The Big Story

AI Has Moved Beyond the Experiment

For many organisations, the first stage of AI adoption was small and informal. The next stage is more visible: putting AI into repeatable work, with clearer expectations about people, data and customer trust.

Why the conversation has changed

AI often entered organisations through low-friction tasks: drafting a first version of a document, summarising a meeting, finding information, creating ideas or helping a team respond to routine questions. These uses can feel private and low risk when one person is trying a tool for themselves. The conversation changes when the same tool becomes part of a shared process or a customer-facing service.

At that point, an organisation needs to know more than whether the tool is impressive. It needs to know what information is being used, who checks the output, what happens when the tool is wrong and how a customer or colleague can get help from a person. The important move is from a collection of individual experiments to a set of uses that the organisation can explain and manage.

That is why the most useful AI question is rarely “Which tool should we buy?” A better question is “What job are we asking the tool to do, and what do we need people to do around it?”

What the EU AI Act means in practice

The EU AI Act uses a risk-based approach. It does not treat every AI tool or every use case in the same way. The obligations depend on the system, the purpose for which it is used and the role an organisation plays in providing or deploying it.

As of August 2026, the European Commission says the majority of the Act’s rules and its enforcement framework have started to apply. The timeline is staged: prohibitions and AI literacy provisions began earlier, rules for general-purpose AI and governance followed in 2025, and some high-risk obligations have later transition dates. The Irish AI Office describes the main roll-out as progressive, with the full set of milestones extending to 2028 for some high-risk systems.

For organisations in Ireland, that means a sensible first step is to map the AI uses already taking place. A simple register might record the tool, the task, the information involved, the people affected, the owner of the process and the checks that take place before an output is used. It can also flag uses that deserve more attention, such as recruitment, employee management, access to important services or decisions that could materially affect a person.

The Act is not the only consideration. Where personal data is involved, GDPR responsibilities continue to matter. An organisation should be clear about what can be entered into a tool, whether the tool is approved, how information is retained and who can access the result. “The system produced it” is not a substitute for responsibility.

The customer experience question

Customers do not experience an AI strategy as a policy document. They experience it as a message, a recommendation, a decision, a queue or a conversation. They may be perfectly comfortable with AI when it makes a process quicker and makes things easier. They are less likely to accept it when they feel misled, or that the AI has made things more difficult to deal with. There is a certain element of the “uncanny valley” feeling that people may get from these situations, and they are quite offputting for the regular customer.

The European Commission’s transparency rules now include situations where people need to be told that they are interacting with AI, as well as requirements around identifying certain AI-generated or altered content. In practical terms, businesses should think about disclosure, the quality of the interaction and the route to human support together. A disclosure that is technically present but difficult to notice will not create much confidence. A chatbot that identifies itself but cannot hand over to someone who can solve the problem will still feel like a dead end.

This is where customer evidence becomes useful. Surveys can show whether people understood what happened. Mystery shopping can test the experience end to end. Focus groups can explore whether an automated interaction felt helpful, confusing or impersonal. Compliance audits can check whether the intended process is actually being followed.

What organisations should do next

• Make the use visible. List the AI tools and features people are already using, including uses that began informally.
• Give each use an owner. Someone should be accountable for the purpose, the information used, the quality checks and what happens when things go wrong.
• Train for the real task. People need practice with the situations they face, not only a general introduction to AI.
• Test the customer journey. Look at clarity, tone, accuracy, disclosure, escalation and the customer’s ability to correct an error.
• Review and improve. Keep a record of issues, near misses and useful feedback, then update the workflow and the training.

The strongest AI strategy is typically in the form of a set of practical choices about where AI belongs, where it does not belong and how people remain able to understand, check and challenge the results.

How Can We Help?

From AI Use to Practical Capability

Buying an AI tool is a technology decision. Making it useful, safe and consistent is a people and customer experience decision.

The implementation gap

An organisation can have access to a powerful tool and still lack the capability to use it well. One person may check every output carefully. Another may assume that a confident answer is a correct one. A third may avoid the tool altogether because the boundaries are unclear and they are afraid of using the tool incorrectly, with consequence. Managers may not know which use cases are taking place, while staff may not know where the line sits between sensible experimentation and unacceptable risk.

