Definitions Are the Foundation, Causality Is the Payload

1. The why gap

Enterprises can report what happened. Most still cannot say why it happened with numbers that a finance team will sign. Gartner's analytics maturity model names four stages: descriptive (what happened), diagnostic (why), predictive (what will happen) and prescriptive (what to do). A dashboard answers the first question. A drill-down hints at the second, but it cannot separate a real cause from a shift in mix.

Finance teams show the cost most clearly. An AFP and APQC study of more than 400 FP&A professionals, run in 2018 and 2019, found that teams spend 42% of their time gathering data and 33% administering processes. Only 25% goes to analysis that adds value. The FP&A Trends Survey 2025 found 46% of time on data collection and validation. Only 17% of organisations rate their data quality as good, and only 18% can run a scenario in under a day.

Automated root-cause analysis changes this when it works. Microsoft's Adtributor found the root cause of revenue anomalies in a large ad system with 95.3% accuracy and cut troubleshooting time by an order of magnitude (NSDI 2014). Uber reports that detecting 100 metric issues took about three months before its uVitals system and one day or less after it (Uber Engineering).

Those results come from engineering systems with a known, stable metric structure. They show the value of structured diagnosis. They do not prove that business causes can be found with the same accuracy. Business metrics rarely have that structure until someone defines and governs it.

2. Definitions drift, and drift looks like a cause

A common view holds that definitions rarely change and causality changes all the time. The second half is right. The first half often fails in practice, and when it fails, causal analysis fails with it.

The Open Semantic Interchange, launched in September 2025 by Snowflake, Salesforce, dbt Labs and others, calls inconsistent metric definitions “one of the most persistent and costly challenges in the data world” (Snowflake). Definitions do not need to change often to do damage. They only need to differ between two tools, or change once without a version record.

When that happens, an attribution engine reports the definition change as a business event:

• Active customer moves from a 30-day to a 60-day window in one system. Churn “improves” and the engine credits the latest retention campaign.
• Net revenue starts to exclude a returns category in one region. That region becomes the top “driver” of growth.
• A join path fans out when a new dimension table is added. Every segment that touches it looks larger, and each contribution is wrong by a different amount.

No causal model can detect these errors from the data alone, because the data is internally consistent. A governed semantic layer supplies the definitions, version history and lineage needed to separate a change in business meaning from a change in measured performance. Causal inference then needs a further layer of evidence. Colrows builds both layers on one model of the business.

3. Three kinds of “because”

Many driver tools present every “driver” the same way. A ranked list of segments looks equally certain whether it came from arithmetic, a regression or a guess. A causal reasoning layer must separate three kinds of answer, because each one needs a different level of proof.

Tier	Example answer	What it rests on	What the reader gets
1 · Accounting identity	“Revenue fell 4.2M. Volume explains 3.1M, price explains 0.6M, mix explains 0.5M.”	A defined formula, such as Revenue = Orders × Average order value.	An exact split of the change. It shows where the change sits in the formula, not why it happened.
2 · Statistical association	“Churn is 2.4 times higher in accounts with more than three open support tickets.”	Patterns in the data, with significance tests.	A confidence level and a sample size. Not proof of cause.
3 · Causal estimate	“The price change caused about 60% of the volume drop, with a range of 45% to 72%.”	A causal model, a comparison that isolates the effect, stated assumptions, and experiments where available.	An effect size, an uncertainty range and the assumptions it depends on.

Judea Pearl's ladder of causation gives the theory behind this split. The first rung is association (seeing), the second is intervention (doing) and the third is counterfactuals (imagining what would have happened) (Pearl and Mackenzie, The Book of Why, 2018). Dashboards and drill-downs stay on the first rung. They show conditional totals for whatever dimension the user clicks, but they do not hold other factors constant. The three tiers are an operational way to classify evidence. They do not map one to one onto Pearl's rungs.

Simpson's paradox shows the risk. A company's overall conversion rate can fall while every channel's conversion rate rises, because traffic moved toward a channel that converts less. A dashboard that shows only the top-line rate tells the wrong story. A tier 1 decomposition separates the mix effect from the rate effect and shows that no channel got worse.

Colrows labels every reason with its tier. A finance reviewer can accept tier 1 reasons as exact, read tier 2 reasons as leads to check, and weigh tier 3 reasons against their stated assumptions.

