AI BI Agent Data & Analytics Agents Tier 2 On-premise Updated September 2026
AI BI Agent

AI Agent for Metrics & KPI Interpretation

Most organisations do not have a dashboard shortage; they have a trust problem. This agent reads your governed metric definitions and dashboard lineage and answers the questions that actually stall meetings: what this number means, what moved it, and why the other report says something different.

Governed Answers use your official metric definitions
Lineage Traced from the tile back to the source table
Reconciled Disagreements explained, not argued about
Attributed Movement broken into contributing segments
Reads
Metric definitions Dashboard lineage Refresh schedules Filter context KPI targets Historical values

What is an AI BI agent?

An AI BI agent is a governed software worker that interprets an organisation’s existing business intelligence layer. It answers metric questions from the governed definitions, reconciles reports that disagree by tracing lineage back to source tables, and decomposes KPI movements into the segments and periods that account for them.

What it does

Answers from governed metric definitions Traces dashboard lineage to source tables Explains why two reports disagree Attributes a KPI movement by segment Flags reports that drifted from definition

What it is not

Not a replacement for your BI platform Not the owner of a metric definition Not a change to any dashboard
The Trust Problem

Two dashboards, two numbers, forty minutes of meeting

A metric layer only works if everyone agrees what the metric is, and in practice the same word means slightly different things on three reports. The differences are usually mundane — a date basis, an excluded segment, a refresh that ran at a different hour — but nobody has the lineage in front of them, so the meeting becomes a debate about arithmetic.

Definitions live in people’s heads

The documented definition and the one implemented in the dashboard diverged during a change nobody recorded.

Filter context is invisible

A tile inherits a page-level filter that excludes a region, and the number is read as though it were company-wide.

Refresh timing is mistaken for movement

One report refreshed at six and another at noon, so a difference in freshness is discussed as a change in performance.

Movement is noticed, not explained

A KPI moved four points and the dashboard shows that it moved, which is the part everyone already knew.

The VDF AI Opportunity

The metric layer, made answerable

Definitions

What This Metric Actually Measures

From the governed definition.

The agent answers from the definition your semantic layer holds — the population, the date basis, the exclusions and the grain — and where the implementation in a specific report departs from it, that divergence is reported rather than smoothed over.

  • Answers from the governed definition
  • Implementation divergence reported
  • Date basis and exclusions stated
  • Owner of the definition identified
Official
Definition Source

Semantic layer

PopulationDate basisExclusionsOwner

Reconciliation

Why These Two Numbers Differ

Traced through the lineage.

When two reports disagree, the agent walks both lineages back to their source tables and identifies the specific divergence — a filter, a join, a date basis, a refresh time — instead of leaving a meeting to adjudicate between them.

Traced
Difference Cause

Back to source

Filter contextJoin pathDate basisRefresh time

Attribution

What Moved It, Not That It Moved

Decomposed by contribution.

A change in a KPI is broken into the segments, products and periods that account for it, with the contribution of each stated, so the discussion starts from where the movement came from rather than from whether it is real.

Decomposed
KPI Movement

By contribution

SegmentProductPeriodMix effect
Run sequence

How the AI BI Agent runs a task

  1. STEP 01

    Anchor on the governed definition

    The question is resolved against the definition held in your semantic layer first, including its owner and last change, so the answer starts from what the organisation has agreed rather than from what the tile appears to show.

    Definition lookupOwnership record
  2. STEP 02

    Walk the lineage

    From the tile in question the agent traces the model, the joins and the source tables that feed it, together with the filter context applied at page and report level, which is where most apparent contradictions actually originate.

    Lineage traversalFilter context
  3. STEP 03

    Compare against freshness

    Refresh schedules and last-load timestamps are checked before any difference is treated as substantive, because a comparison between a report loaded overnight and one loaded at midday is a timing artefact rather than a finding.

    Refresh checkLoad timestamps
  4. STEP 04

    Decompose the movement

    Where the question is about a change, the delta is split across segments, products and periods with each contribution quantified, separating genuine shifts in performance from mix effects that move the aggregate without moving anything underneath.

    Contribution analysisMix separation
  5. STEP 05

    Report divergence honestly

    If the implemented logic no longer matches the governed definition, the answer states both and identifies which reports are affected, leaving the correction to the metric owner rather than quietly preferring one.

    Drift detectionOwner routing
Integrations

Systems the AI BI Agent connects to

Scoped, per-tenant credentials Every call written to the audit log No data copied to a third party
Specification

Inputs, outputs and runtime

Ingests
Metric or KPI questionGoverned metric definitionsDashboard lineageRefresh schedulesHistorical values
Produces
Definition answer with ownerReconciliation of differing reportsMovement attribution by segmentDefinition drift report
Triggered by
Stakeholder questionScorecard reviewReported discrepancy
Human oversight
Metric owners approve any definition change
Models
Open-weight LLMs you host — Llama, Qwen or Mistral class
Typical latency
Seconds for a definition, minutes to reconcile
Deployment
On-premise or sovereign cloud with egress control
Data residency
Metric values stay in your BI environment
Where it pays back

Where the BI Agent pays back

Metric Definition Questions

Answer what a KPI includes and excludes from the governed definition, with its owner named.

