It was 2 a.m. on a Wednesday. I had just finished reviewing a quarterly roadmap for an agentic AI initiative that had taken months of planning. Then I watched a video of an AI agent assembling an advertising operation in a single sitting.
I watched it twice. The contrast with the roadmap was uncomfortable, but the useful comparison was narrower than the headline. The video showed how much production work a person could coordinate through an agent. It did not show a functioning client business over time.
That distinction matters to anyone buying agency services, running an agency or managing an internal team. A campaign can be ready to launch while the business responsible for it is barely ready to begin.
What the demonstration establishes
Brett Malinowski's demonstration presents a sequence of competitor research, brand development, website creation, advertising production and outreach using browser control. The reported run takes about 45 minutes. I am treating that as the creator's demonstration, not an independently timed or audited operating result.
The sequence is worth studying because it crosses application boundaries. Work that normally passes between a researcher, designer, website builder and campaign operator can be brought into one coordinated session. The person decides the objective and sequence; the agent performs much of the visible execution.
What we cannot infer is equally important. We do not know the preparation behind the recording, the interventions outside the edited sequence, the maintenance burden or the eventual campaign results. A deployed advertisement is evidence of a completed action. It is not evidence that a client acquired profitable customers.
The work that begins after launch
An agency client pays for an ongoing commitment. Someone must notice when inventory changes, when an advertisement attracts the wrong audience, when a claim is disputed and when the budget is being spent without a useful result. Someone must be able to explain what happened and agree what to change.
Those obligations remain even if the first creative assets become cheaper. They may become more demanding when a system can create and release many variations quickly. A small mistake reproduced across a campaign needs detection, correction and communication, not another prompt that makes the next image look better.
This changes the staffing question. Instead of asking how many specialists can be replaced, ask which decisions need specialist judgment and how often those decisions occur. An experienced media buyer may spend less time constructing campaigns and more time interpreting weak performance. A designer may produce fewer routine variants and spend more time defining what the brand should sound and look like in a new context.
There is also work that should simply disappear. Copying approved information between systems is a poor use of someone's expertise. Removing that task can improve the service without pretending the relationship itself has become autonomous.
A campaign cost is not an agency cost
Consider an illustrative comparison, not a client result. A team currently spends ten hours preparing a campaign and another six hours reviewing, monitoring and handling changes. An agent reduces preparation to two hours, but review and operation still require six. The saving is eight hours out of sixteen, not fourteen hours out of sixteen.
If the agent adds two hours of checking and repair, the net saving falls to six. If the client requests twice as many variants because production is easier, the saving may appear as more output rather than fewer paid hours. None of these outcomes is necessarily bad. They answer different business questions.
The comparison should include software and inference charges, permissions, integration, review, correction and the ongoing owner. Media spend belongs in the campaign economics too, even though automating production does not remove it. Calling marginal delivery cost almost zero hides the part of the service the client will still expect someone to provide.
An agency can respond in several ways. It can lower prices for standardized production, reinvest time in better analysis, serve additional clients within its review capacity or sell a narrower automated service with clearly defined boundaries. The choice depends on what customers value and what the business can reliably support.
Release authority deserves its own design
Researching a competitor, preparing a draft and spending a client's advertising budget should not share the same permission boundary. A workflow can let an agent gather evidence and create variants while reserving release and budget changes for an accountable person.
The reviewer needs more than a polished preview. They need the source inventory, the claims used, the intended audience, the approved budget and a record of what will be published. If a vehicle price or product specification has changed, the agent should stop or refresh the source rather than improvise.
The same applies to outreach. Finding a business and sending a message are different actions. A useful research result does not automatically authorize contacting everyone it identifies. Delivery rules, client agreements and applicable advertising and data requirements must be checked for the actual service being offered.
I would also make rollback part of the release plan. Who can pause a campaign? How quickly can they remove a misleading asset? What does the client see when an error occurs? These are ordinary service obligations made more urgent by faster production.
Buy the service the demonstration leaves out
When evaluating a supplier, ask for one campaign followed through its whole lifecycle. Look at the source brief, review, launch, monitoring, exceptions and the final performance discussion. Ask which actions the agent took and which decisions a person made. A supplier that can explain those boundaries is easier to trust than one selling an impressive speed record.
The same test applies inside an enterprise. Pick one production-heavy workflow and see whether the team can reduce execution effort without losing the checks that make its output useful. Start with a reversible, bounded assignment. Compare completed work and total effort, including the corrections people would otherwise absorb quietly.
The demonstration changed my view of what a small team can produce. It did not change what I would expect that team to stand behind. The opportunity is to make dependable service less expensive to deliver. Dependability is part of the product.