The AI maximalist position is seductive: models keep improving, costs keep falling, and eventually every task that a human does today will be done cheaper and better by software.
We think this misses something important.
Not because AI isn't improving — it is, faster than most people realize. But because the value of what humans bring to the table isn't fixed. It shifts as the surrounding system changes. And as AI takes on more of the work, the specific things humans do well become more valuable, not less.
Three things AI still can't do
There's a useful framework for thinking about cognitive work in layers. At the base is pure computation — pattern recognition, information retrieval, text generation. AI has this covered and is getting better at it daily.
But above that are two layers that remain genuinely hard for AI:
Judgment under uncertainty. Not the kind of judgment where you weigh probabilities and pick the highest-confidence answer. The kind where the stakes are high, the context is ambiguous, and you need to be able to explain your reasoning to another person and stand behind it. Compliance decisions. Medical triage. Legal interpretation. The situations where "the model said so" isn't a sufficient answer.
Physical presence. AI can see through cameras and speak through speakers, but it can't walk into a building, shake a hand, read the energy in a room, or verify that a lock has been changed. The physical world is still entirely human territory.
And then there's a third, underappreciated one:
Trust. There are things people will tell a human that they won't tell a bot. A phone call from a real person lands differently than an automated message. When a customer has a complaint, being heard by a human matters — not because the human necessarily resolves it better, but because the interaction itself is different.
The agent bottleneck
Here's where it gets practically interesting for anyone building AI products today.
Agents are getting remarkably capable. They can research, write, plan, and execute multi-step workflows. But they keep hitting the same walls: bot detection on websites, phone trees that require a human voice, physical verifications that require someone to show up, decisions that require explainable judgment.
Right now, when an agent hits one of these walls, the workflow stops. A human has to step in manually, context switches, and the whole thing becomes harder to run at scale.
The bottleneck for most agentic systems isn't AI capability — it's the handoff to humans when AI capability runs out.
That handoff is currently unstructured, ad-hoc, and impossible to automate. The agent has no standard way to say "I need a human here." And the human has no standard way to receive the task, complete it, and hand results back in a format the agent can use.
That's the gap Humant is designed to fill.
Humans as infrastructure
Twilio built the infrastructure for programmatic phone calls. Before Twilio, adding telephony to your app meant buying hardware, negotiating with carriers, and maintaining a system you didn't want to think about. After Twilio, it was a POST request.
We think human labor is at the same inflection point.
Right now, "hiring a human to do a task" means job postings, contracts, onboarding, management, and all the overhead that comes with it. It's expensive, slow, and doesn't compose with software systems.
What if it were a POST request instead? What if your agent could dispatch a phone call, a web navigation task, or a judgment call to a vetted human worker — and get structured results back via webhook — the same way it calls any other API?
That's the vision. Not AI replacing humans. AI and humans working in the same system, each doing what they're actually good at.
The human dividend
There's a social dimension to this that we care about deeply.
The worry about AI and work isn't unfounded. Automation does displace some jobs. But the response to that can't just be "retrain everyone" — the timeline doesn't work, and the disruption is real.
We think there's a different path: building systems that expand the range of tasks humans can be paid to do, rather than contracting it. A worker in rural India who has a phone and good judgment can complete a voice call task for a company in San Francisco. A student who needs flexible hours can pick up navigate-and-extract tasks between classes.
This isn't charity. It's an economic model that works because human judgment, presence, and trust are genuinely valuable — and increasingly scarce relative to the AI systems that need them.
What we're building toward
We're starting with two primitives — voice calls and web navigation — because those are the walls agents hit most often, and they're well-defined enough to deliver reliably.
But the vision is bigger. Every place an AI agent currently fails gracefully is a place where a human-in-the-loop primitive could exist. We're going to build those primitives, one at a time, until the handoff between AI and human is as clean and composable as any other API call.
The most important question in tech isn't what AI can do. It's who it serves.
We're building Humant to make sure that answer keeps getting broader.