Dashboard Reconciliation

Explain why two reports show different values for what is supposedly the same measure.

KPI Movement Attribution

Break a change in a headline metric into the segments and periods that actually produced it.

Pre-Meeting Briefing

Summarise what changed on a scorecard since the last review and what the likely drivers were.

Dashboard Onboarding

Explain to a new joiner what each tile measures, at what grain, and when it last refreshed.

Definition Drift Detection

Flag reports whose implemented logic no longer matches the governed definition they claim to use.

Comparison

AI BI Agent vs chatbots and SaaS copilots

Every organisation with more than one reporting tool eventually discovers that its hardest analytics problem is not computing a number but establishing which of two existing numbers is the one people should use.

  Generic chatbot SaaS copilot VDF AI
Definition source Industry convention Report label Your governed definition
Disagreeing reports Cannot see them Shows both Names the specific cause
Filter context Unknown Often missed Read from the report
Refresh timing Ignored Ignored Checked before comparing
Movement Restates the change Shows the change Attributed by segment
Definition changes Invents one Not applicable Routed to the metric owner
Where values are read Pasted Vendor tenancy Your own BI environment
Controls

Governance and controls

The authority of a metric comes from a single owned definition, so anything that answers questions about metrics has to defer to that ownership rather than becoming a second, more convenient source of truth.

Internal data governanceGDPRISO 27001SOC 2

Definitions not overwritten

The agent cannot change a metric

Owner named in answers

Every definition cites who owns it

Drift reported, not resolved

Divergence goes to the metric owner

Row-level security honoured

Existing BI restrictions still apply

Read-only dashboard access

No report or tile is modified

Freshness disclosed

Answers state when data last loaded

Evidence it leaves behind

Definition citation record Lineage trace output Reconciliation findings log Drift notification trail
ROI snapshot

What changes after rollout

Shorter Meeting time spent debating which number
Traceable Every figure tied back to its source
Explained Movements attributed to contributing segments
Surfaced Reports whose logic has drifted from the definition
Audience

Who runs the AI BI Agent

Head of business intelligence

Gets a standing report of where implemented logic has drifted from the governed definition, which converts a suspicion that the estate has decayed into a specific list of reports to correct.

Commercial leader in a review

Arrives at the monthly review knowing which segments account for the movement on each scorecard line, so the conversation starts at the cause rather than at whether the chart is right.

Finance business partner

Can settle a disagreement between an operational dashboard and the management pack by pointing at the specific date basis that differs, instead of rebuilding both numbers by hand before the meeting.

FAQ

Questions about the AI BI Agent

What is an AI BI agent?

It is an agent that makes an existing BI estate answerable: explaining what a metric measures from the governed definition, reconciling reports that disagree by tracing their lineage, and attributing a KPI movement to the segments that produced it.

How is an AI BI agent different from a generic chatbot?

A chatbot describes what a metric like churn usually means. This agent reads your semantic layer and your dashboard lineage, so the answer is what your organisation has agreed the metric is.

Can an AI BI agent run on-premise on BI and metric layer data?

Yes. Dashboards carry commercial performance at segment level and the metric layer describes how the business measures itself, so both stay inside the environment they already live in.

What does an AI BI agent produce, and in what format?

A definition answer with its owner and divergences, a reconciliation naming the specific cause of a difference, and a movement attribution by segment with each contribution quantified.

Where does an AI BI agent fit in a governed AI programme?

It interprets the governed layer rather than replacing it. Open exploration belongs to the data analyst agent, query construction to the SQL agent, and definition changes remain with the metric owner.

Does it replace Power BI, Tableau or Looker?

No. Those platforms remain where dashboards are built, governed and served. This agent reads what they hold — definitions, lineage, filter context, refresh state — and answers questions about it in language. The value is in interpretation and reconciliation, which is precisely the work a dashboard cannot do for itself because it can only show you its own answer.

What if we have no semantic layer or governed definitions?

It still traces lineage and reconciles reports, but it will say that the definitions are implied by the implementations rather than governed. In that situation the drift report becomes the most useful output, because it enumerates the places where the same metric name has been implemented differently, which is usually the evidence needed to justify building the governed layer in the first place.

Can it change a metric definition when it finds an error?

No. It reports the divergence, identifies the affected reports and notifies the owner. A metric definition is a business agreement with consequences for targets, incentives and external reporting, and an agent that could adjust one to resolve an inconsistency would be making that decision on everyone’s behalf without anyone noticing.

How does it attribute a KPI movement?

By decomposing the delta across the dimensions the metric is sliced by and quantifying each contribution, separating volume, rate and mix effects where the data supports it. Where the movement is concentrated in a small population, that is reported with the population size, because a large percentage shift in a thin segment is a different finding from the same shift across the base.

How does this differ from the reporting agent?

This agent answers questions about metrics as they are asked; the reporting agent produces recurring packs on a schedule. They overlap at the point where a monthly pack needs commentary, and in practice the reporting agent uses this one for the explanation of each movement while owning the assembly, the consistency across periods and the distribution.

Settle the number before the meeting

See the AI BI Agent reconcile two dashboards and attribute a KPI movement.