The most dangerous phrase in process automation is "human in the loop," because it describes a location, not a power. A human who reviews machine output without time to review it, authority to reject it, or context to judge it is not in the loop; she is in the loop's documentation, present for liability purposes, absent for all others. As agentic systems take over multi-step work, invoices to campaigns to code deployment, the real design question is harder and more specific: not whether humans stay involved, but where a human gate belongs, what powers it holds, and who mans it.
The case for gates is not sentimental. It comes from the best field experiment yet run on AI-assisted knowledge work. Dell'Acqua and colleagues randomized 758 Boston Consulting Group consultants across tasks with and without GPT-4. Inside the frontier of the model's competence, AI users completed 12.2% more tasks, 25.1% faster, at roughly 40% higher quality: automation earning its keep. But on tasks deliberately chosen to sit outside that frontier, AI users performed worse than the control group, because the model produced plausible, confident, wrong answers and people accepted them. The authors call the boundary "jagged": invisible, irregular, and impossible to map from inside a workflow. A pipeline that runs at machine speed across a jagged frontier will, by construction, cross it without noticing. The gate is what notices.
Where the gates go
Three placement rules follow from the evidence rather than from taste.
Put gates where errors are expensive and irreversible. Reversible, low-stakes machine errors are a cost of doing business, usually cheaper than the human labor of preventing them. Irreversible ones, funds moved, offers made, records deleted, reputations spent, justify the friction of a stop.
Put gates where the frontier is known to be jagged. Wherever the model's failure mode is plausible-but-wrong rather than obviously-broken, the BCG result says unreviewed output is a trap, and the review must be genuinely adversarial, not a skim. This is also where the reviewer needs the most expertise, which has an organizational implication: gates are senior jobs, not clerical ones.
Put gates where the people doing the work say judgment is the work. A 2025 Stanford audit crossed the preferences of 1,500 workers across 844 tasks with expert assessments of AI capability, introducing a Human Agency Scale to express how much human involvement each task warrants. The finding was consistent: workers want low-value tasks automated outright, and on high-value tasks they prefer shared agency, machine execution with human judgment and responsibility retained. Strikingly, the same study found investment flowing into automating tasks where workers do not want automation, a misalignment that predicts both failed adoption and quietly bypassed controls. Workers are, in effect, drawing the gate map themselves; companies mostly are not reading it.
What a gate needs to be real
A gate is defined by its powers, and there are three without which it is scenery. The power to stop the pipeline, unilaterally, without escalation theater. The visibility to judge: which is why gates depend on the accountability infrastructure researchers have proposed for agents generally, identifiers, monitoring, and activity logs, so the gatekeeper can reconstruct what the system did and why (Chan et al., 2024). And a name. Anonymous review pools diffuse responsibility; a gate is one accountable person whose signature means something, the way an engineer signs a bridge.
There is an organizational failure mode to design against explicitly: the rubber stamp. Volume pressure converts reviewers into approvers; approval rates drift toward 100%; the gate becomes latency with a salary. The countermeasures are unglamorous, sampling for depth instead of reviewing everything shallowly, tracking overturn rates as a health metric (a gate that never rejects is dead), and protecting gatekeeper time as jealously as production time.
The line that holds
It is tempting to frame all this as a temporary concession, gates as scaffolding to remove when the models improve. The jagged-frontier result argues otherwise: as capability grows, the frontier moves but stays jagged, and the plausibility of wrong answers improves along with the accuracy of right ones. Better models make weak review more dangerous, not less, because there is less obvious wrongness to catch a skimming eye.
So the durable principle is about accountability, not capability: machines can hold tasks, but only humans can hold responsibility, because only humans can answer for a decision to other humans. A pipeline with no point at which a person can say no is a pipeline whose failures belong to no one. The companies automating well are not the ones with the fewest humans in the flow. They are the ones who can say, for every consequential step, exactly which human can stop it, and who have made sure that saying no is still a career-safe thing to do.
Sources
- Dell'Acqua et al. (2025). Navigating the Jagged Technological Frontier. Organization Science. RCT, 758 BCG consultants: +12.2% tasks, 25.1% faster, ~40% higher quality inside AI's frontier; worse than control outside it due to plausible wrong answers. https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838
- Shao, Zope, Jiang, Pei, Brynjolfsson, Yang et al. (2025). Future of Work with AI Agents (WORKBank). arXiv. Workers prefer automation of low-value tasks and shared agency on high-value ones; current agent investment partly misaligned with preferences. https://arxiv.org/abs/2506.06576
- Chan et al. (2024). Visibility into AI Agents. ACM FAccT '24. Agent identifiers, real-time monitoring, and activity logs as preconditions for accountability over delegated agentic work. https://doi.org/10.1145/3630106.3658948