Skip to content
IvarAutomation
Back to Insights

AIagentsassistants

AI Agents as Assistants: What They Can Actually Do

A well-built AI agent can run entire workflows and make its own decisions. Here's how to match what it's allowed to do to the risks involved.

From chat window to a colleague who does things

A couple of years ago, "AI at work" meant a chat window for most people. You asked a question, got an answer, and pasted it somewhere yourself.

An AI agent goes a step further and does the work itself. It can read an email, look the customer up in the CRM, check an order status and write a reply. It takes several steps in a row, using the tools it has been given access to.

In demos it often looks like a digital employee. In everyday use, the result depends on how the agent is set up: what it may do on its own, what it should hand over to a person, and how you can see what it has done.

What makes an agent different from regular automation

A classic automation follows a fixed path: if A happens, do B. Every step is decided in advance.

An agent is instead given a goal, such as "gather what we need to answer this customer question", and a set of tools it may use. That might be searching documents, reading from a business system or creating a draft. The agent works in a loop where it takes a step, looks at the result and decides on the next step. So it decides for itself which tools to use, in what order, and when the task is done.

That makes the agent flexible. It can find a path you hadn't anticipated, but it can also pick the wrong one. That's why it needs boundaries.

Where agents already earn their keep

An agent can prepare work for a person, but it can also do the work itself and make its own decisions. How much it should decide on its own depends on what's at stake. A decision where a mistake is easy to put right is often one the agent can make by itself. A decision where a mistake could be costly or hard to undo should go through a person.

Pulling together information from several sources

A lot of office work is searching for information about a customer: invoices, the latest contract, what the salesperson promised in an email. An agent can go through several systems and gather it in one place, so whoever makes the decision doesn't have to go looking.

Sorting and prioritising incoming requests

An agent can read incoming requests, work out what they're about and route them to the right queue with a short summary. This is a decision the agent can make on its own, since a misrouted request gets spotted quickly and can be moved.

Answering directly or drafting, depending on the question

The agent can answer simpler front-desk questions directly, such as opening hours, where a delivery is or how to book an appointment. Emails that could have serious consequences for the business are a different matter, and the agent shouldn't send those on its own. That could be a complaint, a pricing discussion or a reply that commits the company to something. In those cases the agent writes a draft that someone at the company reads and sends.

Weaknesses we always account for

Agents have some well-known weaknesses. Whenever we build an agent, we account for them from the start.

Answers that sound convincing but are wrong

A language model phrases things just as confidently whether the answer is right or not. So we build in checks that make sure the answers come out right. The agent returns structured answers that can be checked automatically instead of free text. We also set up guardrails for what the agent may and may not do, meaning which permissions it has and which tools it's allowed to use.

Long chains can get fragile

Every step an agent takes involves a new judgement, and a small early mistake can get carried all the way through. That's why we evaluate the agent on real examples and test that the whole chain produces the right result before it goes into use, and again whenever something changes.

Cost and time vary

Because the agent chooses how many steps to take, the same kind of task can take a different amount of time and cost a different amount from one run to the next. That makes budgeting harder than for a regular rule-based flow.

Being able to review after the fact

An agent becomes hard to review if nobody can see what it read, which tools it used and why it chose what it did. That's why we set up observability with tracing and logs from the start. Every run can be followed step by step, so when something goes wrong you can see exactly where and why.

How to set the right boundaries

The most important design question has less to do with which model you choose and more to do with what the agent may do on its own and what requires a person.

We usually split it into three levels:

  1. Read: the agent can fetch and summarise information. Low risk, can usually run freely.
  2. Suggest: the agent drafts, proposes a category or prepares an action that a person approves.
  3. Act: the agent sends, books or changes something in a system. This needs clear rules and guardrails. How strict they should be depends on what a mistake could lead to. The agent can send a reply about opening hours on its own. Anything involving money, contracts or commitments in the company's name should go through an approval step.

A few principles make a big difference to how well it works:

  • Give the agent the least privilege it needs. Where it only needs to read, it should only have read access. That limits what can go wrong.
  • Let ordinary code handle whatever can be described as rules, such as amount thresholds, dates and required fields. Use the agent where interpretation is needed.
  • Set up tracing and logs for every run, so every tool call and every decision can be followed afterwards.
  • Measure before you expand. Let the agent run in suggest mode for a while and see how often its suggestions are approved without edits before giving it more responsibility.

An illustrative example

Imagine a small wholesale business where customers email in questions about their orders. They want to know where their delivery is, change a quantity or have an invoice explained.

Without an agent, someone in the office opens each email, finds the order in the business system, checks delivery status with the carrier and writes a reply.

With an agent, the flow might look like this:

  1. The email arrives and the agent reads it.
  2. The agent identifies the customer and the order, and fetches status from the business system and the carrier. It only has read access here.
  3. If the question is only about where the delivery is, the agent replies to the customer directly with the current status.
  4. If the question is about changing the order or about an invoice, the agent drafts a reply and a short internal note on what it checked. A person in the office reads it, adjusts it if needed and sends it.

The searching, which used to take the most time per request, is done in both cases. The simple questions get answered without anyone having to touch them. When a customer wants to change an order, the agent prepares the change and a person approves it.

Start narrow and let trust grow

The most common mistake is starting too broad. "An AI agent that runs our admin" is hard to build and hard to evaluate. A better first step is a single, well-defined task where it's easy to see whether the agent is doing a good job.

A few questions to ask before you begin:

  • A good first project is often a task that involves a lot of searching and little judgement. Which task in your business fits that description?
  • Which systems does the agent need to read from, and does it ever need to write?
  • How will we know whether the agent did a good or bad job?
  • What's the worst that could happen if the agent gets it wrong here?

The answers determine how much the agent can be allowed to do on its own and which guardrails are needed.

How we help

When we build agent solutions, we use rules where possible and AI where interpretation is needed. How much the agent gets to decide on its own is matched to what a mistake would cost. It often turns out that part of what was imagined as an "agent" is better solved with ordinary automation, and that the agent does the most good in a smaller, well-defined spot.

Wondering where an AI agent could take load off your team? Contact us and we'll go through a concrete task together and work out what it makes sense to hand to an agent.

Next step

Does any of this sound familiar?

Tell us how things look for you, and together we'll see what can be automated.

Free of charge · No obligation