31 tools, and none of them builds a second.
An agent has no access to the database, only to a list of tools. Each one has a name, a declared input and an effect fixed in advance — and the writing ones dispatch, without exception, into a permitted action.
manage_dunning_stop
Writing
31 declared
Navigation & view 6
Open, keep, name — no data written.
- open_object
- remember_objects
- forget_objects
- manage_list
- +2
Writing 15
Each dispatches into a declared action.
- manage_dunning_stop
- batch_dunning_stop
- manage_promise_to_pay
- manage_letter
- +11
Analysis & plans 5
Analyse, plan, summarise.
- analyze_object
- create_plan
- summarize_session
- manage_plan
- +1
Reading & questions 5
Read back, resolve filters, ask instead of guess.
- read_comments
- read_artifacts
- ask_choice
- list_filters
- +1
manage_dunning_stop Writing
Declaration
- Input
- one document, one customer, or a named set (max. 25)
- Effect
- sets or removes a manual dunning stop
- Permission
- checked before execution, refused when in doubt
- Confirmation
- from two objects on, always a card
- Batch
- the yes runs as a batch with an account per object
Four groups, one list, no access beside it.
The list is the boundary: what is not declared as a tool an agent cannot do — and it cannot add one either. A new tool is development work with a test and a release, not a setting.
Navigation & view 6
Open, keep, name — no data written.
- open_object
- remember_objects
- forget_objects
- manage_list
- highlight_actions
- save_view
Writing 15
Each dispatches into a declared action.
- manage_dunning_stop
- batch_dunning_stop
- manage_promise_to_pay
- manage_letter
- manage_reminder
- write_comment
- rewrite_letter_text
- rewrite_reminder_text
- manage_customer_group
- create_clarification
- resolve_clarification
- manage_debt_collection
- manage_watch
- export_view
- download_document
Analysis & plans 5
Analyse, plan, summarise.
- analyze_object
- create_plan
- summarize_session
- manage_plan
- manage_batch
Reading & questions 5
Read back, resolve filters, ask instead of guess.
- read_comments
- read_artifacts
- ask_choice
- list_filters
- resolve_filter_values
The names are untranslated here because they are the same in the product documentation. Anyone checking whether an agent can do something searches for the name there.
And there are more every week: new tools come from teams that need them — declared, documented and enabled per company, never as a quiet addition.
Request yoursWriting happens only through a declared action.
An action is a contract rather than a function: it belongs to an object type and carries a kind, a context, a permission and — where it is consequential — a confirmation. The same action runs whether it came from the toolbar, from a sentence in the chat or from a plan.
- Kind
- Immediate, form, export, analysis or menu — the kind decides what an agent can trigger at all.
- Context
- On the object, on a selection or on the whole list. What the context does not allow is not offered.
- Permission
- Every action names its right. If the acting person lacks it, the execution refuses — not the agent.
- Confirmation
- Consequential actions carry a confirmation card. It is part of the declaration, not a courtesy of the interface.
- Batch
- Actions over sets are declared as such and run with an account per object.
The agent can propose and order the declared actions. It cannot extend them, and it certainly cannot grant a release.
Where a button is not enough, a form appears.
Some actions need details nobody should guess: a letter's template, the delivery channel, the date of a promise to pay. Then the tool opens a form, prefilled with what the agent understood — and the person confirms or corrects.
- A form, not an assumption
- Template, texts, documents, delivery — visible, changeable, and only then a draft.
- A question as a card
- Where a detail is ambiguous, a choice with clickable options appears instead of a guessed match.
- Prefilled, not predecided
- What the agent understood is in there. What becomes of it is the person's decision.
Query › Letter · Dunning level 2
Form- Template
- Dunning level 2 – Standard
- Text
- Auto-drafted, paragraph 2 adjusted
- Documents
- 3 of 3 selected
- Delivery
- Email with PDF attachment
What no tool can do, not even deliberately.
These points are architecture rather than settings — they cannot be configured, which is what makes them checkable.
- No tool grants a release. There is none that could.
- No tool changes a setting or a template. It can only name which ones exist.
- No tool writes around an action. The direct path is an action too.
- No tool builds another. The list is code, not a runtime catalogue.
- No tool bypasses a permission. It is checked before execution and refused when in doubt.
- No navigation as an action. A jump changes nothing.
What IT asks about tools.
Not as a setting. A tool is code with a declaration, a test and a release — which is precisely why anything binding can be said about the list at all. What you control are the permissions and the actions your object types carry.
No. Every writing tool dispatches into a declared action and its executor — the same one a click in the toolbar triggers. There is no second write path, and the agent cannot extend the list of actions.
The execution refuses, with the ordinary error message — the action decides that, not the agent. It is checked before execution and closed when in doubt, so also when the agent believed the right was there.
When the action declares one — sending a reminder, opening or closing a clarification, handing a document to collection, and from two objects on as a rule. The card is part of the declaration, not a setting of the interface.
It asks with a card offering the candidates to click. That is why a wrong match is rare: guessing is not provided for in the tool, asking is.