Agents make the work fast. Clients pay for whoever answers when it breaks.
A campaign can be approved and still miss its launch. The landing page is built, the copy is signed off, and the media team has booked the traffic. Then a market lead spots an offer that doesn't apply locally. Someone has to find out which versions are affected, fix them, and get the right page live before the first ad sends customers to it.
The creative work is done. The customer is still waiting.
That stretch of work is where the agency holding company WPP wants to put AI agents. On October 20, WPP Enterprise Solutions, its business transformation unit, announced a partnership with Gradial, a Seattle startup that builds AI agents for marketing operations. Gradial supplies the agents, and their remit covers quality assurance, accessibility checks, compliance validation, content migrations and publishing. WPP leads the integration and deployment, runs the platform for clients as a managed service, and has set up a dedicated Agentic Marketing Operations Center of Excellence to deliver it.
I don't doubt the agents are fast. Speed just isn't what I'd pay for in that last stretch.
What a brand actually buys there is the person allowed to say yes, and the person who picks up the phone when the wrong page goes live. The uncomfortable part for agencies is that the work they are now automating is the work that used to produce those people.
Under the proposed setup, a brand keeps its content management system and its digital asset library. WPP says it plugs Gradial into the systems a client already runs instead of replacing them. That sounds like the easy version. It's the hard one, because the brand now has to translate how it already works into permissions an agent can use.
An agent might be allowed to fix a broken link in a draft and send it for review. Changing the terms of an offer on a live page could require someone in that market to sign off.
Both jobs run through the same login. Access is not authority, and someone has to write down the difference before the agent finds it.
Take a page flagged because an image has no description. The agent can write one in seconds. A reviewer still has to decide what the image is there to communicate and whether the proposed description does that job. If the page needs that person's approval, it can sit in the sign-off queue all afternoon, even though the agent assembled it in minutes.
WPP's release says early deployments cut content execution timelines by up to 90%. It doesn't give the baseline, the number of deployments, or the typical result. WPP's Gradial partner page goes further and claims that automating brand, legal, and accessibility checks cuts approval times by up to 85%.
The build was never the slow part. I doubt 90% survives a brand's own sign-off chain, and I'd ask to see the 85% proven first, because approval is exactly where I'd expect the speed to disappear. Someone also has to be on standby when the agent gets stuck. I want to see that person on the invoice.
WPP's Q3 2026 trading update, published October 29, shows why. Revenue less pass-through costs fell 2.1% year over year on a like-for-like basis. WPP Media grew 0.3%, and WPP Production grew 2.5%. WPP Creative, the division that includes Enterprise Solutions, fell 5.1%, worse than its 3.5% decline in the second quarter.
So the unit selling Gradial sits inside the only one of WPP's three reported divisions that shrank, and the shrinking sped up.
WPP said the decline was concentrated in its integrated creative agencies. It gave no separate growth rate for Enterprise Solutions, though CFO Joanne Wilson told analysts the unit did slightly better than Creative as a whole. CEO Cindy Rose called client demand healthy for WPP's AI-enabled services, powered by WPP Open, the company's agentic marketing platform. The quarter ended before the Gradial deal was announced.
A recurring fee for running a workflow could make an agency less dependent on the volume of assets it bills for. That is the appeal for WPP. But a client who expects automation to lower the bill still needs someone on call who knows the brand and can make sense of a local exception. If one person can cover ten accounts, the managed service is a business. If every brand needs its own, it's a staffing plan with a software invoice.
Most agencies have attached the agents to the org chart they already had. In an October 30 report from the trade publication Digiday's Media Buying Summit, Digiday reporter Sam Bradley described agencies putting agents to work while leaving their team structures, job descriptions and career paths largely as they were.
Gila Wilensky, who leads data and media in North America for the agency Monks, said AI fluency is now a baseline expectation. She said Monks isn't hiring fewer junior people, but it is rethinking what it asks them to do, and piecing together spreadsheets is coming off their plates. Digiday reported that early-career work may shift toward quality assurance instead. Wilensky went further and predicted a full-time job of "AI checkers," a role she said doesn't exist today.
So the job of checking the agents doesn't exist yet, and the work that would have trained people for it is being automated now.
I made a version of this argument in October about the work that taught people to catch the wrong logo ["Nobody Left to Catch the Logo"]. The old tasks were tedious, and they still taught someone to recognize a bad input or an inconsistency before it reached a client. An agency that automates them has to give junior staff another way to build that judgment. A supervised review of a flagged page can do some of it, especially if the junior employee has to explain why the page should be held and then watches how the decision gets made. Otherwise, the agency will be asking people to catch errors nobody taught them to see.
If you're evaluating the service, pick a workflow that already gives you trouble. Give the pilot a real release with a late change that affects only one market, and watch whether the agent keeps the change inside that market. In a test environment, rehearse what happens when an approval is missing or the wrong offer goes live.
Before anything runs, you and the agency should agree on who can stop a release and who has to restore the previous version.
Measure from the moment the change is logged to the moment the corrected page is verified live, including the wait for approval. Compare that with similar releases under your current process, and count the rework and the cost of human review.
If the agency can get the right page live sooner, at a cost you can accept, and tell you who picks up the phone when it goes wrong, it has something worth paying for.
If it can't name that person before the pilot starts, the 90% belongs to someone else's deployment.