Agents aren't autocomplete tools anymore. They can handle multi-hop problems, work through ambiguity, and complete tasks that would have required a specialist twelve months ago. But capability alone doesn't determine outcomes. How you scope the agent does.
"As long as you scope these agents correctly, you give them the right tools, you give them the right goals, you give them the right environments to run around in, in my opinion, there is no ceiling," Huang added
The ceiling comes when you ask it to do too much at once.
Do this: Before deploying any agent, define the exact task, the success criteria, and the tools it has access to. Treat it like onboarding a new hire who needs context.
How to Pick Your First AI Use Case
The instinct for most teams when starting with AI is to reach for the most impressive use case available. That's usually the wrong move. A more durable approach starts with a simple filter: is the cost of doing this manually meaningful, and is the failure mode survivable if the AI gets it wrong? If both answers are yes, the next question is whether you have a clear enough success metric to know fast whether it's working. Without that signal, you can't iterate, and iteration is the whole game.
Companies that move carefully on use case selection tend to scale faster because they don't spend months debugging agents that were set up to fail.
Concrete examples worth testing:
- personalizing sales collateral at volume
- giving customer support reps an AI copilot to cut ticket resolution times
- qualifying inbound leads through voice or chat agents
These are meaningful, measurable, and survivable if imperfect. Coyne also cautioned against stringing too many autonomous agents together too early. Why? Compounding errors erode user trust fast, and that's hard to recover from.
Do this: Before committing to any AI use case, ask two questions first. Is there real cost to automate, and is the failure mode survivable? If both check out, define a single, clean metric that tells you whether it's working before you build anything.
Generative Content Is Already Doing Real Work
AI-generated content—voice agents, video avatars, personalized outreach—isn't a test case anymore. Portfolio companies across these firms have moved past pilots and are running generative content as core GTM, at scale, with the kind of results that make it hard to argue for doing it any other way.
ElevenLabs is using its own voice model to qualify inbound SDR leads at scale, across multiple languages and handling the qualification workload that would otherwise require a team of reps. Synthesia avatars have delivered corporate earnings calls, front-line worker communications, and interactive website experiences that route visitors through a lead funnel. Personalized AI video is opening sales sequences. The common thread across all of it: audiences aren't filtering for whether something was AI-generated. They're filtering for whether it's relevant and useful.
