An AI agent needs a clearly defined job
We build AI agents for tasks such as order processing: they read data from your business software, check it and prepare the next steps. Before writing any code, we agree with you which changes they can make, which need approval and who takes over when information is missing.
What is an AI agent?
An AI agent is a program that uses an AI model to carry out a task in several steps. It reads the information it needs, determines the sequence of operations within agreed limits and acts in your business systems. Its job goes beyond answering a question: it must reach a defined outcome, which may be a proposed change ready for approval.
The problems to solve
The same manual steps, every day
Every day, someone copies order details from your business software into a spreadsheet, looks for a confirmation email and checks the same screens again. As orders increase, so do these manual steps.
“AI agent” does not define the task
Processing orders and writing promotional content are different jobs, even when both are assigned to an “AI agent”. Without a defined task and an expected outcome, you have no basis for judging whether the project worked.
The demo works, but everyday use falls short
A demo using selected examples can make an agent look ready. Everyday work brings incomplete orders, duplicate customer records and documents in unexpected formats. Handling these cases has to be part of the task from the start.
What we build
A task defined from input to handover
We document what the agent receives, what it must deliver, which changes it can make and which need approval. We also define how to handle missing data and who takes care of exceptions and maintenance after launch. We agree on this scope with you before writing any code.
The data it can read and change
We check how the agent can access your business software, including Odoo, as well as your email and document storage. We build the agreed connections, specifying which data the agent can read and which it can change.
Limits and approvals
An agent can prepare a change without applying it. We agree which operations it can complete on its own and which require a person's approval, typically those involving prices, quantities or commitments to a customer.
A record of every step
We log what the agent read, which operations it carried out and what it changed. That record helps you trace an error, answer a customer's question and review what happened long afterwards.
What we do not do
We do not let agents make decisions about people
Decisions about a person — a job application, a credit limit, a formal report — stay with whoever is accountable for them. The agent prepares the information for review; it does not make the decision.
We do not automate a process nobody can describe
If everyone follows different rules and exceptions are the norm, an agent adds to the confusion. The process needs to be clarified first, and we will tell you that.
We do not hand over an agent without provision for maintenance
An agent connected to your business systems needs maintenance when a software module, the AI model or a procedure changes. Someone must be responsible for that work, and the project must account for it. Otherwise, the project is not finished.
The work your team still needs to do
An agent can reduce repetitive work, but your team still has to check its output and handle the cases it cannot complete. That takes time and skill. The project is worthwhile only if the overall benefit holds up once this work is counted too.
Exceptions that need human judgement
An order with a discount outside the agreed rules, a customer asking for an exception, a figure that does not add up: the agent passes these cases to the designated person. Someone must review them and decide what to do. These are also the most sensitive cases.
Checking the agent's work
At launch, every operation is reviewed; later, reviews move to spot checks. This needs someone who understands the process and can spot an error even when the result looks plausible. That time must be included when assessing the project.
Maintenance when something changes
A field changes in your business software, the AI model is updated, a procedure is revised: each time, the agent has to be retested on the cases that matter. This is recurring work, with someone responsible for it and time set aside.
What to measure
- How often is the task you want to assign to the agent repeated each week, and how long does it take each time?
- Out of a hundred cases, how many does the agent complete on its own, and how many does it pass to a person?
- How much time does reviewing the agent's work take in the first few weeks and once it is in routine use?
- How many errors reach the customer, and how long does it take you to notice?
- How many hours a month are spent on maintenance and retesting when systems or procedures change?
Facts and sources
Where a decision produces legal effects concerning a person, or similarly significantly affects them, and is based solely on automated processing, the General Data Protection Regulation gives that person the right not to be subject to it (Article 22).
Questions we often hear
- What is the difference between an AI agent and a chatbot?
- A chatbot is designed for conversation; the agent we build has a task to complete in your systems. They can work together: the chatbot gathers the request, and the agent carries out the agreed operations.
- Can it change orders, prices or customer records?
- Only with your authorisation, documented in the agreed scope. As a rule, when a change involves prices, quantities or commitments to a customer, the agent prepares it and a person must approve it before it is applied.
- What happens when it gets something wrong?
- The logs help trace the steps and identify where the error needs correcting. Before launch, we agree which operations require human review, particularly where an error could reach a customer.
- Will it integrate with your current business software?
- We check how your software allows data to be read and changed. Some systems have existing connections we can use; others need integration work. We establish this before proposing the project.
- Where is the data processed?
- We agree with you where processing will take place and record it in the contract. On our own website, for example, all model processing takes place on AWS Bedrock within the European Union.
- Can the project start with a single task?
- That is how we recommend starting. We measure the work before and after introducing the agent, including time spent on reviews and maintenance. The comparison helps you decide whether to continue and assess the ongoing cost.
- What does it cost to maintain?
- It depends on how often your business software, AI model and procedures change, and on the checks required. We estimate this work when defining the task and agree who will handle it after launch.
Tell us about a task you repeat every day
Tell us who does the task, how often it comes up and where it gets stuck. That helps us work out what an agent could take on, when it would need approval and what your team would still need to do.