AI agents for business are software that works through a task in several steps: read a request, look things up in your systems, draft a response, propose the next action. They are useful for a specific kind of work, and risky outside it. If you are looking for a European partner to build one, this is how we decide where an agent belongs, what it needs around it, and how to put it into production without handing it the keys.
The essentials
- An agent is a language model that plans and calls tools and APIs in steps. It is neither a chatbot nor a fixed workflow.
- Good candidates are multi-step processes with loose rules, where a person can review the output.
- Where the rules are fixed, or a mistake cannot be undone, plain automation is cheaper and safer.
- Most of the work is data access and integration with your ERP and CRM, not the model.
- Start with one process where the agent drafts and a person approves. Widen its autonomy only when the numbers hold.
What an agent is, and what it is not
A chatbot answers questions in a conversation. An automation follows a fixed recipe: when A happens, do B. An agent sits between the two. It is given a goal, chooses which tools to use, looks at what comes back, and decides the next step.
The tools are ordinary: a customer lookup, a case search, a stock check. The model decides the order. Your systems supply the facts and carry out the actions.
Our earlier piece on adding AI to existing software covers single steps: retrieval, classification, drafting, extraction. An agent chains several of those and picks its own route. More value, and more places to go wrong.
Where AI agents for business work well
The processes that suit agents share three traits. They take several steps. The rules leave room for judgement. And a wrong result gets caught before it does damage. Four that come up often:
- Case triage and drafting. The agent reads an incoming request, finds the customer and their history, judges urgency, and drafts a reply for a person to send.
- Order intake from email. Orders arrive as free text or PDFs. The agent matches item numbers, checks prices and stock, and creates a draft order in the ERP for approval.
- Internal knowledge with lookups. Staff ask about a customer or a delivery. The agent searches documents, queries the CRM and ERP, and cites its sources.
- IT and service desk. The agent sorts tickets, asks for missing details, suggests known fixes, and handles simple requests such as an access reset after sign-off.
Where plain automation is the better choice
A language model is not deterministic. The same input can produce two different outputs. That is acceptable for a draft. It is not acceptable for a tax calculation.
If the rules are fixed and known, write them as code. Plain integration is cheaper, faster and easier to test. Ledger postings, price calculation and threshold-based approvals belong here.
The same goes for actions that cannot be reversed: payments, deleting records, changing contract terms. An agent can prepare them. A person decides. Often the right design is a mix, with fixed code for the rules and an agent for the parts that need reading and judgement. Telling which is which is part of our AI development work.
Permissions and tool access
An agent should be able to do what the task needs and nothing more. Give it its own service identity with its own permissions, not an admin account and not a staff member’s login. Then you can audit it and switch it off.
- Read before write. Start with read-only access. Grant write access tool by tool, once you have seen how the agent behaves.
- Approval on actions. Anything that changes data, sends something to a customer or moves money goes through a person until you have evidence it does not need to.
- Narrow tools. A tool called “create draft order” is safer than one called “run query”.
- Treat input as data. An email can contain text that tries to give the agent new instructions. Nothing it reads should widen what it is allowed to do.
Integration, cost and evaluation
An agent is only as good as the tools it is given. Checking stock or creating an order needs stable APIs into your systems. If they are missing, the first project is often API development, not AI. Many companies keep their data in Business Central or Dynamics CRM. Both have APIs, but fields, permissions and business rules need mapping before an agent touches them.
Each task costs a number of model calls, and an agent with many steps can take seconds or minutes. Measure cost per task and latency from the first prototype.
Build a test set before you build the agent: real examples from the process, each with the answer an experienced person would give. Run the agent against it on every change of prompt, model or tools. Log every step: what the agent received, which tools it called, what came back and what it proposed. Without that log, you cannot explain a bad result.
On regulation, personal data falls under GDPR as usual: a purpose, a data processing agreement, and control over where data is processed. When personal data is involved, choose a setup where it is processed inside the EU. The EU AI Act adds transparency duties, such as telling people when they are dealing with an AI, and asks you to know which risk category your use falls into before go-live.
Frequently asked questions
A chatbot replies in a conversation. An agent works toward a goal over several steps and uses tools along the way: it looks things up, compares, and prepares actions in your systems.
For simple, low-risk tasks, yes, once your test set and logs show an acceptable error rate. For anything that changes data or moves money, keep a person approving, at least at first.
Rarely. The agent uses the systems you have, through their APIs. The work is making those APIs reliable.
Measure the process before the agent arrives: time per case, error rate, volume. Measure the same afterwards, together with the share of drafts staff approve without edits.
Where to start
Pick one multi-step process with a clear owner. Measure how it runs today. Build the agent so it only drafts and proposes, and have a person approve everything. Track approval rates and errors. Then give it more autonomy one tool at a time, as the numbers allow. We run this from process selection to production as part of our AI development services.
Have a process in mind that might suit an agent? Get in touch and we will look at it with you.
