Growth Marketing Officer AI

Invoked when: monthly unit economics review, new product launch requiring profitability model, or CAC is trending above target threshold

• –Reads borrower cohort data — acquisition cost by channel and segment, loan utilisation, repayment behaviour, NPA rate, fee income, and repeat loan uptake — and builds a CAC/LTV model that tells the growth team which acquisition segments and channels are generating profitable customers, which are acquiring borrowers at a loss, and what the CAC ceiling is for each segment before the unit economics break down.
• –Computes LTV not just as loan revenue, but as the full borrower lifetime value accounting for repeat loan probability, cross-sell potential, referral value, and NPA risk adjustment — because a borrower with a high first-loan CAC who takes three more loans and never defaults has a materially higher LTV than a borrower acquired cheaply who defaults at first exposure.
• –Identifies the maximum sustainable CAC by segment — the point at which acquiring one more borrower in that segment still generates positive lifetime value after credit losses, funding costs, and ops costs are accounted for. This CAC ceiling is the constraint within which all channel optimisation and experiment design operates.
• –Cannot build actuarially calibrated LTV models from raw loan ledger data. It needs pre-aggregated cohort performance data injected at invocation — cohort vintage, channel, segment, loan size band, observed repayment behaviour, and observed NPA rate. The CRO AI's credit performance data is the essential input to any LTV calculation.

Invoked when: weekly performance data available, budget reallocation is being considered, or a channel shows a significant performance shift

• –Reads channel-wise performance data — spend, impressions, clicks, lead volume, lead-to-application conversion, application-to-sanction conversion, sanction-to-disbursement conversion, and CAC — and applies the CAC ceiling constraint from the unit economics model to produce a channel performance ranking that is bounded by profitability, not just volume or CPL.
• –Identifies three distinct performance failure modes by channel: volume saturation (a channel that was performing well is now showing declining conversion as the addressable audience is exhausted), quality degradation (CPL is stable but lead-to-funded-loan conversion is falling because audience targeting has drifted), and efficiency erosion (performance is steady but a newer channel is now delivering the same volume at lower CAC). Each requires a different response.
• –Models the marginal return on additional spend in each channel — what does the next ₹10 lakh of spend in Channel A deliver in funded loans versus Channel B — to identify where incremental budget generates the most funded loan volume within the CAC ceiling. Explicitly accounts for diminishing returns: a channel that is efficient at ₹50 lakh/month may become inefficient at ₹80 lakh/month as the best audience segments become saturated.
• –Does not access platform interfaces, adjust bids, change audience targeting, or move budget between channels in any live system. Channel optimisation outputs are recommendations for the human performance marketing team to implement through their platform tools and approval workflows.

Invoked when: new experiment is being planned, a running experiment has reached its target sample size, or the experiment roadmap needs prioritisation

• –For new experiment design: reads the hypothesis, the funnel stage being tested, the available traffic volume, and the minimum detectable effect the team cares about — and produces an experiment brief specifying the control and variant, the primary metric and guardrail metrics, the required sample size for statistical significance at the target power level, the minimum runtime to avoid novelty effects, and the stopping rules that prevent premature conclusion.
• –Maintains the experiment roadmap from the experiment log — which tests are running, which have completed, what was learned, and which hypotheses were confirmed or refuted. Flags when a proposed new experiment duplicates a prior test, conflicts with a currently running test on the same funnel stage, or tests a hypothesis that was already disproved and should not be re-tested without new evidence.
• –For completed experiments: reads the final results and applies statistical significance testing — was the observed difference real or within random variance? If significant, what is the effect size, the confidence interval, and the practical significance (a statistically significant 0.1% conversion improvement on low-traffic stage may not justify the implementation cost)? If not significant, was the test adequately powered, or did it stop too early to detect a real effect?
• –Does not launch experiments, implement winning variants into production, or access experimentation platforms directly. Experiment governance is the analytical oversight layer — every experiment launch and implementation decision is made and executed by the human growth team in their experimentation tools.

Invoked when: new acquisition funnel is being designed, a persistent conversion problem cannot be solved by optimisation within the current funnel structure, or a new channel requires a tailored funnel variant

• –Maps the end-to-end acquisition funnel — from first paid or organic touchpoint through lead capture, pre-qualification, application, credit check, sanction, and disbursement — stage by stage, specifying the conversion objective at each stage, the data collected, the drop-off risk, and the recovery mechanism (retargeting, nudge, alternative pathway) for applicants who exit at each point.
• –Designs funnel variants by segment and channel: a borrower coming from a high-intent Google search campaign has different intent, context, and patience than a borrower entering from a social media discovery ad — the funnel architecture for each should be distinct, not a single universal flow applied to all traffic sources regardless of intent signal.
• –Identifies the bottleneck stage in the current funnel — the single stage where improving conversion by 10% would have the largest impact on funded loan volume — and recommends whether the bottleneck is best addressed by a funnel architecture change (structural) or an A/B experiment on the existing structure (optimisation). Not every conversion problem requires a funnel rebuild.
• –Checks every proposed funnel stage against the onboarding regulatory compliance requirements from the DPO AI and the Onboarding Head AI — a funnel stage that collects data without the required DPDP consent, or that defers a mandatory KYC step in a way the RBI KYC norms do not permit, cannot be implemented regardless of the conversion improvement it might deliver.