Marketing teams are buying agents that can read six ad platforms, write a Monday report, and propose the next budget move before coffee. That is useful. It is not the same as letting the agent spend.
LangChain open-sourced the paid-media agent it runs on its own accounts. In six months, paid media went from zero to about 20% of marketing pipeline. Cost per qualified lead fell about 30% while monthly spend rose about 60%. The operating detail that matters for every CMO is simpler than the stack: the agent can analyze and recommend, but a human still has to approve before any campaign change runs, and code verifies the change after it lands.
That is the pattern to steal. Not the vendor. The gate.
Throughput is not the scarce input
Production scarcity is mostly gone. A smaller team can draft more ads, more variants, and more reports than last year's larger team. Domo CMO Mark Boothe puts the trap plainly: unlimited output pointed at strategy that still moves at human speed. Volume without a decision rule is a factory without a product owner.
Agents amplify that trap when write access is open. A recommendation that quietly edits keywords, geo, or creative becomes a budget decision that nobody owns until the weekly review finds it.
Separate read from write
Treat the agent like a strong analyst with a locked wallet.
- Read and report on a schedule. Pull spend, clicks, and warehouse outcomes. Publish what changed and why.
- Propose a typed change. Name the campaign, the field, the current value, the proposed value, the reason, and how you will know if it worked.
- Require a named human approval. Not a chat reply. An explicit approve action from someone authorized to move spend.
- Verify after apply. Confirm the platform state matches the approved plan. A "we sent the call" log is not a receipt.
LangChain's separation is the useful one: anyone can ask about performance; only designated people can approve mutations. Analysis stays continuous. Spend stays accountable.
Three questions for Monday's growth review
- Which agents can change live campaigns, budgets, audiences, or creative without a human approve step?
- Who is on the approve list, and is that list smaller than the ask list?
- For the last ten agent-proposed changes, which ones have a verified receipt in the ad platform?
If you cannot answer those, you do not have an agent program. You have unsupervised spend with a chat UI.
This week
Pick one paid surface the team already trusts. Turn off silent writes. Keep the weekly analysis. Require approval for every mutation. Put verified receipts next to any slide that celebrates agent throughput.
The agent can draft the plan. A human still has to own the dollar.