A policy is important, but a policy on its own does not change a working day. People need to see what the guidance means in the situations they actually face: a customer complaint, a sensitive document, a rushed deadline, a recruitment task, a difficult decision or an answer that does not look quite right. Capability is built when people can recognise the risk, choose an appropriate response and repeat that response under pressure.

Where Customer Perceptions fits

Customer Perceptions can help an organisation see what AI adoption feels like from the outside. The important question is not simply whether customers approve of AI as an idea. It is whether the interaction they have is clear, fair, accurate and easy to resolve.

Different research methods answer different questions. A survey can measure understanding and confidence across a broad group, regarding the AI use. Mystery shopping can test the real path a customer takes through a website, contact centre or service desk, especially with new AI pathways developing. A compliance audit can check whether the required AI tools and programmes are behaving correctly, are set up properly and regulations are being followed. Competitor analysis can show how other organisations are utilising AI effectively and with proper use. Focus groups can explore the language that makes automation feel reassuring or frustrating.

Where Optimum Results fits

Optimum Results can help organisations turn broad AI intentions into practical learning. Optimum Results has in the past, and scheduled for the future, AI adoption and policy courses. These courses are aimed at individuals who may be using AI in sales, marketing, or any general business practice.

We provide education on the latest AI adoption news and methods, show what has worked fr companies and job roles in the past and work through practical examples of how individuals would benefit most from including an AI stack into their workflow.

We inform course participants on the fair and correct use of AI applications, relating to EU AI Acts, Irish guidelines, and overall global standards of practice.

A collaborative example

Imagine a service team begins using AI to draft responses to routine customer enquiries. The tool is not allowed to make the final decision, but it can suggest a structure and pull together information from an approved knowledge base.

Optimum Results might help the team develop and practise the new workflow. Staff could work through a correct answer, an incomplete answer, a confident but inaccurate answer and a message that uses the wrong tone. They could agree what must be checked, what information must not be entered, when a case should be escalated and how the customer should be told what is happening.

Customer Perceptions might then test the experience. A sample of interactions could be reviewed for clarity and accuracy. Mystery shopping could test whether a customer can reach a person when the automated route cannot solve the issue. A short survey could ask whether the customer understood the answer and knew what to do next. A small focus group could explore whether the language felt helpful or evasive.

The findings can improve both the workflow and the learning. Perhaps the tool is fine for a first draft but unreliable when a customer is upset. Perhaps the team needs a clearer handover rule. Perhaps the wording is technically correct but sounds cold. The value comes from closing the loop: measure what is happening, understand why, train the people involved and measure again.

This is an illustrative example, not a description of a named client project.

What practical support looks like

• Start with one workflow. Choose a real task where the benefit and the risk can both be described clearly.
• Prepare the people around it. Explain the purpose, the boundaries, the checks and the handover points before the tool becomes routine.
• Measure the experience. Use staff feedback and customer evidence to understand what is working and where friction is appearing.
• Make improvements visible. Update the guidance, prompts, knowledge sources and training when the evidence shows a problem.
• Keep ownership clear. AI may assist with a task, but a named person remains responsible for the outcome and the customer journey.

AI adoption becomes more reliable when people can answer five plain questions: What is this for? What information can it use? What should I check? When should I stop? How can a person take over? Those questions connect training, governance and customer experience in a way that a tool list never will.

If your organisation is reviewing AI use, begin with one process and one honest conversation with the people who deliver it. That is usually enough to find the first useful improvement.

Put It Into Practice

Five Questions to Ask Before You Expand AI Use

Before a successful trial becomes a standard process, pause long enough to ask what the organisation is learning, what could go wrong and what people need in order to stay in control.

The questions below work for a small business testing a generative AI assistant, a public body reviewing an automated service, or a larger organisation adding AI to an existing workflow. They are not a replacement for legal advice or a formal risk assessment. They are a practical starting point for a conversation between the people who use the tool, the people responsible for data and technology, and the people who understand the customer experience.

Question one

What problem are we solving? Describe the problem in terms of the work, not the tool. Is the aim to reduce repetitive administration, help a team find information, make responses more consistent, improve accessibility or support a customer more quickly? What does “better” look like, and how will you know whether it has happened?

