AI agents can now control Google Home devices and handle much of the setup work required to get a business onto WhatsApp.
This comes after Google opened access to its Google Home Model Context Protocol server on September 16, which allowed compatible agents such as Claude and ChatGPT, to control connected devices and access event history.
Likewise, Meta introduced a WhatsApp Business MCP a day earlier, which gave agents a way to handle parts of account and messaging setup through conversation.
In both cases, you can usually tell whether the agent did what it was asked to do because the device either responded or the account was registered.
But that gets more complicated once agents move into creative work, where the result is harder to judge.
Video editing, in particular, presents its own challenges as the requested changes can be made correctly and still leave the piece worse off.
Here, a cut can remove context, miss the strongest part of a recording, or feel wrong for the audience.
Descript already lets people bring that work into Claude or ChatGPT through its MCP, where users can describe the edit they want and have the agent work with the platform, then decide whether what comes back is ready for a client or publication.
In this interview with DesignRush, Descript’s Head of Growth Marketing, Daisy Zhang, explains why the work left after an agent finishes matters just as much as the edit itself.
"A successful agent edit isn't just one that technically completes the task. It needs to get the content meaningfully closer to publish-ready while reducing the amount of work a person has to do afterward," Zhang says.
Who Is Daisy Zhang?
Daisy Zhang is Head of Growth Marketing at Descript, where she leads user acquisition and activation across paid channels including Google, Meta, YouTube, Reddit and ChatGPT, and is building AI-native workflows for performance marketing. Before Descript, she spent nearly five years at Chegg, where she built the company's TikTok influencer program from the ground up and won a Platinum AVA Digital Award for its "Edutainment" campaign. She previously led growth marketing at Udemy. Daisy holds a degree from New York University and is based in the San Francisco Bay Area.
Why AI Editing Review Time Matters
An editor still has to make judgment calls after the agent finishes.
They need to check that the story still holds together, that the speaker's meaning hasn't changed, and that the result still works for the audience and brand.
"Creative work is inherently subjective, so the goal isn't to remove the human from the process," Zhang says.
"It's to let an agent handle more of the execution while keeping the person focused on judgment. Is this the right story? Did it preserve the speaker's intent? Does it feel right for the audience and brand?"
A quick edit can still leave a reviewer with plenty to do, from checking the original material to making sure nothing important changed along the way.
The final story choices, anything sensitive to the brand, and the last check before publication, however, still need to stay with the person responsible for the work.
An agent can still handle much of the work before then, including cleaning up material, finding sections, removing parts of a recording, and creating clips.
There probably isn't one handoff point that works for every edit. What matters is whether the agent has actually taken enough work off the editor before they step back in.
But leave too many judgment calls until the end, and the reviewer can lose that time trying to work out what changed and why.
How Claude and ChatGPT Change the Editing Workflow
Descript is seeing people begin with the assistant rather than the editor.
Someone can ask Claude or ChatGPT to clean up a recording, find strong moments, remove sections, create clips, or help turn a longer piece of content into something new.
The request starts with what the person wants to make instead of which video editing control they need to use.
"It starts to shift editing from ‘I need to know how to use this tool’ to ‘I need to know what I want to make,’" Zhang says.
A person can stay in Claude or ChatGPT, describe the result they want, and have the agent work with Descript to carry out more of the edit without opening the editing interface and performing every step themselves.
"The interface can then become part of the conversation itself," Zhang adds.
That also means the person making the request may not see every choice as it happens.
Cleanup work or clip creation may need little intervention before review, while cuts that could change the meaning of an interview need closer attention.
What Agencies Should Test Before Client Work
"Start with a representative but low-risk project and evaluate the entire workflow, not just whether the agent successfully executes the prompt," Zhang says.
The test needs to feel enough like a real client project to show where the workflow starts to break down.
That includes the setup before the agent begins, whether the edit still reflects the original intent, how easy it is to see what changed, and how much fixing is left afterward.
"Teams should look at how much manual setup is required, whether the output preserves the original intent, how easy it is to review what changed, how much correction is required, and whether the workflow actually saves time after review is included," Zhang says.
A workflow that completes the prompt but leaves an editor with substantial repair work may save less time than expected.
If most of the review can stay focused on judgment rather than rebuilding the edit, more of the production work has actually been removed.
However, clients still expect the same quality and consistency regardless of how much of the production happened through an agent.
"That's especially important for agencies because efficiency only matters if they can maintain the quality and consistency clients expect," Zhang says.
And while Google Home and WhatsApp Business offer easier ways to tell when an agent has finished its job, creative work still ends with someone deciding whether the result is ready to publish.
So ask yourself. Once the review and corrections are included, how much work did the agent actually take off the team's plate?
