Artificial intelligence in SMEs: case study with human review

This hypothetical case shows how Artificial Intelligence can be part of a SME process to sort information, prepare a draft, and maintain a human review before action. It is not a general implementation guide or describes AVSISTEC results.

The starting point should not be "we want to use AI", but "we want to reduce a repetitive task without losing control over it." To go through overall objective, scope, data, solution and validation, see the business guide to AI.

Scenario initial situation

Imagine a service company that receives queries by mail and form. A person reads each message, identifies the type of request, searches for information in several tools and prepares an initial response. The process may work, but it depends on reviewing messages one by one and repeating similar steps.

This is a example hypothetical. It does not describe a project or a result of AVSISTEC.

In such a case, artificial intelligence could be considered as a part of a workflow: interpreting the content of a query, proposing a category, recovering authorized data and generating a draft for human review. The decision to send a response, modify a record or undertake a commitment would remain defined by the company.

Observable problem: repeated tasks and invisible criteria

The problem is not necessarily that artificial intelligence is lacking. The process may have repeated manual steps, scattered information, or criteria that only one person knows.

Before assessing a solution, it is useful to observe the current work:

• What action the process starts.
• What data is received and from where.
• What decisions are made repeatedly.
• What exceptions require human intervention.
• What tools are involved?
• What output should occur: a notice, a response, a record, a task or a document.

A task is a candidate for review when there is a recognizable pattern, but that does not mean that it needs to be fully automated. If the context is ambiguous, the data is insufficient or the error would have relevant consequences, it is appropriate to keep a human revision before executing actions.

Objective analysis: separating results, scope and solution

A useful conversation about artificial intelligence distinguishes three levels:

For example, the objective may be that each query reaches the appropriate person with the necessary information. The scope may include entry from a form, classification of the application, creation of a task and internal notice. The technical solution is decided after understanding these conditions.

This order avoids converting a technology into the target. It also allows detecting that, in some processes, clear rules or well-defined integration may be sufficient without incorporating artificial intelligence.

Questions to define the objective

Before asking for a proposal, a company can answer these questions:

1. What specific task takes time or causes coordination errors?
2. Who is doing this today and who will review the result?
3. What information does the system need to act?
4. What actions can you propose and which ones need approval?
5. What happens when information is missing or the result is unclear?
6. How will it be verified that the flow is in line with expectations?

The answers need not describe a technical solution. They serve to turn a general idea into a problem that can be analysed.

Proposed scope: design a controllable flow

An initial scope should describe the entire path, not just the part that generates text or classifies data. In an automation with artificial intelligence, for example, they may intervene:

• The origin of the data, such as a form, mail or document.
• The rules that determine when the flow begins.
• Information available to the system.
• The task of artificial intelligence.
• The limits of trust or situations that must be scaled up to a person.
• The next action: create a task, update a record, prepare a draft or send a notification.
• The record of what happened so I could review it.

It is also appropriate to prioritize. The first version can concentrate on a single high-value task and leave for later additional integrations, new categories or more sensitive automatic actions.

A hypothetical example of scope

An advisory service wants to sort the requests received by mail. The first version,hipotetica, could be limited to:

• Detect if the message corresponds to a regular query or an incident.
• Propose an internal label.
• Create a task for the responsible person.
• Prepare a summary of the original message.
• Keep the human review before responding to the client.

Automatic sending of replies, modification of sensitive data and any decision requiring professional judgement would be outside this first scope. Delimiting these exclusions helps to avoid ambiguous expectations.

Solution and verification: integrate, review and adjust

A solution with artificial intelligence must be able to be tested under real conditions of use. To do this, the company can define acceptance criteria before starting the flow.

For example:

Exempla hypothetical: when a complete query comes from the form, the system creates a task with the summary and proposed category. If data is missing or cannot be clearly sorted, the task is marked for revision without performing additional actions.

This concrete criterion what should happen, what is expected when everything goes well and what happens when an exception is made. Moreover, it turns the revision into part of the design, not an improvised correction.

The verification should include representative samples: routine requests, incomplete messages, ambiguous texts and situations that do not fit the categories envisaged.The aim is not to assume that the system will be correct in all cases, but to define when it can attend and when it should ask for intervention.

When processing customer data, internal documents or personal information, the company should review what information is incorporated into the flow, who accesses it, what tools are involved and what requirements apply to its activity.

Transferable learning

Artificial intelligence makes more sense when it is incorporated into a process that the company can already explain. These ideas apply to different contexts:

• Start with a specific task, not a fashion tool.
• It defines who maintains control and what decisions are not delegated.
• Describe the exceptions before automating the main action.
• It separates the essential from what can be expected in a later phase.
• Check the flow with observable criteria.
• Document data, integrations and responsibilities.

If you have identified a repetitive task that needs better coordination, classification or preparation of information, you can consider it as a automations project. Explain the current process, the result you are looking for, and the tools you already use to request a budget or assess the scope.