This question protects a team from adopting AI simply because it is available. If the only answer is “we need to use AI”, the organisation has not yet defined a useful test. Write down the current process, the pain point and the measure that matters. The measure might be time saved, fewer avoidable errors, clearer information or a better handover. It might also be a safety measure, such as keeping a person involved in every decision.

Question two

What information is being used? Identify what the system will see, what it may retain and who else could access the information. That includes customer details, employee information, commercially sensitive material, internal documents and anything that could identify a person when combined with other details.

Use the minimum information needed for the task. Remove names or other identifiers where they are not necessary. Check whether the tool is approved for the type of information involved, what the supplier says about retention and whether the organisation has a clear basis for processing personal data. A useful rule is simple: if you would not send the information to an unknown external recipient, do not put it into an unapproved AI tool.

Question three

Who checks the output? Name the person or role that is responsible for checking the result before it is used. Be specific about what the check covers: facts, calculations, tone, fairness, accessibility, confidentiality, policy and whether the answer actually addresses the customer’s need.

Human oversight should not mean clicking “approve” because the system has already made the work look finished. The person checking the result needs enough time, knowledge and authority to change it or reject it. If the work is high volume, the organisation should be honest about whether the proposed checking process is realistic. A control that exists only on paper will not protect the customer or the staff member using it.

Question four

What could go wrong? List the plausible failure modes before the workflow expands. The output might be inaccurate, out of date, biased, too confident, poorly phrased or based on the wrong information. A customer might not realise that they are dealing with an automated system. A member of staff might share information that should have stayed private. A recommendation might affect someone who has no clear way to challenge it.

The point is not to imagine every possible disaster. It is to identify the few failures that would matter most and put a control beside each one. That might be a trusted source, a second check, a limited pilot, a clear escalation route, a record of the decision or a rule that the system cannot be used for a particular task.

Question five

How can a person challenge it? Decide what happens when the AI is wrong, when the customer disagrees or when a staff member is uncomfortable proceeding. There should be a visible route to a person who can explain, correct or review the outcome.

This is both a governance question and a customer experience question. People are more likely to trust automation when they can see where human judgement remains available. The route does not need to be complicated, but it does need to work. Test it from the outside: can a customer find the handover, understand what will happen next and get an answer without repeating the entire story?

Turn the questions into a small review

Take one live use case and write the answers down. Ask the people who do the work to challenge the assumptions. Review a sample of outputs. Speak to a small number of customers or colleagues who experienced the process. Then decide what should change before the use becomes wider.

A 30-day review can be more useful than a long document that no one revisits. Look at the quality of the outputs, the time involved in checking, the issues raised by staff, the customer’s ability to get help and any near misses. Keep what is working, change what is not and set the next review date.

The goal is not to slow every experiment down. It is to make sure that expansion follows learning. When a business can explain the purpose, the information, the checks, the risks and the route to a person, it is in a much stronger position to use AI with confidence.

Thank You!

We hope you enjoyed reading our Newsletter content. If you would like to sign up for our quarterly Newsletter, feel free to click here to fill out the form.

If you have any suggestions on topics to cover, or any feedback in general, contact us through: [email protected] and/or [email protected].

References and further reading

These sources support the factual and regulatory points across the three articles. The regulatory position is changing in stages, so check the current guidance again before publication. The copy is intended as editorial support rather than legal, regulatory or data protection advice.

• European Commission AI Act overview Open source  Risk-based framework, obligations and implementation timeline.
• European Commission transparency and enforcement update Open source  Transparency, disclosure and content-marking context from August 2026.
• AI Office of Ireland business compliance guidance Open source  Irish overview of milestones and practical compliance resources.
• AI Office of Ireland for business Open source  Irish business guidance, support links and the national implementation context.
• EU AI Act Compliance Checker Open source  Official beta tool for an initial assessment of which rules may apply.
• Data Protection Commission Ireland Open source  Irish supervisory authority and data protection resources for organisations.
• Reuters Microsoft drafts code of conduct to keep its AI under human control Open source  Media reading on correction, shutdown, understandable AI and human control.