Why marketing agents need business context before they touch your ad accounts

Picture a campaign for a skincare brand’s hero serum. Overnight, its ROAS dropped from 4.1× to 2.8×. An agent connected to the ad platform sees the drop and has three obvious options: cut the budget, pause the campaign or leave it alone and wait.

Each of those is the right answer for some business. If the serum has eleven units left and the restock lands Thursday, you should stop spending on demand you cannot fulfil. If a competitor dropped their price to ₹1,299 at 1:20 PM yesterday and your conversion rate fell 24% in the hours after, the fix sits in pricing, and cutting spend would only lose share. If the team raised the budget from ₹50,000 to ₹65,000 at 9 AM, the dip may simply be the platform re-learning, and the right move is to wait two days.

The ad platform shows the same 2.8× in all three cases. The number is identical; the decision depends entirely on context the platform never sees.

A number is only half a decision

Experienced media buyers carry this context in their heads. Before touching a budget, they glance at stock, remember the target the founder set last month, check whether a sale is running and scroll the client channel for anything new. They do it so quickly that it barely feels like work, which is exactly why it is easy to forget when designing an agent.

The underlying discipline is often called context engineering: deciding which information reaches an agent, in what shape, how fresh and under which access rules, at the moment it decides. A better prompt cannot fix a stale stock number or a missing instruction.

What goes wrong when agents only see the ad account

The failure modes are predictable, and each one maps to a missing piece of context.

• Scaling into a stock-out. The campaign looks efficient, so the agent raises the budget. The product sells out in two days and the extra spend buys clicks on an unavailable listing.
• Pausing during a planned push. Launch-week numbers look weak against steady-state targets. The agent pauses a campaign the team deliberately funded to build awareness.
• Undoing a human decision. The client asked to hold spend on one service line until next month. The instruction is in a chat thread; the agent never saw it and “fixes” the low spend.
• Chasing the wrong cause. Conversion fell because the landing page broke, but the agent only sees the ad account, so it rotates creatives and adjusts bids.
• Contradicting other agents. A reporting agent and a bidding agent compute ROAS on different windows, and the team ends up arguing about which number is real.

None of these require a weak model. Each comes from asking a capable model to decide with part of the picture.

The six kinds of context a marketing agent needs

• Goals and targets. The target ROAS, CPA or cost per lead for each campaign or product line, and who set it.
• Product and inventory. Which products each campaign promotes, current stock, days of cover and restock dates, across your store and marketplaces.
• Market signals. Competitor prices and promotions, seasonality and sale calendars.
• Money. Budgets, pacing against the month, and any spend caps agreed with a client or the founder.
• Team rules and past decisions. Playbooks (“don’t scale under a week of cover”), instructions from Slack or email, and what happened last time a similar problem appeared.
• Permissions. Who may see what, which changes an agent may make on its own and which need approval, and from whom.

The fifth point is easy to underrate. Meetings, messages and decisions are context in their own right. For growth teams, that means the client channel, the founder’s notes and the reasons behind last month’s budget changes.

How a context-aware decision flows

Why this belongs in a shared layer

The quick fix is to paste context into the agent’s prompt: targets, a stock export, the latest client note. It works for a demo and fails within a week. Stock changes hourly. Targets change monthly. Each new agent needs its own copy, and each copy drifts.

A shared context layer keeps one current version that every agent reads. It also makes the context cheaper to use. An agent that reads typed facts from a prepared layer needs a fraction of the tokens it would spend calling raw connectors for the same question, because the layer returns the fact instead of the raw material. A raw insights export for one ad account can run to thousands of rows, while the decision needs a handful of facts.

There is a consistency benefit too. When the reporting agent, the bidding agent and the morning brief all read ROAS from the same definition, the team stops debating which number is real and starts discussing what to do about it. A new agent added next month inherits the same targets, rules and history on its first run, instead of being taught them again through a fresh prompt.

A shared layer is also where permissions live. AI activity should run under the same security policies as the people using it, and by default the actions an agent proposes should be staged for human review. That is the right default for anything that spends money or reaches a customer.

Where to start

• Pick one decision you make every week. Budget increases are a good first choice: frequent, costly when wrong and easy to check.
• Write down everything a good operator checks before making it. That list is your context requirement.
• Connect the sources that hold those facts, and define the metrics once.
• Let the agent recommend first. Require approval for every change, and keep a record of what it saw.
• After two weeks, review the run history. Where its recommendations were consistently right, consider letting that specific action run on its own.

Meerkats gives marketing agents this context from one place: ad platforms, marketplaces, your store and team knowledge, kept in sync, with approvals built into the workflows. Whether you use it or build your own, don’t let an agent touch a budget until it can see what your best operator would check first.

Questions people ask