we want to apply AI to this process” may reflect very different situations: locating information scattered across hundreds of documents, automating manual tasks governed by stable rules or supporting decisions shaped by incomplete information, changing circumstances and coordination across multiple systems.

Sometimes, the difficulty lies in the design of the process itself: redundant approvals, unclear responsibilities or data being entered more than once.

A single solution is unlikely to work equally well across such different scenarios. Before choosing one, it is important to clarify:

What outcome does the process produce, and for whom?
Where do delays, errors or avoidable costs occur?
What decisions are made, and what information do they require?
Which parts follow stable rules, and which require interpretation of context?
What data is available, what is its quality and under what conditions can it be used?
What level of human involvement needs to be retained?
How will improvement be measured?

When the need affects the organisation more broadly, a wider assessment is advisable. An evaluation of data and AI maturity can help assess the organisation's starting point across business, organisational, data and technology dimensions.

The same need may call for automation, integration, analytics, AI or process redesign.

Understand the process from five perspectives

Viewed simply as a sequence of tasks, a process may appear straightforward. However, many of its challenges lie in less visible elements: the systems that support it, decision rules, coordination between teams or how knowledge is retained.

To build a sufficiently broad view, it is useful to analyse the process from five complementary perspectives:

1. The value-generating activity

The analysis should define what the process transforms, its inputs and outputs, and the outcome expected by the user, customer or business area. Without that reference point, optimising an individual task may not improve the overall result.

2. Operational support

This perspective covers the data, applications, resources and infrastructure that support the process. This is where disconnected systems, duplicated information, unstructured documents or manual tasks used to compensate for a lack of integration between applications often emerge.

3. Rules and decisions

It is necessary to identify who makes decisions, which criteria are used, how cases are prioritised and what happens when an exception occurs. This helps distinguish between decisions that can be handled through rules and those that require interpretation of less structured information, consideration of changing circumstances or human judgement.

4. People and teams

Roles, capabilities and coordination between teams also shape the outcome. A technically viable initiative may lose value if ownership is unclear, if it adds extra steps to day-to-day operations or if the people expected to use it are excluded from its design.

5. Governance and evolution

It is important to define how the process will be measured and improved and, where its impact or regulation requires it, how decisions will be documented and audited. If data or AI models are involved, their quality, security, traceability, intended use and compliance must also be considered.

■ Incorporating Data Governance and AI Governance from the outset makes it possible to anticipate requirements and reduce the need for changes once the solution has already been built.

Analysing the process from multiple perspectives helps uncover barriers that are not visible when looking only at the sequence of tasks.

From friction to a well-defined opportunity

Identifying friction shows where a problem occurs, but does not explain its causes or the right response. “The team takes too long to respond”, “errors occur when processing requests” or “it is difficult to find the right documentation” are symptoms. To turn them into opportunities for improvement, it is necessary to understand what is causing them.

To assess an opportunity, it is useful to capture:

The process and users affected
The problem observed
Its likely causes and the evidence available
The business outcome to be improved
The data and systems involved
Operational, regulatory and security constraints
The indicators that will be used to assess the result

■ Defining each opportunity using these elements makes comparison easier and avoids building up a list of ideas that are disconnected from day-to-day operations. The same framework can be used to compile an initial portfolio of initiatives of different kinds, such as data engineering, integration, visualisation, advanced analytics, machine learning, generative AI or an AI agent capable of using tools.

A well-defined opportunity makes it possible to compare different initiatives using common criteria.

Three processes, three different responses

The following hypothetical scenarios show how similar points of friction can arise from different causes and why diagnosis should come before technology selection.

Scenario 1: improvement does not require an AI model

A company's procurement team takes several days to approve some requests. The initial assumption is that the delay is caused by the volume of documentation, leading to a proposal to use AI to review it.

The analysis reveals a different cause: the same data is entered into two applications, several approvals are requested sequentially even though they could be carried out in parallel, and most decisions follow stable rules based on amount, cost centre and supplier type.

The more proportionate response would be to simplify the workflow, integrate the systems and automate the rules. Introducing an AI model could add complexity without addressing the underlying cause of the delay.

—If the delay is caused by process design, adding AI may increase complexity without addressing the root cause.

Scenario 2: generative AI makes knowledge easier to access

A technical support team consults manuals, procedures and incident records distributed across different repositories. Its professionals spend too much time locating the correct version of each document and gathering the information they need.

Accessing the documents is only part of the problem: the relevant content also needs to be selected, connected and presented in a context that is useful to the specialist. A generative AI application connected to corporate sources could help locate, summarise and contextualise that information. It should show the sources used, respect access permissions, work with up-to-date documentation and require specialist validation for sensitive responses.

