On a short instruction, GPT-4o got it right 91% of the time. On a longer version of the exact same task, it got it right 1% of the time. Not 70%. Not 40%. One percent โ statistical noise (PNAS Nexus, 2026).
That collapse is the single most important number a Head of Operations can carry into a 2026 agentic-AI decision, because it overturns the assumption every rollout is quietly built on: that if an AI agent handles a task well in a demo, it will handle a slightly harder version of that task slightly less well. It won't. It will handle it near-perfectly right up to a threshold, and then fail almost completely. The question for operations is not whether to trust an agent. It's knowing exactly where the cliff sits โ and putting a human-in-the-loop gate in front of it.
The finding: a cliff, not a ramp
The study comes from Suketu Patel and Jin Fan at CUNY, published in PNAS Nexus and surfaced publicly in June 2026 (PsyPost, 2026). Their method is elegant precisely because it's old. They ran the Stroop task โ the classic cognitive-psychology test where you name the ink color of a word while the word itself spells a different color โ on frontier language models, and they scaled the load by lengthening the list of conflicting items from a handful up to forty.
The results are stark. GPT-4o answered incongruent trials correctly 91% of the time on short five-item lists. Push the list to twenty or forty conflicting items and its accuracy fell to 1%. Claude 3.5 Sonnet held its footing longer โ but it, too, eventually dropped, to roughly 10% at forty-item lists (PNAS Nexus, 2026).
Read the shape of that curve, not just the endpoints. Performance did not decay gracefully as the task got harder. It held near-human levels, then fell off a cliff. This is the finding that should reorganize how operations thinks about agent reliability: the danger zone is invisible from the demo. A clean pilot on a simple case tells you nothing about where the model breaks, because the breakage is discontinuous.
Why this is executive control failing โ not intelligence
It would be easy to file this under "AI still makes mistakes" and move on. That reading misses the mechanism, and the mechanism is the whole point.
The Stroop task doesn't measure knowledge or reasoning horsepower. It measures executive control โ specifically the ability to hold a goal in mind ("name the ink color") and inhibit a stronger, automatic competing response ("read the word"). In human cognitive neuroscience, attention decomposes into distinct systems, and executive control is the one that governs goal maintenance under conflict. It is a different faculty from raw pattern recognition.
Here is what the models revealed about themselves. Transformer architectures are extraordinary at the prepotent response โ the fast, automatic, statistically likely answer, the machine equivalent of reading the word instead of naming the color. What they lack is robust inhibition when the context grows long or fills with conflicting signals. The authors frame this as a missing capacity for the executive control that genuine general intelligence would require (PNAS Nexus, 2026).
For operations, translate it this way: an AI agent is not a junior employee who gets tired and sloppy in proportion to the workload. It is a system that maintains one goal beautifully until the number of competing constraints crosses a line, at which point goal maintenance doesn't degrade โ it evaporates. The failure isn't "worse output." It's the model quietly optimizing for the wrong, easier objective while producing fluent, confident text that looks exactly like success.
What "long and contradictory" looks like on your floor
Twenty conflicting words in a lab is abstract. Your real workflows are worse.
Think about what you actually ask an agent to do. Reconcile an invoice against a contract that has three amendments, a special-terms clause, and an exception someone emailed last week. Route a customer complaint under a policy that says one thing, a manager's standing instruction that says another, and a promotion that overrides both until Friday. Draft a compliance response that must satisfy the regulator, the legal team's risk posture, and the account manager's relationship โ three goals that don't fully align.
Every one of those is a Stroop task with the volume turned up. Long context, multiple simultaneous constraints, and a strong "obvious" answer that happens to be wrong once you account for the exceptions. These are exactly the conditions the study shows push executive control off the cliff. And they are also precisely the tasks mid-market operations is most eager to automate, because they're the tedious, judgment-heavy ones that eat a team's hours.
That is the trap. The tasks with the highest automation appeal overlap heavily with the tasks most likely to trigger silent collapse. The agent will demo flawlessly on the clean invoice and fail near-completely on the one with three amendments โ and it will fail confidently, which is the dangerous part.
The business cost of ignoring the cliff
This is not a theoretical concern, and the market is already pricing it. Gartner projects that more than 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls (Gartner, 2025).
The Stroop finding tells you why so many will fail. Teams pilot an agent on a curated, low-conflict slice of the work, see 91%-style accuracy, and scale it into the messy, high-conflict production reality โ where the same agent operates past its collapse point and quietly produces 1%-quality output on the hardest cases. The costs that follow aren't labeled "AI failure." They show up as reconciliation errors caught three steps downstream, compliance responses that need full human rework, and the erosion of trust that eventually gets the whole initiative shelved. The project doesn't die from a dramatic incident. It dies from accumulated silent failures on exactly the cases that were supposed to justify it.
The organizations that keep their agentic projects out of that 40% won't be the ones with better models. Everyone has access to the same frontier models. They'll be the ones who designed for the cliff instead of pretending the curve was a smooth ramp.
Where to put the gate
The instinct, when an agent underperforms, is to reach for better prompts. This study says prompting is the wrong lever. You cannot prompt your way past a structural absence of executive control; you can only relocate the cliff slightly. The durable lever is architectural: a hard human-in-the-loop gate placed before the collapse point, not after the incident.
Concretely, that means three moves this quarter.
1. Map your workflows by conflict load, not by task type
Stop sorting candidate tasks into "simple" and "complex." Sort them by how many competing constraints they carry and how long the relevant context runs. A high-volume task with one clear rule is a safe automation target. A lower-volume task with three overlapping policies and a stack of exceptions is where the cliff lives โ regardless of how routine it feels.
2. Set the gate before the threshold, empirically
For any agent workflow with conflicting constraints or long context, insert a mandatory human review point. Don't guess where the model breaks โ test it the way the study did. Feed the agent progressively messier versions of a real task and watch for the discontinuity. Place the human checkpoint on the near side of it.
3. Instrument for silent failure, not loud errors
The collapse doesn't announce itself; the output stays fluent. So you cannot rely on the agent to flag its own low-confidence cases. Build sampling and spot-audit into any high-conflict workflow, and treat "the demo worked" as the beginning of validation, not the end of it.
None of this requires new technology. It requires treating agent reliability as a property of the workflow design, not a property of the model โ and accepting that the model's competence has an edge you can find but cannot prompt away.
The decision for this quarter
Pull the list of processes your team is planning to hand to an AI agent in the next two quarters. Next to each one, write down two things: how many conflicting rules or exceptions it carries, and how long the context an agent would have to hold. The tasks that score high on both are not your quick wins. They are your cliff cases โ and the study says they'll pass the pilot and fail the job.
For those, the move is not a better agent. It's a human-in-the-loop gate placed deliberately in front of the collapse point. The 40% of agentic projects headed for cancellation will mostly be the ones that mistook a demo for a guarantee. An agent that aces the clean case and collapses on the messy one isn't a tool you've deployed. It's a liability you haven't found yet. Go find the cliff before it finds you.
