IT agents. A person in IT decides. Tickets and logs stay.
Katonic runs AI agents inside your own data centre or private cloud. Agents read your tickets, your runbooks, your policies and your change records, prepare the work, and stop for a named person in IT.
Tickets, logs, access lists and runbooks stay on infrastructure you control.
Tickets, logs, access lists and runbooks stay on your infrastructure
A model you serve
Read and change are separate tools
A tool that changes something waits for approval
Tickets, access requests, change requests, alerts and licence statements run on reading, comparing and drafting. That is where an agent helps most, and where a reset done for the wrong caller or a change made unapproved costs most.
What the rules ask
Access follows a written policy. A change goes through the change process. A reset is a person's act after the company's own checks. Every action is logged to a person.
Of 14 public IT workflow templates we opened, 3 have a step where a person approves before a change, 4 only route a request or notify a person, and 7 have no such step. That is a count of the templates we opened, not of the market.
Some suppliers of AI for the service desk run on the customer's own infrastructure, as Katonic does. Running there is not enough: what matters is which tool waits for a person, whether reading and changing are separate tools, whose permissions the agent uses, and what is recorded.
National cyber security agencies have described callers who pose as staff and ask a service desk to reset a password or a sign-in factor. Public reports describe agents with change rights removing a live database. So the agent nearest a reset lists the checks and resets nothing, and an agent that only needs to read is given no tool that changes anything.
Tools must be chosen one by one: we ran Katonic's importer on 158 public interface descriptions from 64 publishers of systems an IT department uses. 132 import, 24 are over the 5 MB import limit and 2 are refused. The 132 carry 31,951 operations, of which 16,969 are sent as changes, and the method alone cannot tell which of those only read. The counts are per file, and some publishers list several versions of one interface. The publishers are identity, device management, ticket, on-call, metrics, source control, cloud, virtualisation, network and collaboration systems. Several of the most widely used service desk and backup products publish no description a file address reaches: the description comes from your own system. From a description over the limit, you import the part you need.
Tickets, logs, access lists and configuration records show how a company's systems are built and who can reach them, and many hold staff data. They have to stay inside the company.
Agents read and draft. People in IT keep every decision about the reset, the access, the change and the incident.
What changes
What does not
A person makes each decision.
Each is a use case an IT department can build in Katonic Studio, in any industry. A name with a link has its own page.
Every agent follows one rule: it prepares, and a named person in IT decides.
Start here · 01
Reads a new ticket, says what kind of request or fault it is, lists what is missing, and proposes the priority and the team. A service desk analyst confirms. The agent assigns nothing and answers nobody.
Start here · 02
Access Request Preparation Agent
Reads an access request, checks it against the role rules and who must approve, and prepares the request for the approver with what is missing. The system owner or the manager approves, and an administrator grants the access. The agent grants nothing.
Start here · 03
Reads an alert, pulls the related metrics and logs, and drafts an incident summary with the runbook step that fits. An engineer carries out any runbook step.
01
It runs on the company's own infrastructure or private cloud,
with a model the company serves. Tickets, logs, access lists and runbooks stay there.
02
Every tool call passes the company's policy.
A policy can allow a tool, record its use, hold it for a person's approval, or block it, for the whole company or for one team.
03
A tool that changes something waits for approval
unless the company has decided otherwise for that tool. That is how no reset, access grant, production change or device action happens on an agent's word.
04
Read and change are separate tools.
When the company imports an API, it chooses which operations become tools. An agent that only needs to read is given the read operations and nothing else.
05
Each statement shows its source.
An agent that summarises or checks gives the ticket, the record or the article, and says when it found nothing. It does not fill a gap from memory.
06
The agent acts as the person using it.
A knowledge source keeps each reader's permissions, so a colleague does not see through an agent what they could not open themselves.
07
Every call is recorded,
with who asked, which tool, and what was decided, and an agent, a tool or a whole server can be switched off at once.
The rules
The rules on access, change, resets and staff data differ by country and by sector and point the same way. Each is a matter of how the system is built.
