Full Autonomy Breaks Down: She's Betting on AI That Calls for Help When Stuck
AI often stumbles when left alone. Foe's founder bets on semi-autonomous mode—when AI gets stuck, human experts step in immediately—rather than chasing pure full autonomy.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Give AI a Phone Line to Rescue Failed Tasks
She started with a specific obsession: if AI can't complete a task, it should learn to call for help like humans do. She notes roughly 40% of the global population and scenarios aren't digitized yet, so many tasks require phone calls; but models weren't trained for collaboration since training data came from Copilot-like "human supervises AI" modes, not AI proactively admitting failure and asking for help. The team had to specifically teach the model "introspection": recognize it's failing, then judge who to ask.
— Shivani PoddarRedesigning the Call Opening Beats Better Voice Models
When Foe launched, roughly 80% of people who received AI calls hung up immediately; now it's 20–30%. The problem was misaligned incentives—the caller (AI) has a clear task to complete so it stays on the line, but the receiver doesn't know why they're getting this call or why they should listen. This forced the team to completely redesign the opening and pacing of each call, not just swap in a better voice model.
— Shivani PoddarPhone Agents Can't Just Add a Better Voice Model—Safety Redesigns Everything
Beyond making AI sound natural and understand conversation turns, the team had to solve two overlooked problems: regulation requires AI to disclose its identity and declare the call will be recorded, which makes listeners instinctively tense; worse, a bad actor could flip it and trick AI into leaking user personal information. Stack these three together and voice agents can't just plug in a better voice model—collaboration and safety logic must be redesigned from scratch.
— Shivani PoddarBuild Your First Product Only for Yourself
Her first version was entirely designed for one "super user"—herself—to handle scheduling for recruiting and investor meetings. She deliberately skipped the OpenClaw approach of keeping data on-device, instead building remote, emphasizing empathy and reliability. Her logic: if the product can't serve her well, it has no right to serve anyone else; if it fails her, it doesn't deserve to be used by others.
— Shivani PoddarHuman in the Loop Means Hiring Experts, Not Gig Workers
Foe's human involvement isn't Craigslist-style crowdsourced gigs but interviewed experts more like people the CEO hired to do work. AI doesn't dump the whole task on humans but uses "subtask delegation": stuck on a CAPTCHA, just hand off that one step and keep going. This way humans don't carry the whole weight, only handle what they excel at and what AI can't yet do, multiplying efficiency.
— Shivani PoddarTrust Has Two Layers: Privacy First, Then Task Completion
She splits trust in two layers: first, AI won't leak user private information to the wrong people—she mentions a user whose address got leaked on Marketplace to the wrong person who showed up at their door; second, whether the task actually gets completed. She thinks users pick an agent based on the second layer, but if the first breaks, trust collapses instantly. This isn't like social media's abstract consequences—this happens to you in real life.
— Shivani PoddarHandwritten Recruiting Notes Show Commitment More Than Polished Emails
As a solo founder, she didn't pursue cold recruiting. Her first people came entirely from her trusted network—including people from startups she knew were brilliant. When she messaged to recruit, she insisted on handwriting, just three words: "Want to get on the rocket?" She thinks that personal touch itself signals how much she cares about bringing them on.
— Shivani PoddarAI Tool Progress Now Lets Solo Founders Skip the Co-Founder Search
She interviewed people who quit mid-venture and became solo founders, concluding: she's not against finding partners, but won't hastily rush-hire one to escape "carrying it alone." She thinks the cost of not shipping outweighs the time spent learning to self-regulate and fill gaps; more crucially, the last few months of AI tool progress are the first time solo founders can patch ability gaps that only co-founders could fill before.
— Shivani PoddarIn their own words · checked verbatim
I'd never be against having a co-founder, but I'd never rush into it just because I wouldn't want to do it solo.
Shivani Poddar0:00
I built the first ever social chit chat bot that was deployed in Alexa back in 2016. And then I built the first Facebook assistant, which was back in 2018, which was a device called Facebook portal.
Shivani Poddar5:07
I wanted to give the agent a phone. You know, if you're failing, just call somebody. About 40% of the world is not digitized yet, believe it or not.
Shivani Poddar6:08
they're not trained to collaborate, because they're trained on dreaded data, which is like co pilot.
Shivani Poddar7:11
our phone call hang up rate where people would listen to us and like hang up was like 80%
Shivani Poddar10:17
Now we're like at the hang up rate, it's between 20 to 30%, which is insane
Shivani Poddar10:17
if the product fails me, then the company doesn't have a right to be successful for this product to be used by anybody else
Shivani Poddar29:06
we didn't want it to feel like uh throw over the fence kind of a relationship between the ai and the human
Shivani Poddar37:27
Figures
| AI voice call initial hangup rate | ~80% | 10:17 |
| AI voice call current hangup rate | 20%–30% | 10:17 |
| Undigitized global population and scenarios | ~40% | 6:08 |
| ARR of Australian insurance company that reached out | $5 billion | 14:22 |
Glossary
- subtask delegation
- AI hands off only one small step to humans when stuck, rather than dumping the entire task.
- human in the loop
- True human experts jump in during AI execution to supplement what AI can't yet do.
How to listen
Founders and product managers building personal AI assistants who are torn between full autonomy and human-AI collaboration.
The podcast's self-promotion segment around 7:30–8:15 can be skipped.