Revenue Delta Attribution Agent: Decompose the Weekly Number Across 5 Systems

Monday morning. Someone opens the dashboard. Revenue is up 8% week-over-week. Or down 12%. The team starts hunting for the reason. Marketing claims the uptick. Engineering blames the deploy. The CEO reads a macro headline and nods.

Revenue movements don't have a single cause. They're the intersection of experiments running, features shipping, incidents degrading the product, ad spend drifting, and external forces nobody owns. Forcing the delta onto one factor isn't analysis. It's theater. And it actively distorts the next decision.

The fix is structural. An agent fires every Monday before anyone walks into the standup. It pulls the delta, queries five evidence systems in parallel, scores each candidate cause, and ships a ranked hypothesis list with explicit confidence levels and an unexplained remainder. The agent doesn't produce the answer. It produces a structured decomposition with uncertainty surfaced — which is the only honest output.

Whoever frames the narrative first owns it. Recency bias drags credit toward the most recent change — the launch that happened Thursday, not the experiment that ramped six weeks ago and hit its traffic target quietly on Saturday. Confirmation bias routes the explanation toward whatever the room already wants true. Marketing sees marketing. Product sees product. Leadership sees vindication.

Forrester named the second failure: false precision in analytics.^[2] End users never see the formulas behind the dashboard, so they make decisions with confidence the data doesn't support. "The new landing page drove the 8%" carries an implied certainty that was never measured. Nobody argues with it because the number sounds real.

The third failure is timing. By the time the meeting resolves a story, the correction window is closed. Budgets got allocated Tuesday. The experiment got shut down Thursday because it "already worked." The agent has no ego, no recency bias, no calendar. It runs the same evidence pass every week — including the Monday after a long weekend when the analyst is heads-down on something else.

Read the third column. That's the entire structural argument.

The A/B platform can tell you, with statistical rigor, that variant B lifted conversion by 3.1%. The news feed cannot tell you whether a competitor's outage drove signups. Treat those signals as equivalent and the agent is producing decoration, not analysis.

The agent should weight evidence by signal strength and say so explicitly. When marketing asks whether the new campaign drove the increase, the honest output is: spend rose 15%, timing aligns, no incrementality test was running, plausibility moderate, confidence low. That sentence is more useful than any point estimate the room would otherwise improvise.

This is also where incrementality tests earn their place. A holdout group study — where a representative cohort is deliberately excluded from a campaign — directly measures the lift the attribution agent can only estimate.^[9] If you run no incrementality tests, your marketing rows will always carry 'medium' confidence at best, regardless of how much spend data the agent collects.

The scoring function above is fast, transparent, and good for weekly cadence. It starts failing when candidate causes interact — when the pricing test and the ad spend increase both hit the same user cohort and you need to decompose which one drove what.

That's where DoWhy's Graphical Causal Model (GCM) module earns its place.^[1] You define a directed acyclic graph (DAG) of the causal relationships you believe hold — revenue depends on conversion rate; conversion rate depends on page experience and offer quality; page experience depends on uptime — then fit the model on historical data and call gcm.attribute_anomalies() to decompose an anomalous week. The output is a per-node attribution score: positive scores are contributing factors, negative scores are factors that reduced the anomaly's likelihood.^[4]

This isn't a replacement for the weekly heuristic agent. It's a deeper investigation tool for the cases where the heuristic returns a large unexplained remainder or high overlap between candidates.

Behavioral economics research in Management Science on overconfidence in interval estimates^[7] shows the effect is driven by neglecting unknowns. Prompt people to consider what they don't know, and overconfidence drops. The agent should bake this into its output schema — not its tone.

Every hypothesis ships with three explicit fields:

• Impact range — a low-to-high revenue bound, never a point estimate
• Confidence level — high, medium, low, or speculative — each with a defined evidence threshold
• Evidence basis — what data supports the call, and what data is missing

When the sum of impact ranges doesn't cover the full delta, the agent reports an unexplained remainder. This isn't a failure mode. It's the audit trail for everything the system can't observe at a weekly cadence — word-of-mouth momentum, brand-perception drift, slow product-quality compounding. Suppressing the remainder makes the report look complete at the cost of being misleading.

The dataMissing field is not optional. Teams that omit it end up with reports that look authoritative but hide their own blind spots. When a cause is 'medium confidence,' the missing data field should say exactly what would have to be true to promote it to 'high' — an incrementality holdout, a user-level event log, an error rate correlated to the specific timeframe. The reader should know what to go collect next.

For durable scheduling, Temporal is a strong fit for this kind of agent — it handles retries, state persistence across API timeouts, and observable execution history in a way that cron alone doesn't.^[10] GitHub Actions works for lower-frequency cadences. Pick the one your team will actually maintain.

The week below: a B2B SaaS company posts a $47K weekly revenue increase, up 9.2% WoW. Z-score of 1.8 against the 13-week distribution. Above the noise floor. Worth investigating. Below the threshold where a single-cause explanation would be defensible.

The hypothesis ranges overlap and their sum exceeds the delta. That's correct — and expected. Attribution isn't an accounting closeout where causes must add to 100%. Causes interact. The pricing test and the competitor outage may have amplified each other in the enterprise segment. Forcing them into non-overlapping buckets produces false precision.

The unexplained remainder of $0–$16K is uncomfortable. It's also accurate. Some portion of every weekly movement is driven by forces the system can't observe at this cadence — word-of-mouth, brand drift, slow product-quality compounding. Naming that portion is what separates the report from a story.

The agent improves only if the loop closes. After a quarter, walk back through the archived reports. Where ground truth has arrived — a test reached power, a feature was measured over 90 days, a channel was paused and isolated — grade the agent's earlier estimate against what actually happened.

Two metrics carry the loop:

Coverage rate. How often did the true cause appear somewhere in the agent's hypothesis list? If a real cause was never even a candidate, the data-gathering layer has a blind spot. Add the source.

Confidence calibration. When the agent labels a hypothesis 'high confidence,' it should be right 80%+ of the time. When it labels 'low,' roughly 30–40% accuracy is appropriate — the label means 'plausible, unverified,' not 'wrong.' If high and low are equally accurate, the scoring function is broken. The labels are decoration, not signal.

The full system isn't week-one work. The first version is.

Week 1. Pull the revenue delta. Compute the z-score against 13-week history. Knowing whether this week is signal or noise is itself a meaningful output — before any cause is investigated. Post it to Slack. That's it.

Week 2–3. Add one evidence category. Experiments if the team runs an A/B platform. Incidents if PagerDuty is wired up. Get the data flowing into a structured report with a single candidate category.

Week 4–6. Add the remaining four categories one at a time. Each new source adds explanatory power and a maintenance cost. Prioritize by signal strength column in the table above.

Month 2–3. Implement the scoring function and confidence framework. This is the hard part — it will need iteration. Start with the default weights (reach: 0.30, timing: 0.25, magnitude: 0.25, corroboration: 0.20) and refine against calibration data, not intuition.

Ongoing. Run the quarterly calibration review. Adjust weights. Add data sources when blind spots show up. If a recurring blind spot won't go away, consider switching from the heuristic scorer to DoWhy GCM for that hypothesis category.

One admission from teams that have shipped this: after three or four months, the scoring model matters less than the discipline of running the process consistently. The benefit isn't that the agent is smarter than a senior analyst for any given week. It's that the agent runs every week without fail — including crunch weeks, the Monday after a long weekend, the months when the analyst is on something else. Consistency beats accuracy. The structured report changes how decisions get made over a quarter, regardless of any single week's verdict.