Rebuilding Retail Campaign Operations with an AI Agent
Multi-channel retail brand (Southeast Asia)
An AI marketing operating layer across the campaign lifecycle — reconciled reporting, stock-aware spend, and coordinated briefing — with a human approval model that loosens as trust is earned. A scoped, benchmarked proof-of-concept.
At a glance
Industry
Multi-channel retail.
Region
Southeast Asia.
Channels
Own e-commerce store, Shopee, Lazada, physical retail, and paid media on Meta & TikTok.
Engagement
AI Marketing Agent — a scoped proof-of-concept, run alongside the client's existing agency.
The problem
The team wasn't struggling to launch campaigns — they were struggling to understand and steer them. Reporting came in each platform's own shape with no view reconciling performance against actual sales, so two agencies could report materially different numbers for the same period. Catalogue ads pulled straight from the web store and kept spending against out-of-stock items, with no way to flag or pause them. As Meta, TikTok and Google absorbed campaign setup into automated products, the agency's manual role shrank while dependency and cost remained. And promotions and launches were decided verbally in planning meetings, with nothing capturing that context for the next campaign to reuse. The operating layer around campaigns — reporting, stock awareness, briefing, memory — was manual, fragmented, and disconnected from the numbers that mattered.
The solution
Rather than build 'AI for marketing' in the abstract, we scoped a configurable operating layer around the client's real workflow — a reusable core tuned to their brand, catalogue, channels, and data. It starts with reporting that compares platforms like-for-like and reconciles campaign performance against real sales; connects stock signals so the agent can flag and pause spend against out-of-stock products; and carries context across the whole lifecycle from signal to decision to brief to creative to distribution to reporting. Because the client wanted proof, not promises, it runs as a proof-of-concept alongside the existing agency, benchmarked on ROAS, CTR and CPP — and its approval model is built to loosen as the agent earns trust, never removing human judgement from decisions that carry real risk.
The outcome
Because this is a benchmarked, in-progress proof-of-concept, success is defined up front rather than claimed after the fact. The engagement is measured against reporting the team can actually use — true like-for-like comparison across platforms, reconciled to real sales — less spend wasted on out-of-stock products, competitive ROAS, CTR and CPP in a live head-to-head against the incumbent agency, faster and more consistent turnaround, and reduced agency dependency over time. Measured results will be published as the proof-of-concept concludes.
The approach
Rather than build 'AI for marketing' in the abstract, we designed a configurable operating layer around the client's real workflow — a reusable core, tuned to their brand, catalogue, channels, and data.
Reporting, reconciled. The first and highest-value problem to solve: a single reporting view that compares platforms on a like-for-like basis and reconciles campaign performance against real company sales data — something siloed agency reporting structurally cannot do. This is the clarity the team had been missing.
Stock-aware campaigns. By connecting merchandising and stock signals into the campaign layer, the agent flags — and can pause — spend against out-of-stock products before budget leaks away. Marketing gains visibility into stock realities that previously lived only with the product team.
A coordinated campaign lifecycle. The agent carries context across the whole loop — signals to decision to brief to creative (agency- or AI-assisted) to distribution to optimization — so each campaign is briefed consistently and each result informs the next.
A measurable head-to-head. Because the client wanted proof, not promises, the proof-of-concept was designed to run alongside the existing agency, benchmarked on the metrics that matter to them — ROAS, CTR, CPP — so value is demonstrated against a real control, not asserted.
The approval workflow: trust, then autonomy
Most 'human-in-the-loop' AI treats approval as a permanent gate. That didn't match how this team actually works — they onboard a new agency by reviewing everything closely for the first few months, confirm it understands the brand and tone, then deliberately speed the process up. We built the approval workflow to mirror exactly that.
Phase one — full review. Every recommendation, brief, and campaign change is surfaced for human approval, with the reasoning and the triggering signal shown transparently. The team stays fully in control while confidence builds.
Phase two — earned autonomy. As the agent proves it holds the brand's tone and judgement, approval requirements are dialed back by category — lower-risk actions flow through automatically, while material budget, launch, and brand decisions stay with a human.
The result is an AI that behaves like a well-trained team member, not a black box: it earns latitude the same way a trusted partner does, and it never removes human judgement from the decisions that carry real risk.
What we built
Reporting the team can actually use
True like-for-like comparison across platforms, and performance reconciled against real sales data.
Less wasted spend
Fewer ringgit spent against out-of-stock products.
Competitive performance in a live head-to-head
ROAS, CTR and CPP measured against the incumbent agency on real campaigns.
Faster, more consistent turnaround
Especially in prioritising stock that needs to move.
Reduced dependency
A clearer, more self-sufficient campaign operation over time.
Interested in a scoped proof-of-concept for your team?
We start with a focused, benchmarked proof-of-concept on your real campaigns, prove the value, then scale what works. Get in touch with Catalyst AI.
Get in touch