AI agents for business: where do you start?

Start with one process. A company-wide AI strategy can wait. Pick work that comes back often, write down when it is good enough and agree on the moments where a person decides. Once that works, you expand.

This piece is for business owners and managers who are looking at AI agents for business and wondering where to begin. We are Studio Northstar, an agentic product studio in The Hague. We build our own products with agents and checks, and we use the same agentic platform to build software for other companies.

What is an AI agent, and what isn’t?

An AI agent is software that uses a language model to decide its own next step, uses tools to take it, and keeps going until a bounded goal is met or it hits a limit. A chatbot that answers one question is not an agent yet. Neither is a fixed automation in which a language model handles a single step.

Anthropic draws the same line in Building effective agents⁠↗. In a workflow, the model and its tools follow code paths written in advance. In an agent, the model directs its own process and decides which tools to use. Simon Willison puts it more briefly: an agent runs tools in a loop to achieve a goal⁠↗.

The difference matters. According to Anthropic, agents trade latency and cost for better results on hard tasks, and errors can compound because the model keeps working on its own. The advice in the same piece: find the simplest solution that works and only add complexity when you need it. Plenty of requests that arrive as “we want an agent” are better served by a workflow. How we build software with agents ourselves is in What is agentic engineering?

Which process do you start with?

Start with a process that comes back often, where you can say what a good result is, and where a mistake can be fixed before anyone is affected. The duller the work, the better the first candidate usually is.

• It happens often. Something that happens once a quarter pays back little and teaches you little.
• You can say when it is right. If two colleagues disagree on what a good result looks like, an agent can’t reach it either.
• The data is digital and reachable. If half of it lives in someone’s head or on paper, that is where you start.
• A mistake can be undone. Let the agent prepare, and let a person send, book or decide.

AI agent examples and use cases that make sense as a first project:

• Sorting incoming questions or requests, enriching them with data from your own systems and routing them to the right colleague.
• Checking documents for missing parts, such as a quote request or a case file, and listing what is still open.
• Preparing research: collecting sources, summarising them and flagging the questions a person has to answer.
• Reviewing code for security risks or mapping out older software. We do this ourselves with a council from ARCS.

What we would not pick as a first project: anything where the agent spends money, signs a contract or makes a promise to a customer. In our own system those actions always need a visible final step by a person.

How do you set the goal and conditions?

Before anything gets built, write down what you want to achieve in the long run, what the result has to meet and which questions are still open. That is how we start every project: we first discuss what you want to achieve, then capture the long-term goal, the conditions for delivery and the open questions in a brief.

We call that long-term goal the north star. An agent works step by step, and on large jobs the original goal can drift out of view while the model still reports that it is done. A written goal and clear conditions give you something to test every step against.

A useful brief for a first agent contains:

1. The goal in one sentence, in the language of your business. For example “every request reaches the right colleague within one working day”, rather than “an agent for the inbox”.
2. The conditions for delivery: when is a result good enough, and who decides that?
3. The limits: which systems and data the agent may use, and what it never does on its own.
4. The decision points: where a person approves the work, rejects it or asks for changes.
5. The open questions and assumptions, so you test them before they end up in the software.

Our page on research shows how we handle that first phase.

Build it yourself, buy it, or have it built?

Buy when your process is standard and an existing product already does it well. Build it yourself when the risk is small and your team wants to learn. Have it built when the process is specific to your company, has to connect to several systems or needs checks that an off-the-shelf product does not give you.

You can now build an AI agent with plenty of tools without being a programmer. For a first experiment, that is a good way to learn what an agent can and cannot do. It gets harder once the agent touches customer data, writes to your own systems or is used by several people. Then it needs what any software needs: tests, access rights, logging and someone who is responsible. That holds for small businesses as much as for large ones.

We do both. We develop and run our own products and build software for other companies, with the same agentic platform and the same checks. What that could mean for your organisation is on the homepage under For companies.

Which checks do you need?

You need at least five. The most important one is a person who decides before anything happens that is hard to undo.

