Build a growth reporting agent

A weekly growth report looks like a natural job for an AI agent: the same metrics from the same few systems, written up in the same shape every Monday. The risk is just as plain. One wrong number in a fluent paragraph, a renamed metric compared with last month's, or a sentence claiming the new campaign "drove" a rise in sign-ups.

This tutorial applies our nine-step method for building a supervised agent to that job without re-teaching it. It covers what is specific to growth reporting: read-only permissions, metric-definition drift, keeping arithmetic away from the model and keeping causal claims out of the draft.

AI Chronicle has not built this agent in any product. The worked agent is a labelled hypothetical; nothing here is a screenshot, a measured result or a tested configuration. Where we describe an API, we report what its official documentation said on 16 October 2026. For background, see our explainer on what AI agents are.

What the agent is for, and what it is not

The job is narrow: once a week, produce a draft growth note that a named person can check quickly. It is not an analyst or a budget optimiser.

Scope box: hypothetical weekly growth reporting agent

The agent may: run approved, saved queries against analytics, ads and CRM with read-only access; compute metrics and week-over-week changes in code; compare definitions against the metric register; flag anomalies and data-quality warnings; draft commentary in a fixed template; save the draft and its evidence table to a review folder.

The agent may not: write new queries at run time; change analytics settings, key events, attribution models or conversion windows; edit campaigns, bids or budgets; create, edit or delete CRM records; send, post or circulate the report; state that a channel or campaign caused a result.

A person always: approves and circulates the report, changes a metric definition, decides what an anomaly means and acts on budgets.

The last "may not" is a language rule, not a permission; it gets its own section below.

Tools and permissions

The method's core rule applies with extra force: set permissions in the connected systems, not just in the prompt. Each source documents read-only access differently.

Google Analytics 4. The Data API's runReport method accepts either the analytics.readonly scope or the broader analytics scope.[1] Grant the read-only one, so the credentials cannot reach Admin API methods that change attribution settings or key events.

Google Ads. Reporting uses queries in the Google Ads Query Language (GAQL).[2] OAuth scopes will not make them read-only:

“Google Ads API uses the same scopes for read-only versus read-write operations.”

— Google for Developers, Understand the Google Ads access model.[3]

The API follows Google Ads user roles instead, including read-only access.[3] The account the agent uses needs a read-only role in Google Ads itself.

CRM. In HubSpot, one documented example, crm.objects.deals.read views deals and crm.objects.deals.write can create, delete or change them.[4] Grant only read scopes.

There is no send tool and no tool that writes a new query. Saved queries make traceability possible: each one's text, version and run time are known.

Why the model should not do the arithmetic

Language models turn a results table into readable prose well. They are less reliable at producing the numbers:

“LLMs often make logical and arithmetic mistakes in the solution part, even when the problem is decomposed correctly.”

— Gao and colleagues, PAL: Program-aided Language Models (2022).[5]

PAL's remedy was to have the model write a program and hand the solving to a Python interpreter.[5] Program of Thoughts did the same and reported an average gain of around 12% over chain-of-thought prompting on the math and financial datasets it tested.[6] On tables, TableBench's authors found even GPT-4 "achieves only a modest score compared to humans".[7] In FinanceBench, GPT-4-Turbo with retrieval answered 81% of a sample of financial questions incorrectly or refused.[8] These are 2023–2024 model configurations, not current products, but the design lesson holds.

NIST's generative AI profile adds that models can "provide logical steps for how they arrived at an answer even when the answer itself is incorrect".[9] A tidy explanation beside a percentage proves nothing about the percentage.

So the design splits the work. Code computes every number into a results table, each row tagged with a result ID and its query. The model writes words around those numbers. A deterministic lineage check then compares every number in the draft with the table; an unmatched or untagged figure blocks the draft. Our comparison of ChatGPT and Gemini for spreadsheet analysis looks at the same problem from the analyst's chair.

AI Chronicle

Where metric definitions drift

Many misleading growth reports have nothing to do with AI. They compare numbers that look alike but are no longer measured alike. The agent catches this with a metric register: one entry per metric giving its definition, source, query ID, API field and date of last change. The official docs give plenty of reasons to keep one.

Names change. In May 2024 the GA4 Data API told developers to replace the deprecated conversions metric with keyEvents.[10] A saved query, a spreadsheet and last quarter's slide can use different names for one concept.

Changes apply from different dates. GA4's schema notes that marking an event as a key event:

“affects reports from time of creation. It doesn't change historic data.”

— Google for Developers, GA4 Data API dimensions and metrics.[11]

A newly marked key event produces a step up that is not growth. Attribution works the other way: GA4 says "changing the reporting attribution model applies to historical and future data".[12] Google Ads conversion-window changes apply "going forward".[13] Log each change and snapshot what was reported, so a restatement is flagged rather than read as movement.

Similar names, different scopes. GA4's activeUsers counts distinct users who visited; totalUsers counts distinct users who logged any event. purchaseRevenue is purchases minus refunds; totalRevenue adds subscription and ad revenue.[11] Google Ads' metrics.conversions covers only actions included in the Conversions metric, metrics.all_conversions covers all, and a conversions_by_conversion_date variant dates by conversion date.[14] Two "conversions" figures for one week can legitimately differ.

