What are AI agents? Can a Maltese firm use one?

What an AI agent is

A chat tool answers what you ask, from what you give it. An agent is given a goal, instructions and tools, and works in a loop: it decides the next step, uses a tool, looks at the result and decides again. Microsoft describes agents in its Copilot Studio in similar terms: they follow your instructions, draw on the knowledge you connect and use tools to take action. It also offers rule-based agents for questions that need the same answer every time.

Some agents start on an event, such as an email arriving, and Microsoft says some can have their own account to work proactively. Those need the most care: nobody watches each step.

An agent is not a fixed automation. Microsoft describes its agent flows as deterministic: the same input always gives the same output. An agent that chooses its own steps copes with untidy input but is less predictable, so good set-ups keep fixed rules for the predictable steps.

What an agent can safely do in a company today

Most of the safe work today is answering, or preparing something for a person, rather than acting alone.

What it does	An example	Who checks it
Answers from documents	Which products must be kept chilled, from the product sheets, with the sheet named	Someone checks a sample
Sorts and routes	Labels each message in a shared inbox and passes it on with a short summary	The person who receives it
Checks	Compares a supplier invoice with its purchase order and flags any difference	Whoever looks after that supplier
Prepares	Reads a delivery note and fills in the details for someone to post	The person who posts it
Drafts	A reply to a routine customer question, or a reorder when stock runs low	The person who sends or places it

Not safe yet, for most companies: an agent that pays, changes prices, edits customer or supplier records or posts to the accounts with nobody approving each step.

What an agent needs underneath

Connected, agreed data

An agent acts on what it reads. If a customer sits in two systems under two names, or sales and finance mean different things by "overdue", it acts on whichever version it reaches, with confidence. So the data comes first: in one place, entered on time, each key word defined once. Where to start with AI lists the questions to ask, and what Copilot Studio can and cannot see shows the problem inside one agent builder.

Permissions

Give it the least access the job needs: read-only access where reading is enough, drafts rather than sends, one mailbox rather than all. Ask whose access it uses, the person asking or one shared account. The OWASP Gen AI Security Project lists this as a risk of its own, excessive agency: too many functions, permissions or too much autonomy.

A person who approves

Every agent needs a named owner, and every step that leaves the company or changes a record needs a person's yes first; Microsoft's agent flows, for example, include approval requests. Keep a record of every run, what set it off, what it read and what it did.

Risks, and where agents stop

Confident mistakes. An agent can pick the wrong record and carry on, faster than a person would. Test it on past cases where you know the answer.
Hidden instructions. An email, web page or file can carry text written to steer an AI. OWASP calls this indirect prompt injection and says it is unclear whether any method prevents it completely. Microsoft says custom agents built in Copilot Studio include protection against it; narrow access and an approval step limit the harm if something gets through.
Running costs. Copilot Studio is paid for in Copilot Credits. Each action an agent flow takes uses them, a flow an agent sets off by itself is charged even where staff hold Microsoft 365 Copilot licences, and when credits run out, new runs stop.

Where agents stop: they do not know what was never recorded, cannot tell which of two figures the business agreed on, and should not make a call someone is accountable for, such as who is given credit terms or what a price should be.

Can a Maltese business use one?

Yes, and being in Malta changes little. As anywhere in the EU, GDPR applies to personal data an agent reads, so check where the tool processes it and the supplier's terms first. Whoever looks after your Microsoft 365 or other systems will need to set its access. If customers write in Maltese, or mix Maltese and English, put those messages in your tests.

How to start small

Pick one routine job that repeats every week, with a right answer you can check.
List what it must read, and check that it is entered on time and means the same thing to everyone.
Look at what you may already have. If you run Microsoft 365, your administrator decides whether staff can use Agent Builder, which makes agents that answer from SharePoint and other connected knowledge.
Set the least access and the approval step, and name the owner.
Test it on past cases, run it beside the person who does the job today for a few weeks, and widen it only on the time saved and errors avoided.

Where 110 comes in

We build the AI tech your business runs on. We connect your whole business and put AI to work making sure every entry goes in properly, so the right numbers are in front of you every morning and your team spends its time on the work, not the reports. Each system loads every night into one warehouse in your own cloud account. We do not switch on AI answers before the figures underneath are connected and agreed.

Reminders, replies and proposals are drafted and queued for a person to approve. AI does not make your decisions. It checks, flags, answers and drafts; your people decide and send. Each system is read under access your administrator controls, and the systems you run are never written to.

More on AI on your own data and business automation.

Common questions

What is the difference between Copilot Studio and Agent Builder?

Agent Builder, in Microsoft 365 Copilot, makes quick agents that answer from knowledge such as SharePoint content, without actions that reach other services. For those, Microsoft's route is to copy the agent into Copilot Studio, its separate studio for agents and workflows, which connects to business systems so an agent can look things up in them and act on them.

Should we buy an off-the-shelf agent or have one built?

An off-the-shelf agent fits a common job inside one tool, such as answering from your documents. A built one fits when the job crosses several systems, follows your own rules or must prepare work for a person to approve. Either way, the data and the access matter more than how it was made.

How much should an AI agent cost?

There is no single price; what it should cost depends on the time and errors it saves. Expect four parts: the platform, through licences or charges by use; any clean-up of the data it reads; the work to connect and test it; and the time of the person who approves its work and looks after it. Price one small job first.

What are AI agents, and can a Maltese business use them?