The useful question about an AI agent in a company is not what it is capable of, but what it concretely does all day. The popular, and often inaccurate, assumption is that these agents are deployed to watch, monitor, or supervise core infrastructure and systems. We want to address this misconception directly: monitoring infrastructure is a deterministic task that is best handled by established tools, and it is not the primary function of our agents.
Machines watch machines. Our agents do the work that would otherwise land on people. This division of labour means the agents hold the recurring work that consumes a team's week without ever being anybody's project. What is left for people is the judgement, not the assembly and the formatting.
What we do not use agents for
Infrastructure monitoring is a deterministic process, and we rely on dedicated tools such as Prometheus and Grafana for metrics and alerting, alongside uptime probes for availability checks. These systems are designed for the specific purpose of observing system health, and they provide a level of certainty that an AI agent cannot replicate. When a system metric deviates, the alert is clear, the rule is simple, and the failure point is immediately visible.
Attempting to use an agent for this kind of monitoring would be inefficient and overly complex. A simple, pre-written rule is cheaper, cannot hallucinate, and cannot be argued against using a natural language prompt. Furthermore, agentic monitoring would only make sense if the agent could autonomously decide what action to take based on what it saw, but our agents do not make such decisions. The same Prometheus and Grafana stack we use to monitor infrastructure is also what we use to watch the agents themselves.
What the agents actually do
The value of an agent lies in the recurring work that cannot be reduced to a single condition, and that would otherwise sit in a person's week. Agents manage the triaging of incoming enquiries, including the decision to decline one and the reason recorded with it. They also handle research and preparing text, checking claims against sources, and collecting public tenders and listings, classifying them and recording them in the ERP.
The agents also check bookkeeping and payroll calculations against the underlying documents. They are also responsible for compiling daily summaries, writing up findings and defects, and generating work logs. These functions ensure that the administrative record is kept current and that information is systematically captured for future reference.
Not everything we run under an agent identity is a large language model. Part of the fleet is a language model with its own identity and its own judgement; part is deterministic code that simply runs under a name. The tender collector is the second kind: a scheduled job, with no model involved. We draw this line explicitly because a reader who works it out for themselves stops believing the rest, and because the two kinds carry completely different risks: deterministic code fails visibly and predictably, whereas a model can be confidently wrong.
None of this removes the need for judgement. It removes the part of the work that consists of assembling, formatting and remembering.
Why an agent has a personnel record
To ensure that the use of an agent is auditable and controllable, it must be treated as a specific entity within the organisation, rather than a shared service account. Each agent is assigned its own distinct identity, including a named user, a dedicated email address, and a personnel record. This prevents the ambiguity that arises when multiple processes share a single set of credentials.
Furthermore, every agent is supplied with a written brief that clearly describes its purpose, the specific systems it is permitted to touch, and, critically, what it must never do. This boundary is supplemented by a human supervisor, and all work that arrives for the agent is treated as assigned tasks, much like any other employee's workload. The agent must log its hours against the specific task it worked on, ensuring that the work is visible as an accountable activity, rather than appearing as an unexplained change in a system log.
The security architecture ensures that permissions are enforced by the platform itself, not by the agent's instructions or prompt. Credentials are held behind a secure gateway that injects the agent's identity for every single call, meaning the agent never holds the raw token for the business system. This design means that if access needs to be revoked, it is a single change in one location, closing every potential door at once, mirroring the security process of offboarding a person. This structure makes the entire process fully auditable: you can trace exactly who did what, because a specific, named entity did it.
Where the human gate sits
Crucially, nothing leaves the organisation or impacts a system without a named person explicitly stating that it is approved. An approved draft is not the same as final authorisation to publish; these are two separate decisions that require two separate, distinct actions. Similarly, no bulk or destructive operations can be executed without real-time human confirmation and sign-off.
The data contained within a record is always treated as data, never as authorisation. If an instruction is embedded within a document, an email, or a ticket, it does not automatically become a command simply because an agent read it. An agent that treats all content as instruction can be steered by anyone who can write into any system it reads.
What this does not solve
Agents are highly effective at consistency and breadth of coverage, but they are inherently poor at knowing when they are wrong. They will produce a well-structured and convincing answer even when it rests on a misunderstanding of the data. An agent also has no way of setting priority, or of telling which work matters most to the business.
Where the work can be written as a deterministic rule, a scheduled cron job or a defined pipeline, it gets written that way and no agent is involved. We scoped one for a platform of ours once and itemising the brief dissolved it completely, which is a story worth reading on its own: the agent we decided not to build. The accumulated rules about how an agent behaves in your organisation are worth more than the choice of model. A model can be swapped out in an afternoon, but the body of operational rules governing its behaviour cannot be changed so quickly.
Questions and answers
What does an AI agent in a company actually do all day?
Our agents primarily handle the recurring, administrative work that consumes human time. They are used for tasks like drafting initial reports, maintaining data hygiene in the ERP, or compiling summaries from various sources. They automate the mechanical processes of assembly and formatting, allowing personnel to focus on judgement.
Should I use an AI agent to monitor my infrastructure?
No. Infrastructure monitoring is a deterministic function best handled by established tools like Prometheus and Grafana. These tools provide clear, visible alerts based on defined rules, which is a more reliable and auditable method than using an agent.
How do you stop an AI agent from doing something you did not approve?
We enforce human gates at every point of output. Nothing leaves the organisation without a named person confirming its release, and no destructive operation can proceed without real-time human confirmation. The system also treats anything the agent reads as data rather than as an instruction, so text sitting in a document or a ticket cannot become a command just because the agent encountered it.
How do you audit what an AI agent did?
Because each agent has its own unique, personnel-level identity, every action it takes is logged against that specific user. The audit trail is therefore specific: you can trace which named entity performed which action.
The decision of where an agent will pay for itself should not be based on looking at dashboards or monitoring metrics. Instead, the focus should be on identifying the recurring, administrative work within a team's week that nobody currently has the time or capacity to do. That is where an agent earns its place, and it is a far less exciting answer than the one the market usually offers.