4. Metric trees give causal analysis its accounting backbone

Causal root-cause methods need a known causal structure before they can attribute anything. Amazon's work on causal root-cause analysis of outliers, which DoWhy implements, asks whether an outlier would have occurred if one variable had behaved normally. It requires a causal graph of the variables and a model of how each one is produced (Budhathoki et al., ICML 2022). A follow-up study calls the need for that full model “a severe bottleneck to its practical application” (arXiv 2406.05014).

A governed metric tree reduces that bottleneck, but it does not remove it. Three structures are at work, and they are easy to confuse:

• Metric tree. The arithmetic relations that define and decompose a metric. Its edges are exact.
• Causal graph. The hypothesised real-world relations, such as ad spend raising sessions. Its edges must be tested.
• Causal inference engine. The methods that estimate the effect of an intervention from data, a comparison and stated assumptions.

Colrows connects the three. The metric tree tells the causal engine which quantities are linked by definition, so the engine does not estimate what is already known and does not mistake an identity for a mechanism.

In the tree above, Revenue = Orders × Average order value is a definition. It tells you exactly how much of a revenue change sits in orders. It does not tell you why orders fell, which could be demand, stock, price, a competitor or several of these together. Average order value is itself revenue divided by orders, so it is a term in the identity, not an independent driver. Ad spend raising sessions, or a price change reducing conversion, is a causal hypothesis. Colrows tests those edges statistically or causally and labels each result with its tier and uncertainty.

5. Attribution that adds up

A finance team will not accept a list of reasons whose parts do not sum to the change. Some methods leave a residual. Others split the change by the order in which factors are substituted, so a different order gives a different answer. Colrows splits the accounting change so that the contributions reconcile exactly to the total, with the same answer in any order.

The example below uses round numbers. Revenue fell from $10.00M to $9.00M between two quarters. Sessions rose from 2.0M to 2.1M, the conversion rate fell from 5.00% to 4.38%, and the average order value fell from $100.00 to $97.83. The example uses the log-mean Divisia index (LMDI), a published decomposition for products of factors that leaves no residual and does not depend on order. All figures are computed from unrounded values and rounded for display.

Factor	Q1	Q2	Contribution to revenue change
Sessions	2.00M	2.10M	+$0.49M
Conversion rate	5.00%	4.38%	−$1.27M
Average order value	$100.00	$97.83	−$0.22M
Revenue	$10.00M	$9.00M	−$1.00M

Sessions grew, so the drop did not come from traffic. The conversion rate cost more than the whole decline, and the three parts reconcile exactly to the $1.00M fall. That is the tier 1 answer. It is exact, and it shows where the decline sits. It does not yet say why conversion fell.

To find out why, Colrows follows the tree down into segments such as channel, region and device, and separates mix effects from rate effects, as in the Simpson's paradox case in section 3. Segments and candidate drivers that move with the decline are tier 2 leads. They are not yet causes.

From leads to a causal estimate

Suppose a price change happened in the same quarter. A credible tier 3 estimate of its effect on conversion must answer five questions, and Colrows shows the answers with the estimate:

1. The causal question. Which price change, for which products and customers, over which period, and which outcome.
2. The comparison. What shows the outcome without the change: a controlled experiment, a holdout region, or a justified observational design such as difference-in-differences.
3. The confounders. How seasonality, marketing activity, competitor prices and stock levels were accounted for.
4. The identification assumptions. Why the comparison isolates the effect of price from other changes in the same period.
5. The robustness. Whether the result holds under other specifications and sensitivity tests. A narrow confidence interval does not protect against a wrong causal graph or an unmeasured confounder.

When the data cannot identify the effect, the honest answer is a tier 2 association with competing hypotheses, not a causal number. Colrows reports it that way.

Two ledgers, kept apart

The accounting ledger and the causal ledger answer different questions, so Colrows keeps them apart.

A $1.00M decline with a $0.60M price effect does not mean other causes explain the remaining $0.40M. The price estimate may overlap with the volume effect or depend on assumptions about customer behaviour. Keeping the ledgers apart stops a reader from adding numbers that do not add.

Ledger	What it answers	Reconciles to the total change?
Accounting attribution	Where in the metric formula the change sits.	Yes, exactly, under a stated decomposition method.
Causal evidence	Which interventions caused part of the change, and how much.	No. Each estimate links to the accounting line it explains and carries its own uncertainty. Estimates can overlap, so they are not forced to sum.

6. Where LLMs belong in causal reasoning

Large language models are good at proposing causes and poor at proving them. Research from 2023 to 2026 shows this split, and the split decides where an LLM sits in a causal stack. Kıcıman and colleagues also note that the models “ignore the actual data”. Their strength is domain knowledge, which is the input that causal analysis most often lacks.

