It runs on your own server with read-only access and records every Finding with a link to its source. Interpretation and decisions stay with you.
AgenticObjects product · Implemented by D-CAT
Role, data scope, permissions and spending limits are defined once, at agent level. Both modes work within the same boundaries, and neither can extend the other's permissions.
The agent works to its own schedule and its own brief. It scans the data, finds what has changed and turns what it finds into persistent records your team shares. By the time the day starts, the analysis is done.
When the same Finding comes up again, no new alert is raised; the record is marked as ongoing.
You ask a question in plain language. The metric is looked up in the semantic model, the query is built from it, the values are retrieved and the sentence is tested against those values. The model does not write SQL.
You can open a Finding and keep asking questions about the same evidence, without starting a new run.
An agentic object is a persistent business record produced by an analytics agent. It stays linked to its evidence, to the run that produced it and to related records. No record is ever deleted, only archived.
How a metric behaved for a specific breakdown in a specific period. Every Finding carries a severity level; which metric, which period and which filter are written by the engine, not the model.
A record for you to evaluate: a suggestion based on at least one Finding. A Recommendation without supporting evidence cannot be produced, and a suggestion whose only basis is “no data” is rejected. The decision is yours.
A cover note that brings a run's Findings together in one readable text. Every Brief is linked to the Findings it covers, and from there the evidence is one click away.
The same record answers a different question for executives, the data team and auditors.
Scheduled agents scan overnight and record what has changed. By the time the day starts, the analysis is ready.
A figure without a source is stopped before delivery. Governance stays with you, and the semantic model is the single source of definitions.
Permissions are sealed when a record is written and checked again on every read. Every run records the model that wrote it and what it spent.
The language model writes the sentence; every number in it is produced by a deterministic engine and verified before delivery. The implementation follows the same principle.
The agent's raw output never goes straight to the dashboard. Every candidate passes rule-based publishing gates, and a figure whose source can't be traced cannot enter the body text. Every candidate that is stopped stays on record with the reason.
Only read-only gateways touch the data. Nothing is copied, there is no write path, and database credentials stay encrypted on your server.
The application, the agent runtime, the semantic layer and the record store run on your server. It connects to your SQL data warehouse and SSAS cube through read-only gateways. The language model can run in the cloud or locally; it has no access to the data path, permissions or query execution, and receives only text.
Permissions, row-level security, spending limits fixed before a run starts, and an append-only audit log. The model doesn't change mid-run, and its name is written on every record.
Three questions at a glance: where does the application run, who touches the data, and where does the AI model sit?
Outside the frame. In the cloud or local. Only text goes out, never data; the model is fixed for each analysis.
Application (end users) and console (administrators)
The agent runtime writes the narrative; the query engine calculates; FinalGuard checks at the gate
One official, versioned definition for every measure
Run and audit log, roles, budgets
SQL and SSAS gateways, both read-only. No writes, no copies; your password never reaches us.
Data warehouse (SQL) and SSAS cube. The data never leaves here.
Query building, query execution, figure checks, permission decisions and spending control all stay outside the model.
A tool with clear boundaries from the start can be evaluated clearly in a pilot, too.
It connects to source systems read-only; it doesn't copy data and has no write path.
It records what changed and where the change is concentrated, together with the evidence; interpretation and decisions stay with people.
Agents work to their own schedule and record Findings in each scheduled run.
It works on top of your semantic model; the language model only writes the sentence.
AgenticObjects creates value on your data, with your definitions. As implementation partner, D-CAT owns the implementation from start to finish.
Read-only connection to your SQL data warehouse or SSAS cube, user roles and row-level security.
Metric and dimension definitions are written in Axoria Data Studio and handed over to AgenticObjects as a versioned .axr file.
For each agent we write, together with you, the executive it serves, its priorities, the areas it looks at, the things it must never claim and its run schedule.
Acceptance criteria are written down before the first connection, and the 4-week pilot is evaluated against them.
After the pilot, new business areas, new agents and new user groups are brought on board step by step.
Model updates, brief adjustments, review of user feedback and operational support.
AgenticObjects doesn't write the semantic model itself; it imports a versioned model. The model is prepared in Axoria Data Studio, D-CAT's data warehouse studio that works with AI agents. If you don't have a model yet, we write one together for the pilot's first area.
The experience behind it: 20 years of Microsoft and SAP business intelligence expertise and 1,000+ projects, with hands-on experience in data warehouse, SSAS and semantic model design.
AgenticObjects product · Implemented by D-CAT
The decision at the end of the pilot doesn't rest on how impressive the interface looks, but on evidence, auditability, cost and real usage signals.
Before the first connection: every figure shows its source in one step, every assumption is written out, the spending limit is set before the analysis starts, and every operation is in the audit log.
Data warehouse connection, user roles and row-level security, import of the semantic model, and the first question-and-answer screen with the evidence panel.
A scheduled scan of a business area you choose. We write the first agent's brief together and set a weekly rhythm; the first Brief arrives this week.
A second business area, executive users, a feedback loop, and evaluation of the pilot against the acceptance criteria written in advance.
It shows which Findings, Recommendations and Briefs your model can support today, and together we decide which business area the pilot should start with.
AgenticObjects product · Implemented by D-CAT
Short answers to the questions we hear most often on this topic.
No. Only read-only gateways touch the data; nothing is copied and there is no write path.
The language model writes the sentence; every number in it is produced by a deterministic engine and verified before delivery. A figure that can't pass the FinalGuard publishing gates, or whose source can't be traced, is not published.
The pilot takes 4 weeks. Acceptance criteria are written down before the first connection, and the pilot is evaluated against them. One business area, one data warehouse and 3–6 executive users are enough.
D-CAT is the implementation partner for AgenticObjects: it sets up the data warehouse and SSAS connection, writes the semantic model in Axoria Data Studio, prepares the agents' briefs and KPIs with you, runs the pilot and provides support afterwards.
The application, the agent runtime, the semantic layer and the record store run on your server. It connects to your SQL data warehouse and SSAS cube through read-only gateways. The language model can run in the cloud or locally; it has no access to the data path, permissions or query execution, and receives only text.
The semantic model is written in Axoria Data Studio and handed over to AgenticObjects as a versioned .axr file. AgenticObjects doesn't write the model itself.
One business area, one data warehouse and 3–6 executive users are enough. D-CAT runs the implementation, and success criteria are written down before the first week.
