Nobody installs AI and walks away

Almost every company treats AI the way it treats software: define it, build it, install it, done. Not even critical software works that way; the ERP and payroll live with change control, monitoring and support. The mistake is assuming AI needs less of that discipline, when it needs more.

A payroll system calculates the same way in year three as in year one, because it executes fixed instructions. An AI agent constantly interprets vendors, documents, policies and systems that change, and if you stop watching it, it drifts. It won't crash. It will get worse slowly, very sure of itself, and you will notice once the damage is already in the books.

AI degrades because it judges a world that changes

A traditional system applies its rules until someone changes them on purpose. An agent reasons about things that change on their own. The vendor sends the invoice in a new format. A policy gets updated. A screen in the ERP moves. A new law comes out. The agent doesn't find out and keeps applying the pattern it learned, with the same confidence as always.

The industry calls this drift (concept drift or behavioral drift), and it is dangerous because nothing seems to break. The agent looks healthy while its decisions rot little by little.

A silent failure builds up in the dark

When a server goes down, someone gets an alert at two in the morning. An agent that drifts delivers answers that are plausible and wrong, which sail straight through the checks a crash would have tripped. If monitoring wasn't designed into the operation, nobody finds out.

The error shows up when it adds up: exceptions rise, a reconciliation stops balancing, a compliance report reads oddly. By then the agent has been deciding badly for weeks. You are not risking an outage you see right away, but an invisible erosion you catch late.

The day-two work nobody budgets for

Running an agent in production already has a name: AgentOps, which Microsoft and IBM formalized in 2026 as the lifecycle management of AI in production. It includes:

• continuous monitoring of what the agent is really doing;

• early drift detection;

• instruction adjustments and retraining when the world changes;

• governing it again when policies and regulation change;

• incident response when it steps outside policy;

• tuning of cost and tokens as consumption grows.

None of that is done once. Installation is where the work begins. A traditional system fails loudly and you notice right away; an agent fails quietly and you notice once the damage has added up.

Why almost nobody can build that team

Operating an agent well is a specialized capability that never rests, and few companies can justify keeping it in-house. Gartner recommends budgeting up to 10 times the price of an AI tool for the work around it: data, monitoring, retraining and change management.

So companies end up in one of two bad places. Either they don't operate the agent and it degrades silently, or they build a full team to run it, the hidden cost that dwarfs the license. The third path is for someone to operate it for you, all the time, and answer for keeping it good, not just for getting it running.

It is what we ask of a digital collaborator before it deserves the name: having a role, having a manager, leaving evidence of everything, making that evidence auditable, and having someone who answers when it fails. The first four make drift visible; the fifth makes sure someone fixes it. At Yunt, every correction from the manager goes into the role's manual and is not repeated, and behind every collaborator there is a team that answers for it, like an airline's mechanics on shift.

Change the question

Stop asking how fast you can install AI and ask who answers for operating it after launch. A launch is a moment; an operation is a commitment. The value of AI is defended every day after it goes live, or it erodes without anyone seeing it.

The model that wins is the one where installing and operating are the same responsibility, in the hands of people whose job is to keep the agent right long after the demo is forgotten. That is the difference between an AI you launched and one you can depend on a year from now.

Sources

• Microsoft and IBM (2026), AgentOps.

• Gartner, associated costs of AI tools.