Task	Result	Source
Say which of two variables causes the other, from domain knowledge	GPT-4 methods scored 97%, 13 points above earlier algorithms	Kıcıman et al., Microsoft, 2023
Formal causal inference across all three rungs of Pearl's ladder	GPT-4 scored 62.03%, against a random baseline of 49.27%	CLadder, NeurIPS 2023
Infer causation from stated correlations	17 LLMs scored close to random	Corr2Cause, ICLR 2024
Answer metric questions through a semantic layer instead of writing raw SQL	100% compared with 52.5% for text-to-SQL, on covered questions	dbt Labs, 2026

The design follows from the evidence. In Colrows, the LLM does four jobs: it translates the question, selects the analysis, proposes candidate causes from domain knowledge, and explains the result in plain language. The attribution engine computes every contribution from governed metrics, and causal estimates come from defined analytical methods. The LLM does not calculate any number.

This removes one source of error, not all of them. An LLM can still pick the wrong metric, the wrong analysis or a biased hypothesis, and a deterministic engine repeats a wrong method every time. Colrows adds three checks on the path from evidence to explanation:

• Grounded hypotheses. Each candidate cause the LLM proposes must map to a variable that exists in the governed model and the data.
• Checked methods. Before a causal estimate is shown, the method's assumptions and data requirements are checked.
• Labelled narrative. The final explanation keeps observed facts, accounting contributions, associations and causal estimates in separate, labelled statements.

One limit applies. The semantic layer benchmarks above measure “what” questions, not “why” questions. The step from better retrieval to better attribution is a sound inference, but the published studies do not test it directly.

7. How Colrows answers why

Colrows puts the causal engine on top of a governed semantic layer, so every reason it reports starts from a governed definition. The LLM sits at the top and talks to the user. It sends governed requests down, and it never produces a number itself.

A question such as “why did gross margin fall in Q2?” moves through the layers in order:

1. The LLM reads the question. It resolves “gross margin” and “Q2” against governed definitions, picks the analysis, and proposes candidate causes from domain knowledge.
2. The semantic layer fixes the meaning. It supplies the versioned definition, the metric tree, the valid join paths and the user's access rules.
3. The attribution engine computes. It splits the change exactly along the metric tree (tier 1), tests segments and candidate drivers statistically (tier 2), and estimates causal effects where a causal model applies (tier 3).
4. The warehouse runs compiled SQL. Colrows runs above the data platforms you already use, so the data stays where it is.
5. The answer comes back with its evidence. The LLM explains the result in plain words. Each reason keeps its contribution, tier, definition, SQL and uncertainty.

Because the semantic layer and the causal engine share one model of the business, Colrows can tell a definition change from a business change. A version number alone does not make two periods comparable. If “active customer” moved from a 30-day to a 60-day window between Q1 and Q2, Colrows flags the break, and the comparison is either restated under one definition or shows the definition effect as its own line.

8. A trust checklist for causal attribution

Any team that buys or builds a causal reasoning layer can test it with six questions. Colrows is designed to answer yes to each one.

Question to ask	Why it matters	How Colrows answers
Do the accounting contributions add up exactly to the observed change?	A gap between the parts and the total fails a finance review. The Shapley efficiency property (1953) states this requirement for fair attribution.	Accounting contributions reconcile exactly, for additive and ratio metrics. Causal estimates link to them and are validated separately.
Can each reason be traced to a governed definition?	A reason with no definition behind it cannot be audited.	Each reason links to the metric definition, its version, the generated SQL and the source tables.
Is each reason labelled as identity, association or causal estimate?	Readers must know which reasons are exact and which depend on assumptions.	Every reason carries its tier label.
Does each estimate show its uncertainty?	A point estimate with no range implies false precision.	Statistical and causal results show confidence ranges and the assumptions they depend on.
Can the result be reproduced later?	Audit and model-risk reviews need the same answer from the same inputs.	Point-in-time reproducibility on definitions and data.
Can a user ask “what if”?	Counterfactuals are the third rung of causal reasoning and the step toward decisions.	What-if questions run on the metric tree and on causal models.

For regulated teams, these checks match existing governance. In April 2026 the Federal Reserve, OCC and FDIC issued SR 26-2, revised guidance on model risk management that replaces SR 11-7. The Federal Reserve expects it to be most relevant to banking organisations with more than $30 billion in total assets (Federal Reserve, SR 26-2). Under the EU AI Act, ordinary business metric analysis is not necessarily high-risk. Classification depends on the system's intended purpose and deployment context, and the Act lists uses such as credit scoring of individuals as high-risk (Regulation (EU) 2024/1689). Transparency and human oversight are good practice in either case.

