AI agents for companies.
AI agents for companies are software systems that handle a recurring process independently - checking incoming orders, say, or assembling an order confirmation - while drawing on your own data. They run inside your private AI infrastructure, assisted with an approval step or autonomous, depending on how much trust the process has already earned.
The problem AI agents are meant to solve
In many companies, a recurring but not especially complicated process ties up a disproportionate amount of qualified staff time. Checking invoices, presorting emails, matching orders, assembling reports - most of this follows clear rules, but demands constant human attention because the rules carry many exceptions. Classic automation fails at these exceptions, because it needs rigid rules and has to be manually updated for every new one. An AI agent can handle unclear, incomplete or contradictory input much the way a person would, while still staying traceable, if it is set up correctly.
The difference from a simple language model in a chat window matters here. A chatbot answers a question and the conversation ends there. An AI agent pursues a goal across several steps, pulling from several data sources as needed, weighing intermediate results, and finally triggering a concrete action in a connected system - creating a record, say, or routing a case to the right department.
How our AI agents are built
An AI agent with us is made up of several parts working together. The foundation is a language model that runs on your private AI infrastructure, as described at private AI infrastructure. Next comes the connection to your data sources through retrieval augmented generation, a method where the model looks up your own documents, databases or systems before every answer instead of relying on general knowledge. That substantially reduces wrong or invented answers, because the model is tied to verifiable sources from your own records rather than formulating freely.
After that comes the tool connection - access to concrete actions such as creating a record, sending a message or triggering a workflow in a connected business system. Each of these actions is individually defined and individually secured, so an agent never gets unlimited access, only exactly the tools its specific process needs.
Last comes the approval logic, which sets when the agent acts on its own and when a person has to confirm. This logic is not a fixed default - it is set individually for each process and can be adjusted during ongoing operations, if it turns out that more or less human review makes sense.
Assisted or autonomous, an adjustable decision
Not every process deserves full trust from day one, which is why the approval step with us is adjustable and never a fixed default.
In assisted mode, the agent prepares a decision - drafting a reply, say, or proposing a booking - and a person reviews and confirms it. This suits new use cases, particularly consequential decisions, and any case where an audit function requires documented human sign-off.
In autonomous mode, the agent acts independently within defined limits and only reports exceptions back. This suits high-volume, low-per-case-risk processes with a clear pattern, once assisted mode has shown over a sufficient period that the agent works reliably.
The move from assisted to autonomous happens gradually, usually for individual sub-steps first, then for the whole process, and stays reversible at any time. If ongoing operations show a rising error rate or a change in the underlying data, we switch the affected sub-step back to assisted mode until the cause is resolved, rather than letting the risk run unwatched.
Typical use cases
AI agents suit processes that are rule-based but carry many exceptions, recur often, and rest on structured or semi-structured data.
In logistics, a typical example is matching delivery notes against orders: an agent reads incoming delivery notes, checks quantities and item numbers against the original order, and flags only the cases with a discrepancy for a person, instead of manually reviewing every single one. In financial services, an agent often handles the pre-screening of incoming applications, gathering the necessary documents and preparing a decision brief. In manufacturing, an agent evaluates maintenance logs and flags patterns that point to an upcoming failure before it actually happens. In public administration, an agent often pre-sorts incoming applications by responsibility and urgency, so caseworkers can focus on substantive review instead of losing time on routing.
A more detailed overview with sector-specific examples is at Industries.
What an agent costs, and what the price depends on
The price for an AI agent depends on the complexity of the process, the number of connected systems and the chosen approval model, not a flat per-user licence fee. A simple agent that queries a single data source and prepares suggestions in assisted mode costs noticeably less than an agent that links several systems and acts autonomously. As with private AI infrastructure, the exact amount is only fixed in writing after the assessment; price frames for orientation are at Pricing.
How we measure an agent's quality
An agent whose reliability nobody measures can neither be improved nor responsibly moved into autonomous mode. So for every agent, before rollout, we define what marks its quality - the share of correctly handled cases, the number of cases returned to a person, and the time a case takes from intake to completion, for example. We track these metrics continuously, not just once at rollout, and disclose them to you regularly, so you can demonstrate the agent's value rather than just assume it.
As with building private infrastructure, introducing an AI agent follows fixed phases. In the assessment, we jointly define which process gets automated and what data it needs. In the pilot, within 72 hours, a first version emerges that your team can test internally, usually in assisted mode. In the rollout, 4 to 10 weeks, this becomes a productive system with the approval levels that fit your process. In ongoing operations afterwards, we monitor the quality of results and adjust the approval logic as needed, as described at Operations.
Traceability as a precondition, not an add-on
An agent that prepares or makes decisions has to be explainable after the fact, or it can neither be revised nor justified to a regulator. That is why every one of our agents logs which sources it drew on for an answer, which action it took, and, in assisted mode, who approved it. These logs are part of the documentation an internal audit or an external auditor can review, and they are also the basis for moving an agent from assisted into autonomous mode, because its reliability can be demonstrated with real data rather than assumed.