Agents only work on facts

Many companies have now rolled out AI. A few colleagues are enthusiastic and have it write emails and summarise documents. Handy, but that is often where it stops. The simple tasks are still done by hand: retyping an order, copying data from one system to another. That is because the AI has no access to your systems and cannot act on its own.

In our previous blog we wrote about the three-legged chair: data, solutions that connect to that data and people with skills. In this blog we look at what becomes possible once those legs are in place.

The first step: from chatbot to assistant with knowledge

The first step is connecting your systems. As soon as AI is connected to your ERP system, your email and your calendar, something changes. Data that was always there suddenly becomes visible. Questions that used to cost an afternoon of searching through your systems, you now answer in a few minutes. Think of questions like: which customers were responsible for which revenue last month? What are rising fuel costs doing to your margins? Which companies have you been in touch with over the past month?

You can start working with it straight away. If you set it up well, it saves your team time.

The second step: from assistant with knowledge to assistant with knowledge and a mandate

After that, the AI can get to work by itself. An agent does not just give an answer, it completes a task from start to finish. It looks at the data, makes a decision, carries out the action and only comes to a person when something does not add up.

That development is moving incredibly fast. According to KPMG, 54% of large US organisations were already actively deploying agents in the first quarter of 2026. A year earlier it was 11%. The models are also improving exponentially. Research organisation METR measures how long the tasks are that AI agents can complete on their own, with a 50% success rate. That length has been doubling roughly every seven months since 2019. Where you could once put a model to work for a few seconds or minutes at most, it now works for hours at a stretch, without anyone watching over it.

The work you are happy to get rid of

Recurring tasks with a clear start and finish are an excellent fit for AI. A few examples. An agent can read every order that arrives by email, check stock and price and stage the order in the ERP. Only the exceptions land on your desk. It can also check whether a new supplier is legitimate, for example whether the bank account number matches the country.

It does not stop there. An agent can check whether incoming invoices match what was delivered, describe, categorise and translate new items for the web shop, prepare work orders and paperwork from an order, and check supplier lists with hundreds of products for price and item data. It handles standard customer questions, such as delivery status, stock and price, largely by itself. The rest goes to a person. That person then plays more of a reviewing role and has time for work that needs human input and genuinely moves the company forward.

You cannot let agents guess

An agent that has to guess makes mistakes, just like a new colleague who is told nothing. What it needs is not complicated, but all of it has to be there: data, reasoning power and clear instructions.

Data is what the agent sees: your orders, stock, prices, customer details and emails, straight from your systems. If that data is wrong or out of date, the agent will carry on neatly working on wrong facts.

Reasoning power is the AI model itself. It reads an email from a customer, understands what is being asked, compares that with your data and decides what the next step is. This is the part that has improved so fast over the past years.

Instructions are your rules: what is true, when may the agent act on its own and when does it hand something to a person? That is its mandate. A standard order it prepares itself. If the price deviates from the agreement, it stops and asks you.

This is how a task runs from start to finish. The agent reads what comes in, pulls in the data it needs, applies your rules and carries out the action. What does not fit, it hands to a person. That last part matters: make sure a person is consulted for exceptions.

The question is what you do with the time

Once the foundation is in place, the possibilities are practically endless. The question is then not whether agents will take work off your team. The question is what your team will do with the time that frees up.