It plans, it explains itself and it waits for your yes.
An agent working in real data is only useful once you can follow how it arrived at a proposal. Four layers sit underneath — context, plan, tools, release — and the third of them is the only one that writes.
Input
Chat
A question in words
Session
Context, history, spend
Orchestration
Plan
Steps with goals
Lists
A selection that stays
Subagents
Analysis, planning, titles
Execution
Tools
31 declared, none builds its own
Actions
Declared per object type
Query path
One route for click and sentence
Record
Customers
Master data, balance, limit
Open documents
Amount, due date, currency
Reminders
Level, deadline, dunning stop
Clarifications
Reason, owner, effect
No path into the record goes around it. One run, from the question to the record — and the one place it stops.
From the question to the release — in that order, every time.
None of the four can skip the next. That is what makes it possible to say anything binding about an agent at all: a proposal without context does not arise, a write without a plan does not arise, and a plan without a release does not run.
-
1
Context
The agent reads the state you see — with your permissions and your data scope. A view is a query rather than an extract: it is re-executed every time anyone looks.
-
2
Plan
The context produces a proposal carrying its consequences — a chain per case whose steps are named one by one, with the expectation that is checked before anything goes out.
-
3
Tools
The agent acts through declared actions. No tool writes on its own, none invents an action, and for the release there is none.
-
4
Release
At the end there is a person. The consequential steps come back as a card, and whatever nobody released stays lying.
All plans › Group today's dunning notices by review need
Dunning agent · less than a minute ago
Re-cut by review group
Send the 3 notices in the “clear case” group, provided they are still dispatchable — an explicit sign-off on the safe bulk send.
and 2 more.
Look at the notice whose dispute block is holding documents back, and decide whether it goes out incomplete today or waits until the block expires.
You watch along while it computes.
While an analysis or a Mahnlauf (dunning run) is being worked out, the Denkweg (reasoning path) runs alongside it: which datasets are being read, which round it is in, what it concludes. No result falling finished out of a black box — and afterwards the same reasoning stays with the artefact.
- Step by step, not at the end
- The path is visible during the work rather than as a summary after it. Anyone who wants to check a recommendation sees what it rests on before it is made.
- Every statement names its figure
- The prose points at the datasets it came from. A statement without a reference is not something the result provides for.
- Nothing is left unread
- The whole data basis is computed before a word is written. So a report cannot say "not considered" — the only gap it can name is what your own permissions do not reach.
No number comes from the model. Not one.
Counts, sums, rankings and dates are computed from the database — permission- and scope-filtered, formatted in one place. The model selects, orders and phrases. It does not calculate.
- From the database
- Every figure arises in a query, not in a sentence. What the model returns is prose and references to computed datasets — never the values themselves.
- Checked twice
- What the model delivers is validated against a declared shape when it is stored, and again when it is rendered. An output that does not fit is not rendered.
- Your ladder, not ours
- Where an evaluation asks about age it uses your company's own aging ladder and its words. Where none is configured the evaluation is left out — a wrong ladder is worse than a missing one.
- A measured nothing is said
- An evaluation that found nothing says so rather than disappearing. But only where the emptiness is itself the finding.
Where something is unclear, it does not guess.
Where several records fit your question, the agent asks back with buttons rather than a numbered list — pressing one is the same as typing that answer, and typing something else always works. Where a value is missing for an action, a short form asks with a fixed subject and exactly the fields the action provides for.
What a result cost is stated beside it.
Every plan, analysis and summary states its own consumption, and above the composer a meter counts the whole session. Support can read the accounting per result — and where a company asked for the recording, the technical history behind it too.
Session
3 results
What the agent explicitly does not do. In full, not curated.
The part against which the rest can be checked. It stands the same way in the product documentation, there alongside seventy-three further recorded limitations.
- No number comes from the model. Counts, sums and rankings come from the data.
- It sends nothing of its own accord. The consequential steps come back as a release card.
- No reversal. An executed action stands in the real data.
- It creates, changes and deletes no setting and no template.
- It never writes a comment unasked and never changes a note field.
- No aging evaluation without your company's own aging ladder.
- One analysis per object at a time — enforced in the database, not by asking nicely.
What people want to know before they put an agent to work.
Then it does not appear. Counts, sums, rankings and dates are computed from the database, and what the model returns is validated against a declared shape when it is stored and again when it is rendered. The model selects, orders and phrases — it does not calculate.
Yes, and while it is coming about. The Denkweg shows step by step which datasets are being read and which round the agent is in; afterwards the same reasoning stays with the artefact, together with the figures it rests on.
The agent asks back, with buttons for the candidates it cannot tell apart — each carrying the detail that identifies it. Pressing one is the same as typing that answer, typing something else always works, and neither path resolves anything by guessing.
No. Every change goes through a declared action and its executor, the same one a toolbar click takes. No tool builds a path of its own beside it, and the agent can propose and rank the declared actions but not extend them.
Every plan, analysis and summary states its own consumption, and above the composer a meter counts the whole session. Support can read the accounting per result, and where a company asked for the recording, the technical history too.