Is Controlling AI Agents Impossible?

Is Controlling AI Agents Impossible?

An accounting manager, HR director, data analyst — all surface the same problem:

“Yesterday, I pasted an email draft into Claude and asked it to finish writing. The same week, our product team wrote a prompt to auto-scan specific features. A utility is running to categorize customer emails, but there’s no spec. HR has Claude Plus on a few personal accounts to summarize CVs. All of these touch data several times a week, with no output metrics.”

This is not an accident. This is the accumulation of small decisions. Each decision is rational on its own, but together they become a mass.

How Many AI Decisions Are Made in a Week?

Consider a typical week at a mid-sized firm:

By week’s end, opened channels:

None has defined responsibility. None has defined boundaries. None has encrypted, segregated data. None has centralized metrics.

Take this week and multiply by the next. Four weeks: 8 new tools, 3 pilot projects, maybe 15-20 hidden account activities. Three months: 50+ AI touchpoints, none explicitly organized.

Why Does This Slip Past Notice?

Three reasons:

1. Rational at Micro Scale

Writing a webhook on Monday works for that Monday. By week’s end, nobody knows if it succeeded. But during that week, it looks like it’s working. The decision-maker doesn’t see the outcome. But by next week, someone else is using it. A third person expands it. The first decision-maker has moved on.

2. Departmental Decisions, Organizational Impact

Marketing tries a SaaS. IT doesn’t know. Finance doesn’t know. Three months later, Marketing has sent data from 3,000 customers to this cloud and no one can answer: was it encrypted? Backed up? Who signed the contract?

By now, these micro-decisions have become a macro data problem.

3. No Instinct for “Humans Are Using It, So Risk Exists”

HR pastes CVs into Claude, gets summaries in two minutes — this looks routine. But it’s happening 50 times a day. External account, no password, no backup, no audit trail. The HR manager thinks, “This is how modern work moves.” The scope of risk is invisible.

Organizational Blindness and Loss of Audit Trail

After three months of accumulation:

This is not “Shadow AI.” This is “Unaudited Decision System.” Human decisions, accelerated by AI, leaving no trail.

To grasp the organizational risk, consider a scenario:

The marketing SaaS gets breached. Customer data leaks. The company faces liability. The auditor asks for explanation. What do you say? “Over three months, a few people used it for testing. I don’t know the exact count, who signed the contract, if data was encrypted, if backups were taken. None of it is documented.”

That answer is a legal risk.

Starting Point

A single action:

Observe the inside. For one week, collect every new AI decision. “This week, how many people accessed how many AI systems, for what?” Start by asking, not punishing.

Most organizations skip this simple inventory. Because the answer unsettles most managers.

The shadow AI problem reveals a management truth: speed is preferred, control is secondary. When that stays true, risk accumulation accelerates.

Controlling the problem means changing how management operates. But most organizations wait for a crisis first.

For readers: In this post, I discussed the scale of shadow AI. Organizational blindness is not a structural issue — it’s a decision issue. If AI agents in your organization go untracked, observe and count. No single metric is inherently risky, but when combined without user consent, data begins to leak.