9. How the approaches compare

Driver analysis tools fall into three groups. BI platforms add statistical driver features to dashboards. Warehouse platforms add contribution analysis inside one engine. Causal AI platforms give data scientists a workbench for causal models. Each group does useful work. The table below uses each vendor's public documentation as of October 2026. “Not documented” means we did not find the capability in the sources reviewed, not that the product lacks it.

Approach	Examples	Method, per public documentation	Governed definitions	Contributions reconcile to the total	Tier labels on each reason
Driver features in BI	Power BI Key Influencers, Tableau Pulse, ThoughtSpot SpotIQ	Regression, statistical insight types and change analysis	Varies. Tableau Pulse uses defined metrics.	Not documented, or only for additive measures	Not documented
Contribution analysis in the warehouse	BigQuery ML contribution analysis	Ranks segments by unexpected difference against expectation	Within one engine	Segment level. Exact reconciliation not documented.	Not documented
Causal AI workbenches	causaLens	Causal discovery, causal graphs and counterfactuals	Not a governed semantic layer	Not documented	Causal focus
Semantic layer with causal attribution	Colrows	Metric trees, exact decomposition, statistical and causal estimation	Yes, versioned, with lineage	Yes, exactly	Yes

Colrows does not replace the BI tool or the data science workbench. It adds a governed definition and attribution layer that those tools, and AI agents, can query. A team can keep its dashboards and still get reasons that reconcile and trace to source.

10. What comes next: from why to what if to what to do

Gartner's Hype Cycle for Artificial Intelligence, 2025 places causal AI among early-stage techniques (Pasqal summary of the Gartner report). In January 2026 Gartner published its first Magic Quadrant for Decision Intelligence Platforms, which evaluated 17 vendors on decision modelling, AI-driven automation and governance at scale (SAS). Both reports describe the same shift: analytics is moving from describing the past toward recommending actions.

Each step depends on the one before it:

1. Why. Attribute a change to its reasons, exactly and with a trace to source. Without this, any recommendation rests on a guess.
2. What if. Run a counterfactual on the metric tree or a causal model: “What would margin be if the price change had not happened?”
3. What to do. Rank actions by their estimated effect and let a person or an AI agent act within governed limits.

Colrows research focuses on making each step reliable before the next one is automated. Our view is that AI agents will make more business decisions over the next few years. Those agents will need governed definitions and reasons that add up before anyone can trust their actions.

To see Colrows trace a change in one of your own metrics, from its governed definition to a reconciled decomposition and a tested causal estimate, write to engage@colrows.com or book a session at colrows.com.

Definitions are the foundation. Causality is the payload.

References

1. AFP and APQC. Five practices of highly effective FP&A teams.
2. FP&A Trends. 2025 FP&A benchmarks and trends.
3. Bhagwan, R., et al. Adtributor: Revenue debugging in advertising systems. NSDI 2014.
4. Uber Engineering. uVitals: an anomaly detection and alerting system. December 2023.
5. Snowflake. Open Semantic Interchange. September 2025.
6. Budhathoki, K., et al. Causal structure-based root cause analysis of outliers. ICML 2022.
7. Root cause analysis follow-up on SCM dependence. arXiv 2406.05014.
8. Kıcıman, E., et al. Causal reasoning and large language models. 2023.
9. Jin, Z., et al. CLadder. NeurIPS 2023.
10. Jin, Z., et al. Corr2Cause. ICLR 2024.
11. dbt Labs. Semantic layer vs text-to-SQL. 2026.
12. Google Cloud. BigQuery ML contribution analysis is GA.
13. Tableau. Pulse insight types.
14. ThoughtSpot. SpotIQ change analysis.
15. Microsoft. Power BI key influencers and Power BI decomposition tree.
16. causaLens. Causal AI product.
17. Federal Reserve, OCC and FDIC. SR 26-2: Revised guidance on model risk management. April 2026.
18. Regulation (EU) 2024/1689 (AI Act).
19. Pasqal. Gartner Hype Cycle for AI 2025 summary.
20. SAS. Gartner Magic Quadrant for Decision Intelligence Platforms. 2026.
21. Pearl, J., and Mackenzie, D. The Book of Why. Basic Books, 2018.