Custom agents

    What your best people know, for everyone who is allowed to use it.

    A custom agent is a specialist of your organisation: built by the team that knows the field, approved by the domain that owns it and used in chat, in Flows and in the AI assistant you already have. Knowledge no longer stays with one person.

    Finance analystFinance · v3livewhen to useQuestions on margin, revenue and costownerFinance control30 days
    Ownership

    An agent is a business asset with an owner.

    An agent belongs to a domain, such as finance or operations. Those who know the field build it. Those who are responsible decide whether it goes live.

    Finance analystowner · Financestop · resume

    A domain owns it.

    Every agent belongs to a domain and has an owner group. That group can stop the agent at any moment and resume it again.

    draft

    The team builds.

    The people in the domain set out what the agent is for, when to use it, which tools it has and which data it may read. A test run shows what it does before anyone else sees it.

    Version 4submittedapprovedapprovedtwo different people

    Someone else approves.

    A steward of the domain reviews every version. Whoever submits a version does not approve it themselves. Only after approval does the agent answer.

    Building

    Six choices make an agent.

    The team fills them in and tries the draft with a test run, with the tester's own rights. An agent looks things up and answers. Changes go through Flows, with the approval that applies there.

    purposewhen to usetoolsdataAnswers questions on marginFor revenue and cost questionsmetricstermscatalogfinance.*test run

    Try first, then submit.

    Purpose

    What the agent is for, in plain language. This steers every answer.

    marginplanningservice

    When to use

    One sentence that tells colleagues and assistants when this agent is the right one.

    Tools

    What the agent may look up: metrics, terms, the catalog or a flow.

    Data

    Which sources and tables the agent reaches.

    this far

    Limits

    How many rows, how long to wait and how many steps per question.

    Model

    Which AI model the agent uses, chosen per agent.

    Always better

    Every improvement is a new version.

    The team that uses the agent sees where it falls short and adjusts it. The live version keeps answering while the new one is reviewed. What was replaced stays on record.

    Coming soon

    Benchmarks per agent.

    Measuring whether a new version answers better than the last, so every improvement can be shown. This is not available yet.

    v1v2v3v4replacedin reviewkeeps answering

    The live version answers, the next one waits for approval.

    The check

    A second agent looks along before you see the answer.

    Agents work together. A router names the right specialist and a reviewer assesses its answer. If it approves, the answer is marked as checked. If it rejects, the answer is not presented as fact.

    checkednot checked

    Checked or not: it says so.

    margin and cost

    Specialist

    Answers questions in one field, with its own tools and data.

    Router

    Reads every question and names the specialist that should answer.

    Reviewer

    Reviews a specialist's answer before it is presented.

    Using

    Available to anyone and anything, if it is allowed.

    An approved agent can be used wherever your organisation works. Who may use it is decided by the organisation, per role.

    Finance analystPlanning agentService agent@

    Pick an agent with @ or call it by name.

    In Datahub

    The Agents page lists every live agent you may use, with when to use it, its domain and its owner.

    @

    In chat

    Pick an agent with @, name it in your message or let the router choose. Every consult shows as a step in the conversation.

    In your AI assistant

    Every agent you may use appears as its own tool in Copilot, Claude or ChatGPT. You get the answer and the query used, never the raw rows.

    In Flows and through the API

    An agent can call a flow that is offered as a tool. Other systems reach agents through MCP or the API.

    Works withCopilotClaudeChatGPT+ any assistant with MCP
    Access

    Two ways to arrange access.

    askeragent

    With the asker's rights.

    By default the agent sees exactly what the person asking may see themselves. No more.

    own accessStewardData owner

    With its own rights, after two approvals.

    Sometimes an agent has to answer from data the asker may not open. Then the agent gets its own rights: first the steward approves the agent, then the data owner approves the access. Who may call it is set per group.

    Control

    Always stoppable. Everything on record.

    reasonSource is being replaced

    Stop at once, with a reason.

    The owner group or an administrator stops an agent in one go. The reason is recorded, and whoever asks the agent hears why it does not answer.

    who askedagenttoolsoutcomeControllerFinance analyst v32 · 14 sansweredquestion text not kept

    Every consult on record.

    Who asked, which agent and version answered, which tools were used and how it ended. The question text itself is not kept.

    asks per week12 weeks

    Usage per week in view.

    For each agent, twelve weeks show how often it was asked. That shows which agents do their job and which need attention.

    31decembernotice 7 days ahead

    An end date where it fits.

    An agent with its own rights can get an end date. The owner group is told seven days ahead.

    Unlimited

    As many agents as the work calls for.

    An agent is a colleague, not a licence. That is why the number of agents carries no price and no limit: a team that needs a second or a tenth agent deploys it without an extra Datahub licence. Agents run on the AI model your organisation already buys, or on Private AI.

    Planningthe team's agents

    Two desks, eight agents: the work sets the number.

    Getting started

    Four starter agents are ready as drafts.

    Every environment comes with them. They only read and work with the rights of whoever asks. An administrator adjusts them to the organisation, picks a domain and submits them for review.

    RevenueFinance

    Governance Steward

    What a term means, where it lives, who owns it and whether it is approved.

    Needs: the brain, the glossary and the catalog.
    open€ 12.400

    AFAS Finance Analyst

    Questions on open receivables and payables, ledger accounts, journal lines and sales invoices. No payroll or employee questions.

    Needs: an AFAS connection.

    Flow Run Triage

    Why a flow run failed, stalled or is waiting, what it did step by step and how often it fails. Never starts a run itself.

    Needs: Flows.
    gross margin · Q331.4%

    Metrics Analyst

    How a KPI is defined, which insight shows it and its value for a period.

    Needs: metric definitions and semantic models.
    Frequently asked questions

    What teams ask before they build their first agent.

    Next step

    Start with one agent for one question.

    An intro call looks at which knowledge in your organisation deserves an agent first and who will own it.