When AI Takes Over Your Routines, Who Takes Over the Exceptions?

I keep coming back to the same question every time someone shows me a demo of an AI agent handling a business process end to end. The demo is always clean. A procurement request comes in, the agent routes it, checks compliance, selects the vendor, generates the purchase order. No human in the loop. The process is fast, consistent, and auditable. Then I ask what happens when the vendor has a compliance exception, or the budget code does not map to any existing category, or the requester needs to expedite because a production line is down and they cannot wait for the standard three-day turnaround. The answer is always some version of "that gets escalated to a human." That handoff is doing more work than anyone admits.

Feldman and Pentland (2003) gave organizational theory a vocabulary I cannot stop using. My study-hub does not contain their original paper, so what follows is my understanding from how the IS literature cites them, particularly Leonardi (2011) and the way I wrote about routine dynamics before. They distinguish the ostensive routine from the performative routine. The ostensive is the routine as designed: the flow chart, the SOP, the process map, the thing you could write down and hand to someone new. The performative is the routine as enacted: what people actually do on a particular day with particular constraints, particular pressures, and particular exceptions that the flow chart never imagined. The two are related. They are not the same. And the gap between them is not error. It is where most organizational value lives.

Leonardi (2011) draws on this distinction directly. He writes that employees can alter the performance of a routine while still maintaining its ostensive qualities, the broad understanding of what the routine should do (Feldman and Pentland 2003). That line captures something essential. The performative routine diverges from the ostensive not because people are lazy or resistant, but because reality does not match the process diagram. The exceptions, the workarounds, the local adaptations, these are not noise in the system. They are the system working. They are how the organization actually functions when the ostensive routine hits something it was not built to handle.

Now think about what an AI agent takes over when it automates a routine. It takes the ostensive. The agent gets the script. It can follow the process map, execute the documented steps, handle the predictable cases. The exceptions, the moments where an experienced person looks at a situation and says "this one does not fit, I need to handle it differently," those stay with the human. Every vendor demo assumes the exceptions are a small percentage and the human handler is a safety net. I think that assumption is wrong in two ways.

First, the exceptions are not a small percentage in most complex organizations. Any domain where the environment is genuinely uncertain produces exceptions as a routine feature. The whole reason experienced people are valuable is that they have seen enough exceptions to recognize them quickly and respond adaptively. The exception rate is not noise around the standard process. It is the standard process.

Second, the exceptions are not just a different kind of task that can be handed to a different kind of worker. The performative routine depends on the ostensive one for its starting point. People improvise from the script, not in spite of it. Remove the routine cases from a purchasing agent's screen, and the exception arrives without the contrast that made it recognizable as an exception.

Giddens (1984) built structuration theory around the duality of structure. Structures, meaning the rules and resources that organize social life, are simultaneously the medium and outcome of action. People draw on structures to act, and in acting they reproduce or transform those structures. As I wrote about structuration theory and IS, this recursive loop is not abstract. It is exactly what happens when a team uses a system repeatedly and develops patterns of use that become the de facto way things are done. Orlikowski (1992) extended this directly to technology. Her argument, which I cited in my post on same tool, different outcome, is that technology is both a product of human action and a medium through which action occurs. The structural properties of a technology are not fixed at design time. They are enacted through recurrent use. Two groups using the same system develop different technologies-in-practice because their structural contexts differ. The technology as designed is the ostensive. The technology-in-practice is the performative.

When an AI agent takes over the ostensive routine, it does not just execute the steps. It becomes a structural participant in Giddens' recursive loop. Baird and Maruping (2021) make this explicit: agentic IS artifacts are not passive tools. They can initiate action, accept rights and responsibilities, and operate with autonomy under uncertainty. Their delegation framework replaces "use" with delegation as the unit of analysis for human-AI relationships. But delegation, as I argued when I wrote about trust and delegation, involves appraisal, distribution, and coordination. When the agent handles the ostensive routine, it does distribution and coordination on the standard cases. The appraisal function, deciding whether this case is standard or exceptional, gets split between the agent's classification logic and the human's judgment. The agent classifies. The human overrides. But the human only sees the cases the agent flags. The structural context that informed the human's judgment has been replaced by the agent's filter.

DeSanctis and Poole (1994) gave us spirit and appropriation. The spirit of a technology is the designer's intent for how it should be used. Groups appropriate those structures, and that appropriation can be faithful or unfaithful. The problem with agentic AI is that the agent's spirit, its optimization criteria, its classification thresholds, its priority rankings, gets enacted at machine speed. The organization's spirit operates at human speed. By the time a human reviews the flagged exception, the agent has already classified thousands of cases as routine and processed them. The agent's outputs become inputs to the next cycle of action. The agent's classifications become how the organization thinks about priority. The agent's patterns reinforce what counts as normal. This is structuration running through the agent, and it compounds faster than any human appropriation can review.

I think the deepest problem is not that agents make mistakes on exceptions. The deepest problem is that the performative routine loses its material when the ostensive is no longer enacted by humans. As I wrote about agentic AI and organizational design, routine decisions are how junior managers build judgment. They are the ostensive practice that develops performative skill. If the agent handles the routine cases, the humans who would have learned from them lose the developmental path that teaches them to recognize exceptions. You cannot train someone to handle exceptions by showing them only exceptions. Exception handling requires the contrast with routine cases, the accumulated pattern recognition that comes from seeing hundreds of routine cases and gradually noticing which ones are not.

This is not an argument against deploying AI agents. It is an argument for understanding what you are actually deploying. You are not deploying a tool that handles the boring parts so people can focus on the interesting parts. You are deploying a structural participant that takes over the ostensive routine, which is not just the documented process but the training ground for performative judgment. The exceptions do not get smaller or simpler when the routine is automated. They get harder, because the people handling them have less context for recognizing what makes them exceptional.

{
"@context": "https://schema.org",
"@type": "Article",
"headline": "When AI Takes Over Your Routines, Who Takes Over the Exceptions?",
"author": {
"@type": "Person",
"name": "Ali Safari",
"url": "https://alisafari.space"
},
"datePublished": "2026-05-19",
"description": "AI handles the ostensive routine beautifully. But organizational value lives in the performative, where people improvise around what the process diagram never anticipated.",
"keywords": ["routine dynamics", "agentic AI", "organizational theory", "Feldman Pentland", "IS theory"]
}