The objective would be to reduce search time without replacing technical judgement. The same principle should apply to generative and agentic AI solutions: identify the use case first, then assess how best to integrate them into day-to-day operations.

—Generative AI adds the most value when the challenge is to locate, connect and contextualise dispersed knowledge.

Scenario 3: an agent can coordinate a multi-step task

A company needs to review supplier onboarding requests. The process requires checking documentation, consulting multiple systems, requesting missing information and routing exceptions to the appropriate team.

When the workflow combines rules, variable documentation, queries across different tools and tracking of each case, it may be reasonable to assess an AI agent with a clearly defined scope of action. The agent could coordinate intermediate tasks and prepare the information needed for decision-making without replacing the process controls.

It could also gather information, flag inconsistencies and suggest the next step. Higher-impact decisions, such as approving a supplier or changing bank details, would retain the appropriate authorisation controls and human oversight. Its usefulness would depend on carrying out its tasks with the required reliability, traceability and control, as well as on the limits placed on its permissions, data sources and autonomy.

—An agent can coordinate multi-step tasks, provided it operates within clear boundaries for control and oversight.

■ The same logic applies to the everyday use of tools such as ChatGPT, Claude or Microsoft Copilot. Some tasks are best handled through an iterative conversation, while others are better suited to an agent that can carry out several steps. In some cases, automating the task takes more effort than it saves, or completing it manually remains the better option. The aim is not to use more AI, but to choose the form of support that adds the most value.

How to prioritise use cases

The scenarios above show that not all opportunities are the same or require the same type of response. Once they have been identified, prioritisation helps focus resources on initiatives that combine expected value, feasibility and practical implementation readiness.

Our Data & AI Consulting helps to prioritise these opportunities through a methodology tailored to each company's criteria and context. The assessment applies minimum criteria across technical feasibility, business impact, ethics, regulation and governance, and makes it possible to compare and prioritise initiatives of different kinds on a common basis.

This work is carried out together with the customer's business subject-matter experts and using a technology-agnostic approach. As a result, the selection process does not start with a predetermined tool, but with the specific needs of each process and the initiative's ability to deliver measurable impact.

To make initiatives easier to compare, these criteria can be grouped into four dimensions:

Business impact: what outcome is expected to improve, for whom, and how will that improvement be measured?
Technical and data feasibility: is there data of sufficient quality, availability and traceability? Can the technology provide the required reliability and integrate with the systems involved?
Risk and governance: what would the consequences of an error be, and which security, ethical, regulatory, oversight and traceability requirements must be met?
Operational readiness: who will own the use case? How will it be incorporated into day-to-day work, maintained and evolved over time?

■ With this approach, the most technically advanced initiative does not necessarily need to come first. A well-defined, measurable use case supported by suitable data may deliver more value than a more ambitious one for which the company is not yet ready.

Prioritisation is not about choosing the most advanced option, but the one that combines impact, feasibility and real implementation readiness.

From use case to roadmap

If prioritisation identifies where to start, the roadmap turns that decision into an initiative that can be delivered. For each priority use case, Use Case Discovery makes it possible to define a technical and functional blueprint, the requirements for its development, the reference architecture, including its technology components and services, the owners, metrics, controls and implementation roadmap. Depending on the project, a minimum viable product can be used to validate assumptions and refine the solution before scaling it.

The roadmap should also capture shared dependencies: several use cases may rely on the same data source, integration capabilities or governance framework. Identifying these dependencies reduces the risk of creating isolated solutions and makes it possible to plan reusable capabilities.

Implementation also requires adapting day-to-day operations and preparing the people who will use the solution: simply having a tool available does not guarantee a change in ways of working. Involving the relevant teams from the discovery stage brings the design closer to operational reality and helps prepare for adoption ahead of deployment. Technical support and training help maintain alignment with the objectives and support cultural change.

A roadmap turns prioritisation into execution by incorporating dependencies, controls and operational adoption.

Conclusion

Before introducing AI into a process, it is necessary to determine what that process needs in order to work better. Sometimes reorganising tasks or connecting applications will be enough; in other cases, analytics may identify patterns or generate forecasts, generative AI may make dispersed knowledge easier to access, or an agent may coordinate actions across multiple tools.

Understanding the process, defining the expected outcome and comparing alternatives using common criteria helps direct investment towards initiatives that are both relevant and achievable, while also making it easier to review results throughout implementation.

AI adds value when it addresses a specific need and can be integrated, measured and sustained in day-to-day operations.

Búsquedas recomendadas