A personal-care brand runs Amazon and Flipkart ads with AI agents that know its numbers and its limits.
Three platforms, three definitions of the same metric, three restatement windows. Reports were pasted into spreadsheets and questions were answered a day late. Now agents run pacing, competitor prices and the weekly search-term review on one catalog, tied to listings and stock. They recommend changes within caps the brand set, a media buyer approves, and every change is recorded with its sources.
The situation
- Amazon reports ACOS on its own attribution window. Flipkart reports RoAS and ROI and restates recent days. Meta reports purchase ROAS on a third window.
- “Which search terms convert?” needed exports from each platform, a pivot, and someone to remember which column meant what.
- An agent pointed at the raw APIs burned through its context on rows and tool definitions, and still mixed the definitions.
- Nobody could say what the agent was allowed to change, by how much, or how to undo it.
Before the agent could act, it had to know
The agent could reason and use tools from day one. It became reliable once this list was written down: which information to trust, what the terms mean, which rules apply, and what it is allowed to do.
- Spend against budget per campaign, every hour
- ROAS by the brand's definition, not the platform's
- Which campaigns are protected and left alone
- Search-term rows from both marketplaces, mapped to one definition
- Stock and price on the listing behind each term
- The category target and the cap per change
- The before state, so the change can be undone
- The cap per change and per week
- That a person approved this plan
- Nothing. There is no write API. The plan becomes a checklist with the exact values for a person to enter.
What they already had, and what we built on it
Where a system already gave the agent what it needed, we used it as it was. We built only what was missing.
The system, layer by layer
How the search-term review runs
- ScheduleEarly in the week, the agent reads search-term reports from both marketplaces, already mapped onto the brand’s definitions.
- ContextIt establishes what changed since the last run: new terms, terms that stopped converting, stock and price on the listings behind them.
- DecisionTerms are ranked by code, not by the model, against the category targets. Each candidate change carries its source rows and freshness.
- PlanOne plan: negatives to add, bids to move within the cap, and the campaigns left alone because they are protected.
- ReviewThe plan lands in the inbox with the evidence. The media buyer edits a line; everything downstream re-resolves. One approval covers the plan.
- ExecuteAmazon changes apply, read back and store the previous values. The Flipkart part becomes a checklist with the exact values to enter.
The pacing automation, in outline
{
"automation": "marketplace-pacing",
"schedule": "hourly",
"scope": { "platforms": ["amazon_ads", "flipkart_ads"], "exclude": "protected" },
"metrics": { "roas": "catalog:…", "pacing": "catalog:…" },
"condition": { … },
"plan": ["amazon.campaign.budget.set", "flipkart.checklist"],
"policy": { "risk": "medium", "requires_approval": true }
}Each stateful automation is a typed contract: a trigger, a condition, the context it may read, the registered actions in its plan, and its policy. The model fills fields. Our code validates, resolves and binds. A person approves what the policy says needs approval. The full contract carries more than we show here.
Who can do what
- Budget and bid changes are capped per change and per week. The caps apply even after approval.
- Launch campaigns are protected entities. Prices and listing content are never automated.
- Every number in a plan names the rows it came from and when they were synced. A value with no source does not render.
- Each agent has its own identity with read rights across the catalog and write rights only on the actions in its plan.
Bring us one workflow. We’ll start there.
We connect the systems you already run, write down what the agent must know before it acts, and start with recommendations before actions. The next workflow reuses what the first one built.