In brief
People stay in control of AI agents when it is clear before the start which defined task an agent takes on, which steps it only prepares and which it may carry out. Agents work within the configured access and permissions. The people responsible review the results and approve them before changes are implemented in connected systems.
People stay in control of AI agents when three things are clear before the start: a clearly defined task, the steps the agent only prepares and the steps it may carry out. Agents work only within the configured access and permissions. The people responsible review the prepared results and approve them before changes are implemented in connected systems.
Start with a clearly defined task
A clearly defined task shows which information, systems and permissions an agent actually needs. It also makes it possible to check whether its results are correct. With Hawk IVO, the introduction follows four steps:
- Initial conversation: We discuss your task, the teams involved and your goals.
- Define the use case: Together, we determine which information and systems are needed and how you will measure the benefit.
- Pilot with real users: We set up the use case and support you during the first weeks of operation.
- Expand: Additional teams, sites or tasks are added step by step.
Prepare or carry out
Agents work in three steps, which are described in detail in What is an Agent Workforce?. For each step, you define how far the agent may go and where a person decides.
| Step | What the agent does | What you define |
|---|---|---|
| Use knowledge | It draws on relevant information in the right company and project context. | Which sources are approved |
| Prepare results | It brings information together and creates drafts and working materials. | Who reviews and selects |
| Carry out work steps | It handles suitable steps in connected systems. | Which connections and permissions apply and what is approved beforehand |
A simplified example from planning shows how the work is divided. Hawk IVO Plan proposes a staffing plan based on rules and qualifications, and an agent summarizes conflicts and prepares options. The people responsible give their approval. Only then are the changes implemented in the connected systems. Read more in the article Shift planning with AI.
Tasks in which the agent prepares and a person reviews the result before it is implemented are well suited for getting started. Which steps an agent can carry out directly in a system depends on the configured connections and approvals.
Approvals and permissions
With Hawk IVO Force, agents act only within the limits of what you approve. Two kinds of settings need to be distinguished. A permission defines what an agent may do in a system. An approval in the workflow confirms a specific result before it takes effect.
The following applies to an agent’s framework:
- Agents use only configured connections and work within the permissions you define.
- Actions in connected systems are carried out in a controlled manner.
- Reusable work instructions are stored as skills.
- Company knowledge comes from approved sources.
Traceability
An agent becomes traceable when it is known what it works with and where it can act. The settings defined during the introduction provide the basis for this. Knowledge comes from sources you have approved, and recurring tasks follow stored skills. Actions are taken only via configured connections. When the people responsible give their approval, the decision remains with a named person. How drafts, reviews and approvals are documented in detail is part of defining the use case.
What the EU AI Act means by human oversight
The EU AI Act (Regulation (EU) 2024/1689) governs human oversight of high-risk AI systems in Article 14. These systems are to be designed and developed in such a way that they can be effectively overseen by natural persons during the period in which they are in use. Under paragraph 4, the persons overseeing them should, among other things, be able to
- understand the capacities and limitations of the system and monitor its operation,
- remain aware of the tendency to over-rely on the output,
- correctly interpret the output,
- decide, in any particular situation, not to use the system or to disregard, override or reverse its output,
- intervene in its operation or interrupt it.
Under Article 26(2), deployers of such systems assign human oversight to persons who have the necessary competence, training and authority. Whether a specific use qualifies as a high-risk AI system is determined by Article 6 and the annexes of the regulation and is a legal question that depends on the individual case. Regardless of this classification, the criteria can serve as guidance when you define who makes decisions and who can intervene in an agent’s work.
Website chat, knowledge article and executing agent
Not every AI application acts in systems. The difference determines what control is needed.
| Aspect | Website chat | Knowledge article | Executing agent |
|---|---|---|---|
| Purpose | answers questions in a conversation | explains a topic | takes on a task through to the result |
| Basis | message, open page, knowledge articles | editorially maintained text | company knowledge, skills, connected systems |
| Effect | non-binding answer | information | steps in connected systems |
| Control | explicit start by visitors | subject-matter approval of the text | permissions and approvals by the people responsible |
Practical example: the AI assistant on this website
The AI assistant on hawkivo.com runs on Hawk IVO Force. Its task is narrowly defined: it answers questions about Hawk IVO, the products and the open page and suggests relevant knowledge articles in the language of the page. It does not carry out work steps in other systems.
You decide whether the chat starts. It only loads when you click “Start chat”. Before that, a notice explains that you are chatting with an AI, that answers may contain errors and are not binding, and which data the assistant receives. It learns which page and which section are open. It does not receive what you type into form fields.
On a small scale, the assistant shows the same settings as an agent in a company: a defined task, a fixed scope of information and a human decision at the start. Read more about this deployment in the case study on the website assistant.
Checklist for getting started
These questions should be answered before a pilot:
- Which clearly defined task does the agent take on, and where does it end?
- How will you measure the benefit?
- What information does the agent need, and from which approved sources does it come?
- Which connections will be set up, and which permissions apply within them?
- Which steps does the agent only prepare, and which does it carry out?
- Who reviews and approves the results before changes take effect?
- How are drafts, reviews and approvals documented?
- Which users does the pilot start with, and by what criteria are further teams or tasks added?
Software vendors who want to bring agents into their own product will find the relevant questions in the article Embedding agents in your own software. If you have such a task in mind for your own company, get in touch.