Where humans stay in the loop

What shipped

The useful shift is not that every team can now “automate everything”. It is that non-technical staff can build small agents that help with repeatable work, provided the workflow keeps a person in control.

For an Irish SME, that might mean an agent that reads a supplier email, drafts a reply, prepares a summary for a manager, or suggests which folder a document belongs in. The agent is not the decision maker. It is decision support, reviewed by a person before operational use.

That distinction matters. Many teams hear “AI agents” and assume they need developers, data scientists or a large platform project. In practice, the first useful agents are often built around plain business rules, sample documents, and the team’s own way of working.

The human review step should be named and designed, not left as a vague instruction. I usually call it Confirm before action. The agent can prepare the draft, extract the fields, suggest the next step, or flag the exception. A person then checks the output, edits where needed, and approves the action before it reaches a customer, supplier, payroll run, booking system, or internal record.

That is the core of good agent-building training: teach people how to create assistants that are helpful, limited and reviewable.

Who should care

This is for owners, operations managers, finance leads, HR managers and service managers who can see the admin burden but do not want a risky AI rollout.

It is also for teams that already use Microsoft 365 or Google Workspace and suspect there is more value in the tools they have, but are unsure where to begin. Tools such as Copilot Studio, Google AI Studio and other agent builders are becoming easier to use, but ease of use does not remove the need for judgement.

The people closest to the work usually know the exceptions, the awkward hand-offs and the phrases that signal risk. They are often the best people to shape the first version of an agent. What they need is a safe method: define the task, narrow the scope, give the agent reliable material, test it with real examples, and put a named reviewer in the path.

Richard recently completed a cross-departmental Copilot Studio workshop for a large organisation. The lesson transfers well to SMEs: the hard part is rarely pressing the buttons. The hard part is choosing a sensible workflow, setting boundaries, and agreeing what a human must approve.

If your team is asking “Can we build this ourselves?”, the answer is often yes, for the right kind of task. If the question is “Should we let it act on its own?”, the answer is usually no, at least at the start.

A worked example

Take a small Irish distributor that receives product enquiries by email. Staff spend time reading the enquiry, checking whether the customer is trade or retail, finding the right product sheet, drafting a reply, and deciding whether the query should go to sales, accounts or technical support.

A first agent should not be given permission to send replies on its own. A safer version would do four things.

First, it reads the enquiry and produces a short summary: who is asking, what they want, whether they mention an order, and whether there is any urgency.

Second, it suggests a route: sales, accounts, support, or manager review. This is a suggestion, not an instruction to move the case automatically.

Third, it drafts a reply using approved wording and the company’s own product notes. If it cannot find a reliable answer, it says so rather than guessing.

Fourth, it presents everything to a staff member at the Confirm before action step. The reviewer checks the summary, corrects the route if needed, edits the draft, and then sends or assigns it.

This is not glamorous, but it is useful. It can reduce blank-page drafting, make routing more consistent, and help newer staff handle common enquiries with more confidence. It also gives managers a clear place to inspect quality before any operational use.

The training value is just as important as the agent itself. Staff learn how to describe a task, spot where AI is likely to help, recognise where it may fail, and decide what should remain human-owned.

That is why this belongs in enablement and training rather than a large technical project. The goal is not to create one clever demo. The goal is to give the team a repeatable way to build small, safe agents for everyday work.

Where I'd start

Start with one workflow where the team already checks work before acting. Good candidates include enquiry triage, internal policy Q&A, supplier email summaries, draft responses, meeting follow-ups, training support, or document classification.

Avoid starting with anything that moves money, changes employment records, gives legal advice, approves credit, or sends customer messages without review. Those may become possible later, but they are poor first training examples.

For the first session, bring three real examples of the work, with any personal or sensitive details removed. Bring the current wording, policy, spreadsheet, folder structure or checklist the team already relies on. Then ask four questions:

1. What should the agent draft or suggest?
2. What sources is it allowed to use?
3. What must it never decide by itself?
4. Who performs Confirm before action?

From there, the team can build a small agent, test it, compare outputs, adjust the instructions, and decide whether it is useful enough to keep. The result should be a working pattern the team understands, not a black box.

If you want your non-technical team to learn this in a practical setting, Artellis offers agent-building training for SMEs. You can also read more general AI workflow thinking in the Artellis insights library.

The takeaway is simple: agents are safest and most useful when they prepare the work, not when they take over the work. Train your team to build around that line.