Using AI-agents to automate operations – how to get started

This is not because they are difficult to automate as such, but because they require one or more integrations and portals. Traditional integration projects has been too expensive and too fragile to be worth the effort.

These tasks are instead handled manually. They repeat. They consume time. They create dependency on key individuals. And they are rarely visible enough for anyone to prioritise solving them.

In other words: It has not been worth it to automate these yet - however the calculation of what is worth to automate and what is not is being changed by AI agents.

In this domain an agent becomes interesting when it can perform concrete actions in the organisation's systems. This could be creating a user, changing a permission, closing a case, initiating a synchronisation, updating the price on a product, or retrieving a status across platforms.

This is where Model Context Protocol (MCP) plays an important role. MCP makes it possible to expose the actions the agent is permitted to use, in a secure and controlled manner.

A recurring pattern in operations and platform teams is the management of users and permissions.

When an employee changes role, project, or team, access rights need to be adjusted across systems such as Microsoft Entra groups, CMS, PIM, e-commerce platforms, support tools, and internal administration systems.

The task is rarely complex, but it requires a technical member of staff, and it repeats.

For instance: An onboarding process for a new employee involving access to three systems typically takes a technical employee 20-30 minutes. In an organisation with 50 onboardings per month, that amounts to 25-37 hours of technical time. Add to this errors from manual data entry, delays when key individuals are absent, and the fundamental risk of permissions being set incorrectly.

This is ongoing friction that rarely appears on a dashboard but is always there.

With MCP, this type of task can be handled by an agent. For example via Microsoft Teams, where a technical member of staff can interact with the agent.

The prompt could be:

• "Onboard this new employee, he needs access to A,Y and Z systems."

• "Same permissions as David".

The agents then with access to a limited set of well-defined actions: find user, update permission, remove access, log the change, performs these changes. The same process is now completed in under two minutes. The agent performs only the actions it has been explicitly authorised to perform, and nothing else - and reports back once completed.

Another example: Add a 5 discount to all products in a certain ETIM-class, or close all support tickes with no activity for 2 months.

Many MCP solutions work in a prototype and fail in production. Three principles distinguish the solutions that hold.

Actions should correspond to what an experienced employee would do. Actions that directly mirror technical API calls force the agent to navigate details and sequences that are irrelevant to the task. This increases the error rate and makes the solution costly to maintain. Actions should represent clear business decisions with validation built in.

Overlapping actions are the classic source of instability.

If it is not obvious which action applies in each situation, the agent behaves unpredictably and becomes harder to manage.

A limited, unambiguous set of actions is easier to test and maintain. 

An agent presented with too much or incorrect context will, over time, produce worse results. This goes for both a single object - avoid unnecessary fields - and lists. A good analogy is a person. A person would also struggle to manually go through hundreds of items in a list (without tools like Excel).

Targeted lookups work far better than broad queries. If additional detail is sometimes necessary, it is better to expose a separate action for that purpose than to include it by default.

Governance places the same demands on agents as on human users: clear access management, consistent logging and audit trails, and reliable stop mechanisms.

Without these mechanisms, the consequence is rarely a dramatic failure - but a gradual erosion of traceability and control that makes the solution too unreliable for production. 

MCP creates value most quickly on tasks that involve manual steps across two or more systems, are carried out by technical staff even though they do not require technical judgement and occur with high frequency and low business risk.

User administration, permission adjustments, and status lookups are good examples. Add a 5 discount to all products in a certain ETIM-class, or close all support tickes with no activity for 2 months are also examples.

A typical path forward begins with mapping the tasks the organisation stopped automating and assessing whether the calculation has changed. A well-defined scope area with high frequency and low risk is then identified. The action set is defined, the necessary governance is established, and the solution is put into production.

This is precisely what we can help you with - from the initial assessment of which processes make sense to automate (Advisory), to the design and implementation of the MCP setup (Build), to ongoing operations and further development (Run).

Not as an isolated technical deliverable but as part of a broader effort to reduce the friction that slows down day-to-day operations.