Humans Stay in the Loop
ojas@ojas
+1.1%
since Oct 2
The idea
I think full autonomy in AI is overhyped and that humans stay in the loop far longer than the market expects, especially in the last mile of agentic systems. This basket owns the human-review companies the market treats as roadkill alongside the workflow, oversight, and eval software built on the assumption a person keeps checking the work, with steadier expert-judgment names as ballast and a cash cushion. It's spread across many names rather than any single bet, meant to be held for years while this mispricing plays out.
What it owns
- Companies that sell trained people checking and fixing AI output, priced by the market as if AI kills them, when the take says AI feeds them. · Concentrix CorporationIt runs human review operations at scale and keeps winning contracts as clients add AI; the take depends on revenue holding steady as people stay in the loop rather than getting automated away., Genpact LimitedIt sells process work that needs human judgment at the end, the layer the take says autonomy can't replace; the premise holds as long as clients keep paying for people to own outcomes., Innodata Inc.It provides the purest lever here: humans labeling and reviewing data for AI companies; the take holds if big labs keep spending on human review the way the market has actually paid for it., TaskUs Inc.It staffs the safety and trust teams that check AI output for platforms; the take depends on money flowing to human review rather than to smarter models alone., IBEX LimitedIt runs customer operations with people in the loop, priced as if full autonomy eats it soon; the take is that day comes far later than the market expects, so it keeps earning in the meantime.30%
- Software built on the assumption that a person approves, steers, or supervises the agent rather than disappearing. · ServiceNow Inc.It builds agents inside workflows where a person approves each step, the deployment shape that stays relevant whether enterprises adopt AI through supervised workflows or by replacing workflows entirely., Palantir TechnologiesIts design centers on humans steering the system while AI proposes and a person decides, especially in government and defense where control won't loosen quickly; contract wins show buyers with real stakes want human control., Salesforce Inc.It sells agents with human supervision baked into the product, the opposite of full-autonomy deployment; the take holds if enterprises actually buy supervised agent systems at scale., UiPath Inc.It orchestrates work across people and bots together, betting that the handoff between them is the product, not a temporary stage; it stays relevant if customers keep building mixed human-and-agent workflows., Pegasystems Inc.It builds workflow software that assumes a person stays in the loop for decisions, a smaller way to own the same bet as larger competitors; it has to keep converting customers to AI features while holding the human-decision core.27%
- The tools teams buy to watch whether agents are actually working, which gets more valuable the less autonomy is trusted. · Datadog Inc.It provides the observability layer where engineers watch what agents do in production; spending grows if agents ship but can't be trusted unsupervised, and shrinks only if agent platforms bundle monitoring for free., Dynatrace Inc.It sells production reliability checks, the tooling that exists because nobody trusts automated systems without oversight; it has to keep growing as AI workloads spread for this premise to hold.13%
- Steady companies selling professional judgment wrapped in software, slow-moving proof that humans plus tools beats tools alone. · RELX PLCIt sells expert data and analytics that professionals use to make judgment calls; the take depends on AI making its data more valuable to human experts rather than rendering it obsolete., Thomson ReutersIt sells tools for lawyers, accountants, and other professionals whose judgment is the product; it has to keep growing its professional segments and weave AI in without cutting out the expert.11%
- Marketplaces where businesses hire people to review, fix, and finish what AI started. · Upwork Inc.It's the marketplace where companies hire humans to check and fix AI's work, priced as an AI casualty when the take says AI creates more last-mile human work instead of less., Fiverr InternationalIt sells humans on demand for exactly the tasks AI gets almost right, and trades like the market thinks almost becomes fully any quarter now; the take flips that mispricing.10%
- Cash9%