• A person at agreed moments. Not at every step, or you no longer have an agent. But before anything that is hard to undo. In our setup the system presents the work at those moments: a person approves it, rejects it or asks for changes.
• A log. Afterwards you want to see which steps the agent took, with which data, and why it stopped. Our work rounds stop with a clear reason: approval is needed, the work is blocked or the budget has run out. Each round saves a checkpoint, so you can inspect and resume the work.
• Permissions outside the prompt. What an agent may do is set in configuration and access rights. A line in the prompt such as “never send anything without approval” is a request to the model. It stops nothing.
• A data choice per task. Decide per task whether the work may run at an outside AI provider or has to stay in-house, and record that choice. We process confidential data on Northstar One, our own machines and models on a private network. More on that in Running your own AI models.
• An independent reviewer. A model checking its own work does not always see its mistakes. Have the result reviewed by an independent reviewer: a different model or a separate context that did not produce the work.

How we set this up in our own work is in ARCS: agents with checks.

What do the GDPR and the AI Act mean for you?

The GDPR applies as soon as your agent processes personal data, including in a pilot. The AI Act sorts AI systems by risk. The heaviest duties, such as human oversight under Article 14, apply to high-risk systems, not to every agent.

The GDPR

According to the Dutch Data Protection Authority⁠↗ (AP), you have to establish in advance whether you can comply with the GDPR when you start using algorithmic systems that process personal data, “also if it concerns a pilot, test or trial project”. For such a project the AP points to a DPIA, an assessment of the privacy risks. The GDPR is technology-neutral: whether a person, a script or an agent processes the data makes no difference to the rules.

The AI Act

The AI Act⁠↗, Regulation (EU) 2024/1689, entered into force on 1 August 2024⁠↗. According to the AP⁠↗, it is being phased in and will be fully in force by 2 August 2027. Since 2 February 2025, a number of AI systems are prohibited and organisations that develop or use AI have to make sure their employees are AI-literate.

High-risk systems, such as AI for recruitment and selection or for law enforcement, face stricter rules: logging activity, managing data properly and making human oversight possible. Lower-risk systems mainly face transparency rules, such as marking generated content. An agent that sorts requests does not automatically fall under the heaviest rules. An agent that assesses job applicants probably does. It is up to your organisation to check which category a system falls into.

How do you measure whether it works?

You measure it by comparing the agent with how the work goes today, on the conditions you set in the brief. So take a baseline before the agent runs. Without one, all you know afterwards is how it felt.

• Quality: how often does the result meet the conditions without correction?
• Corrections: how often does a person step in, and why? The reasons tell you what to improve in the agent.
• Time: how long does a task take now, including the review?
• Cost per task: model usage, maintenance and the time of the people who check.
• Stop reasons: how often does the agent get stuck, and on what?

Feelings are a poor yardstick. In 2025, METR⁠↗ had sixteen experienced open-source developers work with and without AI tools. With the tools they were 19% slower, while they believed they were 20% faster. METR has since qualified that result. In an update from February 2026⁠↗, METR writes that developers are probably sped up more by today’s tools than in early 2025, but calls its own new data an unreliable signal of by how much. What people experience and what you measure can be far apart.

How does your team learn to work with it?

Your team learns by working with agents on its own tasks, with a shared grounding in what is and isn’t safe. That grounding is also a duty: the AI Act requires organisations that use AI to make sure their employees are AI-literate.

That is why we built the Studio Northstar Academy, a private learning environment with fourteen modules and 97 chapters, in Dutch and English. Your organisation works with one licence and chooses which modules the team follows. Progress is tracked per participant. Modules cover, among other things, working with AI safely and responsibly, the step from workflow to solution, and capturing ways of working with Agent Skills.

We also run hackathons for teams, such as the NextEnergy AI Hackathon 2026.

Want to know which process in your organisation is the best first candidate? Get in touch and tell us what you want to achieve. We will look together at whether an agent, a workflow or something simpler fits. What we do for companies is under For companies.