The data is still moving. GA4 says processing "can take 24-48 hours", and attribution credit for key events can change for up to 12 days.[15] A Monday-morning run should mark recent days provisional.

The data may be incomplete. GA4 Data API responses carry flags: subjectToThresholding when data may be withheld, dataLossFromOtherRow when rows were rolled into "(other)", and sampling metadata.[16] runReport returns 10,000 rows by default and at most 250,000.[1] The agent puts a warning beside any affected figure.

Attribution is credit, not cause

GA4 describes attribution as "the act of assigning credit" to ads, clicks and other touchpoints; under a data-driven model the credit can be fractional.[12] Google's advertising docs separate this from causation. Its Conversion Lift page contrasts standard attributed conversions with incremental conversions from a controlled experiment, which it says measure "the true causal impact of your advertising".[17]

So the rule: attributed figures are reported as attributed, under a named model and window. "Drove", "caused" and "because of" are allowed only when the source is a controlled experiment, and the reviewer still decides.

The model will not apply this judgement unprompted. In the Corr2Cause benchmark, seventeen language models inferring causation from correlational statements achieved "almost close to random performance".[18] Spend and sign-ups rising in the same week is exactly what a model will narrate as cause and effect. Our guide to AI for marketing covers the checks marketing teams should apply to AI output more broadly.

The worked example: a weekly growth note

This is a hypothetical design example. It describes a plan, not a system AI Chronicle has built or run. No outputs, accuracy figures or time savings are claimed.

Imagine a small subscription software company whose growth lead spends Monday mornings pulling GA4, Google Ads and CRM figures into a note for the founders. The agent's job is to have a draft and its evidence ready when she sits down.

Inputs. Five saved, versioned queries: sessions by channel group (GA4-01); trial sign-ups as the sign_up key event (GA4-02); Ads cost (ADS-01); Ads conversions, with the conversion-date variant (ADS-02); CRM deals created and closed-won (CRM-01). The range is the last full week; each response's timeZone and currencyCode are recorded.[16]

Reconcile. Code checks each query against the register. If GA4 sign-ups and CRM trials disagree, both appear with their definitions. The agent neither picks one nor averages them.

Flag. The team writes the anomaly rules: a weekly change beyond a set threshold, a quality flag, a logged definition change, a failed or empty query. Each flag states what was observed, not why.

Draft and check. The model fills a fixed template: headline figures, changes, warnings, open questions. Every number carries its tag, such as "Trial sign-ups: 412 [R2 ← GA4-02 v3]" (an invented figure). Percentages come from code, which distinguishes percentage points from per cent. A failed lineage check or a banned causal verb returns the draft once, then goes to the reviewer marked. The growth lead edits and circulates it herself.

Her review time is the real measure of whether the agent helps; see how to measure AI productivity without hiding review work.

Test cases for this domain

The method's test step applies unchanged. These rows are for the hypothetical agent; the Result column is empty on purpose.

Failure modes specific to growth reporting

• Silent definition drift: renamed metrics, new key events, changed attribution. Defence: register, change log, snapshots.
• Mangled numbers: rounding, recalculation, transposition. Defence: code computes; the lineage check blocks.
• Hidden incompleteness: thresholding, "(other)" rows, sampling, row limits. Defence: read response metadata every run.
• Provisional data shown as final. Defence: a freshness rule and label.
• Double counting: one sign-up as GA4 key event, Ads conversion and CRM trial. Defence: add across sources only where the register allows.
• Causal over-claiming. Defence: the causal-language rule.
• Alert fatigue: thresholds so tight that reviewers stop reading. Defence: tune the rules during staged rollout.

When not to build this agent

Not if your team has not agreed what its core metrics mean; the agent would automate the disagreement. Write the register first, which is useful without AI. Not if a dashboard already answers the question. Not for board packs, investor updates or regulated filings without finance controls. And not if the data needs write access or nobody has time to review: an unreviewed note with confident commentary is worse than none.

Questions readers commonly ask

Can the agent write its own queries?

Not in this design. Saved queries keep numbers traceable and data access bounded. A person writes and reviews any new query, approves it and updates the register.

Should it explain why a metric moved?

It can list what changed in the same period and pose open questions. Explaining why is the reviewer's job; attributed conversions are not causal evidence.[17]

Is an anomaly flag a problem?

Not necessarily. It means a figure crossed a rule the team wrote: a real change, a tracking fault, a definition change or noise. A person decides which.

Can it send the report once it seems reliable?

Treat that as a separate decision with its own evidence. For a report that informs budgets, a person between draft and audience is cheap insurance.

The next step

Growth reporting checks numbers against queries. The next tutorial checks invoices against purchase orders and receipts: build an invoice-exception investigation agent that explains mismatches and leaves every payment decision to a person.

About these sources

All pages were accessed on 16 October 2026. Official documentation establishes what a product or API is described as doing; it does not show how well an agent built on it would work. AI Chronicle has not built, run or tested the example agent. The scope box, tools table, test cases and example figure are illustrative.

Related reading

• Build a research agent with source verification
• Build a SaaS onboarding investigation agent