Learn about AI · AI agents
What is an AI agent, and what can it do in a business?
Question or task: that is the whole difference
Ask a chatbot a question and you get an answer. Give an agent a task and you get a result. "What does article 4471 cost?" is a question. "Find out where order 48213 is and reply to the customer" is a task. It requires someone to find the order in the shop system, look the parcel up with the carrier, judge whether something is wrong, and write a reply. An agent can do all four in sequence.
What is new is not that computers can carry out steps. They always could, as long as the steps were identical every time. What is new is that the agent can cope with the steps varying: the customer writes "where is my parcel" with no order number and from a different e-mail address than the one they ordered with, the carrier reports "delivery attempted", and the same e-mail also asks about a return.
How an agent works
- Task
An employee, an e-mail or a scheduled time starts the task.
- Plan
The agent works out which steps are needed and in which order.
- Tools
It looks up, reads, writes or creates in the systems it has access to.
- Check
It checks its result against the rules you have given it.
- Hand-over
The result goes to a person for approval, or is carried out if you have allowed that.
"Tools" is the decisive word. An agent can only do what you have built a connection for: read the shop system, look a parcel up with the carrier, write a draft in the mail client, create a return label. Each tool carries its own permission, and those permissions decide what the agent could ever do. An agent without write access cannot change anything, however wrongly it reasons.
An example: customer service at a webshop
Take a webshop selling home accessories, with 12 employees and 250–400 orders a day. Customer service gets around 150 e-mails a day, and three out of four are about the same things: where is my parcel, I want to return, can you change the address, when is this back in stock. Two employees spend most of the day looking up orders in the shop system, checking tracking with the carrier and writing more or less the same e-mail over and over. Response time is a day on a normal day and three days after Black Friday.
An agent takes over the routine. When an e-mail arrives, it finds the order by name, e-mail address or order number, looks the parcel up with the carrier and writes a reply with the current status and the expected delivery time. If the customer asks to return, the agent checks that the item is within the return window and creates a return label. If the customer asks something that is not in the systems, or is clearly unhappy, the agent passes the e-mail to an employee with a draft and what it found. The employee decides. The agent does not.
- Start with read access
In the first month the agent only writes drafts. The employees read every draft before it is sent, and count how many they correct.
- Give write access to the boring part
Once the numbers hold up, the agent is allowed to send parcel-status replies and create return labels within the window itself. Refunds and complaints still go to a person.
- Log everything
Every step the agent takes is written to a log with a timestamp, what it read and what it decided. That log is what makes it possible to find a mistake and to trust the rest.
- Keep the rules somewhere you can read them
"A refund over 1,000 kr goes to a person" and "never say a parcel is lost before the carrier has confirmed it" live in plain text that the customer-service lead can change without asking a developer.
Response time on the three common questions drops from a day to a few minutes, on Saturday evening too, and that is not the main point. The main point is that the two employees spend their day on the customers who actually have a problem.
What an agent is good at, and what it is not
Good at
- Tasks that repeat and follow rules you can write down
- Lookups across several systems that today need four tabs open
- First drafts: replies, cases, orders, which a person approves
- Working when nobody is at work: nights, weekends, holidays
Not good at
- Judgement you cannot explain yourself: "we can usually tell whether a customer is serious"
- Tasks where a mistake is expensive and hard to undo, without an approval step
- Systems with no proper access: if the agent has to click around a screen, it becomes fragile
- Processes that are already unclear; an agent does not tidy up a messy process
What it takes on your side
- A task described precisely enough that a new employee could do it from the description. If you cannot write it down, the agent cannot carry it out.
- Access to the systems through their API or database, with a user that can only do what the agent needs to do.
- An approval step from day one. It can be removed later, when the numbers show the agent gets it right.
- A log, and someone responsible who reads it during the first month.
- A way to measure: how many tasks, how many corrected by a person, how long before and after.
Common misunderstandings
| What we hear | How it actually works |
|---|---|
| The agent makes decisions on its own | It only does what its tools and permissions allow. Deciding what it may do is yours, and it can start at "nothing without approval". |
| It's just a chatbot with a new name | A chatbot answers. An agent looks up, writes and creates in your systems. The connections are what make the difference. |
| It replaces an employee | It takes over the part of a job that is routine and lookups. The judgement and the responsibility stay where they are. |
| If we describe the task well enough, it can handle anything | The description helps, but the agent is still limited by the tools it has and by the quality of your data. |
What we would advise against
We would advise against giving an agent write access to the finance system in week one, however well it performs in a demo. And we would advise against starting with the most complicated process because the gain there is largest. Start with the process whose rules are clearest, and let the agent earn its trust on that one.
Frequently asked questions
What is the difference between an AI agent and ordinary automation?
Ordinary automation requires every step to be identical every time. An agent can handle variation: different formats, different wording, missing information. If your process really is identical every time, ordinary automation is cheaper and more predictable.
Can the agent do something wrong?
Yes, like a new employee. That is why it starts with read access and an approval step, and why everything is logged. The damage is limited by the permissions, and the mistake can be found in the log.
Which systems can an agent work in?
Any system with an API or a database we can get access to: shop systems such as Shopify, WooCommerce and Magento, carrier tracking, finance systems, CRM, e-mail, calendars, inventory. Systems that can only be operated through the screen are possible but fragile.
How long does it take to build an agent?
The first, well-bounded task can typically be in testing within a few weeks. What takes time is the connections to your systems and the rules for what it may do.
Does our data have to leave Europe?
Not necessarily. The language model can run with a European provider or on your own server. Your systems and data stay where they are; the agent only gets access to what it needs.
Read next
Want help getting from explanation to practice?
Have a no-obligation chat with us about where it makes sense to start. We reply within 24 hours.