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.
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
What it is not
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 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
Semantic layer
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.
Back to source
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.
By contribution
How the AI BI Agent runs a task
- 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 - 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 - 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 - 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 - 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
Systems the AI BI Agent connects to
Metric and lineage
Analysis
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 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.
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 |
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.
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
What changes after rollout
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.
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.