Sonya Huang, a partner at Sequoia Capital, invested in OpenAI in 2021, before ChatGPT existed. She now says the agentic AI bottleneck has shifted: the capabilities have arrived, and the constraint is no longer execution. “The capabilities are finally there,” she says.
Her framing is worth sitting with: “In a world of agentic AI, you have a team of interns, a whole company, at your fingertips. People with entrepreneurial vision and passion, the ideas guys, get to run the world.”
That is not a tech prediction. That is an organizational claim. And she is right, with one important qualifier.
What Changes When the Agentic AI Bottleneck Moves
For most of the last decade, the constraint on building was execution. You needed engineers, designers, and operators to turn an idea into a product. That capacity was expensive, scarce, and slow. Agentic AI is compressing that constraint. Coding is already largely agentic. Computer use, the ability to hand an agent a screen and a task, is next. When those capabilities land broadly, the gap between conceiving something and shipping it collapses.
As a result, the bottleneck moves. It does not disappear. When execution gets cheap, the premium shifts to judgment: knowing which ideas are worth building, which problems are worth solving, and which markets are worth entering. That has always been the harder work. Agentic AI makes it the only work that matters competitively. Organizations that generate and evaluate ideas well will widen their advantage. Organizations good primarily at grinding through execution will lose the protection that slowness once provided.
Huang’s investment lens reflects this. She backs founders who can adapt as the state of the art changes every three months. Not the people with the best current solution, but the people most likely to hold the right position as the landscape shifts beneath them. Her portfolio includes Harvey, doing exactly this in legal services: targeting the analytical and document-intensive work lawyers do at volume, so their judgment and client work remain the scarce thing.
The pattern generalizes. Pick any category. Look closely at where human effort accumulates most densely around repetitive, structured, high-stakes work. That is where agentic AI lands first and hardest. The question is not whether your industry is affected. It is where the bottleneck moves when it is.
Source: Fast Company, June 18, 2026
Get new articles from Robin Green delivered directly.
Insights on AI leadership, the future of work, and human collaboration. No noise. Unsubscribe any time.
Subscribe free
Robin — as you say, the question isn’t whether an industry is hit. It’s the ground — where the bottleneck moves. “Human effort” is the commodity constrained at the bottle’s neck: value lost to reduced productivity. Huang’s positioning applies directly to the ground therapy holds — or risks losing — as things shift.
Therapy’s Hit: Same Bottle, Two Necks
“Great, I can reduce the time-sink of producing psychotherapy notes with AI” — it is truly great! → Not the highest stakes, not the highest value.
All therapists find themselves, at one point or another, in therapy with a specific client — bottlenecked, stuck. → Yes, the highest stakes and highest value.¹
Huang’s read is right for the reason you gave. She isn’t backing the best current answer — she’s backing whoever holds position when the ground shifts again in three months. Harvey works because law splits clean: document review is volume, client judgment is scarce. Automate the first and the second stays untouched.
That split is what makes the Harvey model portable, and valuable.
The Ground: Law = Therapy
Therapy has no such split. Look at where human effort piles up densest around repetitive, structured, high-stakes work — you said that’s where agentic AI lands hardest. That’s not an adjacent case for clinical work. It’s the definition of a stuck client and therapist.
And that is judgment + therapeutics (how the client, the “industry,” moves) = outcomes: good v. bad, stuck v. processing/healing, value v. productivity.
That equation is the practical meat of our industry — and it’s exactly where the ground has to move. iVASA.ai, a working, fully literate AI-driven therapeutic platform, is built to hold it.
Every turn, the agent reads where the client actually is — not a “CBT” script cycling through, but a live field read gating a posture overlay tied to the client’s stage in the work. The agent’s posture moves because the client moved, not because a timer did.
High-stakes²: War is the only other. In therapy, that’s not even a question.
Structured: Our system is built on it — identifying the contradiction, how it’s restaged across relationships, jobs, silences, session after session, until the client names it. That’s not chaos. It’s pattern recognition.
Repetition: This isn’t noise to our architecture. It’s the object our architecture is built to detect.
Loosening the note-writing bottleneck is a convenience. Catching the client stuck in front of you, session after session, and knowing the difference between stillness and stall — that’s the actual practice. That’s the bottleneck iVASA was built against.
#MentalHealthAI #ClinicalAI #AISafety #DigitalMentalHealth #PsychoContextualAnalysis
¹ The Therapist’s Paradox: that’s how we make our money. “Fixing” a client means we have to find another. We’re incentivized not to do this — to keep the bottleneck intact.
² Acute risk monitoring (suicidality, psychosis, SIB, delusions, mania) stays separate from all of it. A second model, not the therapeutic one, reads every utterance against fixed structural definitions — narrating versus moving to enact, baseline versus an acute break. It has exactly two moves: hold and return judgment to the client, or escalate to a human, bounded and non-discretionary. The agent never sits in the middle grading its own homework on whether the person in front of it is drifting toward becoming the case that isn’t stuck anymore — the one that’s over.
Mathew, the Harvey split is the right lens, and you have named exactly where therapy resists it. Law has a seam: volume work on one side, judgment on the other. Therapy collapses that seam. The structured, repetitive, high-stakes work IS the clinical work. That is not overhead you can separate out; it is the practice itself.
The architectural decision you describe on acute risk: a bounded second model with exactly two moves. That is the kind of constraint that separates serious clinical AI from the note-writing convenience tier. The incentive structure footnote is honest in a way most clinical AI marketing is not. Worth a longer conversation.