AI in the ad stack

The last module is the least settled. Two shifts are underway at once: AI surfaces becoming inventory, and AI agents gaining the ability to change accounts.

Agents with write access

During the week of 24–28 August 2026, AI agents gained write access to live ad campaigns. That is the line worth marking. Read-only agents were a reporting convenience; an agent that can change budgets, bids and targets is a new actor in the account.

Two practical consequences:

• Change history becomes the audit trail that matters. "Who changed this" now includes non-humans, and diagnosing a performance drop starts with reading it.
• Scope should be earned. Trial agent tooling in a test account or with read-only scopes until change attribution is proven. This is the same caution you would apply to giving a new contractor account access, and it is not cynicism.

AI surfaces as inventory

Search is becoming an answer surface, and ads are following. AI Mode is being promoted as an entry point from ordinary Search and is now completing transactions — as of August 2026 it books hotels with ten US partners, with hotels keeping merchant of record, and surfaces flight alerts across 180 countries.

The strategic question this raises for advertisers is not about bidding. It is that when a booking completes inside the search surface, the competition shifts from winning the click to being one of the integrated partners. That is a business development question wearing a media buying costume.

What stays human

The parts automation is worst at are the ones that need judgement about the business:

• what a conversion is worth, and which conversion to optimise toward
• the offer and the creative angle
• whether reported performance is incremental
• when to stop

None of those are settings. All of them determine whether the automation is pointed at something worth buying.

What trips people up

• Granting agent write access to a production account because a demo was impressive.
• Assuming AI surfaces are someone else's problem if your category is not yet affected.
• Believing automation reduces the need for measurement. It raises it, because errors scale faster.

You have got this when

You can say what you would let an agent do in your account today, what you would not, and what evidence would change your mind.