Trust center →This is a summary of published rules, not legal advice.
An agent reaches your records and systems in one of four ways, all through the same tool gateway and the same approval rules.
Connection · 01
A change request, an export from the asset register, a licence statement from a supplier, an incident chat log.
Connection · 02
Runbooks, how-to articles, IT policies, past incident reviews and closed tickets, indexed with each reader's permissions kept.
Connection · 03
↗Give Katonic the description of an interface you already run, and choose which operations become tools. This is the route to the ticket system, the directory, the device manager and the asset register. From a description over the 5 MB import limit, you import the part you need.
Connection · 04
↗Servers from the catalogue, deployed inside your own cluster, for tickets, on-call alerts, metrics and logs, chat channels and reporting databases.
System names are left out on purpose. You supply the address and the sign-in. Nothing here is a tested integration with a named product. The catalogue has no connector for a directory, a device manager, an asset register or licence records: those are reached through your own API. Where a system you connect is a hosted service, the calls to it go to that service, as they do today.
Five ways to build an agent. Every one ends on the same governed runtime.
A guided builder. Step by step, no code.
A conversational builder. Service desk leads describe how they check a reset request against the procedure or a change request against the change policy. The agent is drafted for a developer to refine.
A workflow canvas. The steps laid out as a flow.
An IDE with the SDK. Full code, for engineers.
Bring your own agent. An agent already built on an open-source framework moves onto the same runtime, with testing, guardrails and monitoring added.
A worked example
A caller phones the service desk, says he is a colleague who has lost his phone, and asks for his second sign-in factor to be reset. The analyst opens a ticket. An agent reads the ticket and the company's reset procedure and lists the checks the procedure requires for this kind of request: a call back to the number on record, and the manager's confirmation. Beside each it says whether the ticket shows it was done, and here neither is on record. The agent verifies nobody and has no tool that resets anything. The analyst rings the number on record, speaks to the manager, notes both checks in the ticket, and then carries out the reset herself. Had a check failed, she would have followed the procedure and the reset would not have happened. The agent read what the company already has, changed nothing, and left a record.
Seven questions to ask any supplier before an agent touches an account or a production system, with our answers.
On infrastructure you control: your data centre or your private cloud. The model is one you serve there. Where a system you connect is a hosted service, the calls to it go to that service, as they do today.
What the person using it can read. Knowledge sources keep each reader's permissions, so a colleague does not see through an agent what they could not open themselves.
No. The reset request check agent lists the identity checks your own procedure requires and whether each is on record for the request. It verifies nobody and resets nothing. A service desk analyst verifies the person and carries out the reset.
A named person in IT decides. A tool that changes something waits for approval unless you have decided otherwise for that tool. No agent grants or removes access, changes production, approves a change, closes an incident or wipes a device.
An agent can call only the tools it has been given. Read and change are separate tools: when you import an interface you choose which operations become tools, and an agent that only needs to read is given the read operations and nothing else. A tool that changes something, where you allow one, waits for a named person's approval unless you have decided otherwise for that tool.
Tickets, on-call alerts, metrics, logs and chat channels through servers from the catalogue, deployed in your own cluster. The directory, the device manager, the asset register and licence records through your own API: you give Katonic the description of the interface you run. No product is named here because nothing here is a tested integration with one.
Who asked, which tool ran and what was decided, for every call. An agent, a tool or a whole server can be switched off at once by your administrators.
Go deeper
Guides and answers
An IT team that wants AI agents on its own infrastructure has more than one kind of product to choose from. The products sold for this fall into four types: agent platforms you run yourself, infrastructure platforms, automation platforms with agent features, and developer frameworks. Our comparison describes fifteen of them with the same fields, read on each vendor’s own pages: where each runs, who operates it, governance, price and limits.
For an IT team, approval before an account, a device or a production system is changed, and separate tools for reading and for changing, matter most.
On-premises AI agent platforms compared: 15 options by type →An IT how-to Q&A agent can run on the first day, because no account or device is touched. Talk to our team about building it on your infrastructure.