Agentic analytics · for finance and BI teams on an enterprise data warehouse · runs in your environment

Your own team of analysts.On your servers, on a schedule, under rules they cannot break.

AgenticObjects is an agentic analytics product that runs on top of your data warehouse. Its agents scan on a schedule, find what changed before anyone asks, and answer when you do. Every number in their sentences comes from a real query, not from the model. Every finding is kept as a record with its evidence and its address.

Start a 4-week pilot How a number is proved

Four weeks. Your data stays where it is. You know how success will be measured before week one starts.

read from your warehouse

revenue_net@last_month

inventory_turnover@last_month

AI Governance

+88.72% 58.53% 4-week average

held/tagged, with its reason

a figure with no source a figure that did not match its evidence

published records

FINDING 1revenue_net@last_month · e1.r1.v.1uevidence chain

FINDING 2inventory_turnover@last_month · 6 branchesevidence chain

RECOMM.based_on → 01M1B399HJ93R8075A761ESV8Nevidence chain

BRIEFcovers 3 findingsevidence chain open

AI writes the sentence. The engine produces every number in it, and checks it before delivery.

continuous intelligence

Continuous analytical attention,
defined once, governed everywhere. One agent definition. Two modes of work.

Role, data scope, permissions and spend limits are set once, at the agent level. A scheduled discovery run produces the persistent records your team shares; a question you type answers you directly. Neither widens the envelope the other works in.

Autonomous analyst

A scheduled run, start to finish: the finding, the evidence under it, and the brief it goes into. Demo data, twenty-seven seconds, no narration — open a derivation and the engine shows its own arithmetic, in its own spelling.

It works before you do.The agent runs on its own schedule, from its own job description. No dashboard can do that: it does not wait to be opened.
Still defensible in six months.Every brief carries an identity, an author and a time, and its evidence stays attached to it.
Recurring is counted, not re-alarmed.The same finding twice comes back stamped ongoing, not filed again.

what one run may publish — at most 6 findings · 3 recommendations · 1 brief

Conversational analyst

A question in plain language, then a follow-up over the same evidence. Demo data, one minute — the answer states what the data cannot separate, and discloses a value the engine corrected against the source.

Checked before you saw it.Metric looked up, query built from it, values fetched, then the sentence tested against them. The count and the time are on the card.
The engine wrote the numbers.The model wrote the sentence around them — dotted figures were written by the engine from the evidence is the product’s own line.
Follow up without starting over.A finding opens in the conversation with its proven figures, each keeping its address. No new run starts.

what comes back with every answer — answer card · evidence chain · assumptions written out

What changes

Answers you can act on. Records you can defend. Runs you can afford.

For the executive

The brief is on your desk before the question forms.

Scheduled agents scan overnight and file what changed. That is continuous intelligence in practice: the analysis is already done when the day starts.

For the data team

Fewer confident wrong numbers in circulation.

A draft answer whose figures have no source is stopped before delivery, and the agent fixes most of them on the next attempt. Ad-hoc requests fall. Governance stays with you.

For the CIO and audit

“What did we base that on?” still has an answer six months later.

Authority is sealed at creation and re-checked on every read. Each run names the model that wrote it, and cannot change model half-way. Spend is capped before the run starts. Nothing is deleted, only archived.

The thing the product is named after

What is an agentic object?

An agentic object is a persistent business record produced by an analytics agent: a finding, a recommendation or a brief. It carries its own identifier. It stays bound to its evidence, to the run that produced it, and to the records around it. The engine decides whether it may be published; the model never can. And nothing is deleted. There is no delete path in the product, only archival. Months later you can open the record and see exactly what it rested on.

Monday Executive Briefing · CFO Agent · 08:10 · scheduled

BRIEF

NEWLast week of August: two points of attention in the revenue mix, one region in inventory turnover.

covers 3 findings · evidence chain open

FINDING

HIGHNet revenue is below its 4-week average; the gap is concentrated in two product groups.

revenue_net@last_month · 01–27 Aug · evidence 3/3 · e1.r1.v.1u

FINDING

ONGOINGInventory-turnover variance is concentrated in six locations.

inventory_turnover@last_month · 6 branches · also seen last week

RECOMM.

TO DOStart a price / inventory review in the two highest-impact locations.

based_on → 01M1B399HJ93R8075A761ESV8N

Illustrative example. Identifier and evidence-address formats shown are the product’s actual formats.

IdentityIts own identifier, and its own address on your system. Cite it in six months and it is still there, at the same place.
EvidenceWhich query, which period, which value, and the address every number was read from. A figure that cannot be proven never enters the body; it falls into a held-values list with its reason.
LinksA recommendation opens the finding it rests on. A brief opens the findings it covers. The engine enforces the direction, so the chain cannot loop.
JudgementTrust tier, novelty and verification are written by the engine only. The model can propose a record. It can never stamp one.

This is what an analytics answer usually is not. You ask a copilot a question, the answer arrives, you close the window, and six months later, in an audit, “what did we base that on?” has no answer. The full anatomy →

Deployment and control

Your warehouse. Defined access. Recorded runs.

The app, runtime, semantic layer and object repository run on your server. The only components that touch data are read-only gateways, and the model sits outside that line: it receives the sealed results it needs to write the sentence, and nothing else.

Governed

Authorization, row-level security, spend limits and an append-only audit log are part of every run, not settings added later.

Private

Read-only gateways; no copy, no write path. Database credentials stay encrypted on your server and never reach us.

Model isolation

The model cannot reach the database, the authorization layer or query execution: it reads text and writes text. And the model you approved is the one that runs. It is chosen before the run starts, never swapped part-way through, and named on every record the run produces.

Budgeted

Steps, seconds and cost are set before a run and freeze when it starts. An agent can narrow its limits, never widen them, and a run that stops early records which limit stopped it.

Bounded

Perceptive analytics means noticing what deserves attention, not reporting everything. At most six findings, three recommendations and one brief per run reach a reader. Overflow is kept and counted, never dumped.

SQL data warehouses through the relational gateway · SQL Server Analysis Services cubes through the SSAS gateway · connector availability for your platform is confirmed in week one of the pilot.

See the architecture →

This is what an analytics agent looks like when it has to show its work.

Perceptive, continuous, governed. Four weeks. Your data. Your acceptance criteria, written first.

Start a 4-week pilot Ask us the 12 questions

The team that builds AgenticObjects: 20 years of Microsoft and SAP business intelligence at D-CAT Technologies, for more than 300 brands. Those were consulting engagements, not AgenticObjects deployments.

Before you do either: what this product is not, and what is still next · the two papers behind it