Celonis, a company that has long sold software for reading the records that business systems keep, argues that AI agents at work need a layer of context: what role a task plays in a bigger decision, and which limits surround it. Its chief scientist, Devavrat Shah, makes the case that without it, agents work in a vacuum. The argument matters well beyond large companies, because every workplace runs on more than its written rules.
An agent without context is expensive in two ways
The first way is cost. An agent that does not know your business can still work out what it needs, the way a new hire can, by asking questions until the picture is complete. A new hire who interrupts everyone all day is a cost, and so is an agent that spends a large amount of computing power on one routine task.
The second way is error, and it is the more serious one. When the agent misses a constraint, the result is a wrong answer delivered confidently. Shah's comparison is a restaurant tip that turns out mediocre versus a financial forecast that sinks a company. Most work sits between those two, and the mistakes are small but frequent.
For example, a distributor pays 1,000 supplier invoices a month. The written rule says an invoice is paid only when it matches a purchase order and a delivery receipt. In practice, one supplier delivers on weekends when nobody logs the receipt, so the warehouse lead emails the clerk and the clerk approves the invoice anyway. If one invoice in ten takes a detour like that, an agent working from the manual stops or pays wrongly on about 100 invoices a month. At 15 minutes each for a person to untangle, that is 25 hours, about three working days, and it is the work the agent was bought to remove.
The detours are already on record
The workaround itself lives in an email, but its effect shows up in the system. Each invoice leaves a trail: received, matched, blocked, unblocked three days later by a named person. Across thousands of invoices, the trail shows how work runs in practice, including the routes nobody drew. This is what Celonis built its business on, and its Context Model gives agents that picture of the work, with limits on what each agent may do.
The data does not have to be tidy for this to work. Shah's view is that the records can be incomplete and messy and still be useful, because the agent is reading patterns, not a perfect file. What the records cannot give is the reason. They show that invoices from one supplier pause and then move on. They do not say that the pause is safe.
The person who knows the reasons becomes the reviewer
That gap is where the human job goes. Someone has to say which detours are legitimate habits and which are old mistakes that nobody removed. The clerk who has handled that supplier for six years can do this in an afternoon, and an agent cannot do it at all. If agents take the repetitive work, as Shah expects, then knowing the exceptions is what makes a person the one who approves what the agent proposes.
You do not need a software platform to start. Take the last 20 cases of one job, such as invoices, refund requests or rota changes. Mark every case that did not follow the written steps, and write one line beside each on why. That list is the briefing an agent needs, and it also shows which parts of the job are routine enough to hand over first. Pick the hardest step, give the agent that, and measure whether it saves time before widening it.
An agent that follows the manual is the white strike in software form. One that has seen how the work really goes, and has a person to ask about the rest, can do real work.