What Unmonitored Agents Actually Cost (And Why the Fix Starts at $0)

Published June 23, 2026

An unmonitored agent looks free. That's the whole problem. It doesn't show up on an invoice, it has no line item, and it sits in production doing something at a cost of exactly zero dollars — right up until the afternoon it costs all of them at once.

Liabilities work like that. They're free until they aren't.

The number nobody put on the agent's budget

Start with the baseline. IBM's 2025 Cost of a Data Breach Report — the number CFOs already trust — puts the global average breach at $4.44 million. That figure actually fell 9% year over year, the first decline in five years, because AI-powered defenses are containing breaches faster than before.

Then the catch: that progress does not extend to the AI you aren't watching. When a high level of "shadow AI" — tools and agents running without approval or oversight — is involved in a breach, IBM found it adds an extra $670,000 to the bill. And the organizations taking that hit weren't unlucky; they were uninstrumented. 97% of organizations that suffered an AI-related security incident lacked proper AI access controls, and 63% had no AI governance policies in place at all.

Read those two numbers together. The premium isn't charged for having AI. It's charged for having AI you can't see.

"Detection time: never"

Here's the lever underneath the dollar figure. The same report clocks the mean time to identify and contain a breach at 241 days — and the entire cost curve bends on that number. The longer something runs undetected, the more it costs. Every time.

Now ask the uncomfortable question about your own fleet: for an agent that emits no events, sends no signals, and appears on no dashboard, what is the detection time?

It isn't 241 days. It's never. You don't detect what you don't measure. An unmonitored agent can degrade, drift, hallucinate, or quietly route customer data through a third-party tool call for as long as it stays unmonitored — which is to say, indefinitely. The breach clock doesn't start until something external trips it: a customer complaint, a regulator's letter, a reporter's email. By then you're not measuring detection time. You're measuring how long the damage compounded before anyone noticed.

The slower, quieter cost

Not every unmonitored agent ends in a breach. Most end in something less dramatic and just as expensive: cancellation.

Gartner predicts that over 40% of agentic AI projects will be scrapped by the end of 2027 — and the three reasons it names are escalating costs, unclear business value, and inadequate risk controls. That last one is the silent killer. An agent project rarely dies in a single catastrophe. It dies in a budget review, when someone senior asks "what is this thing actually doing, and what is it costing us?" — and the team that built it can't answer with anything more rigorous than a shrug.

You cannot defend a line item you cannot measure. The agents that survive the 2027 culling will be the ones whose owners can show, on demand, exactly what they did and what it was worth. Monitoring isn't overhead on that project. It's the evidence that keeps the project alive.

What the fix costs: $0

Here's the part that should irritate every CFO who's ever been quoted a six-figure governance platform. The first and most valuable move costs nothing.

VeriSwarm's Gate runs on a free tier: unlimited event ingestion, 5,000 trust decisions a day, up to ten agents. You instrument your agents to emit events, and they stop being invisible. Every agent that sends a single event shows up named, counted, and trust-scored — identity, risk, reliability, and behavior tracked continuously instead of assumed. Each one is sorted into a policy tier — allow, review, or deny — that shifts automatically as its behavior shifts.

That's the whole trick. It converts "detection time: never" into "detection time: the next event." The liability that appeared on no balance sheet becomes a monitored asset on a dashboard — for the price of zero dollars and an afternoon with the SDK.

The rest of the ladder is just insurance math

When you outgrow free, the pricing follows the same logic a CFO applies to every other risk: you buy the control that's cheaper than the incident it prevents.

Pro, at $49 a month, raises the ceiling to 100,000 decisions a day and adds portable identity credentials and intelligent LLM routing. Max, at $299 a month, adds the controls that turn "we think it's fine" into "we can prove what it did": PII tokenization that strips sensitive data before it leaves your infrastructure, a kill switch that actually stops an agent on command, and an immutable, hash-chained audit ledger that produces exactly the evidence a regulator or an auditor will ask for.

Run the comparison the way you'd run any insurance decision. One avoided incident with that shadow-AI premium attached — $670,000 over baseline — pays for roughly 187 years of Max. The free tier costs less than the meeting you'd schedule to decide whether you can afford it.

The asset on the other side of the ledger

Unmonitored agents are a liability precisely because nothing is measuring them. The cost is real; it's just deferred and unbilled. A scored agent is the opposite — a known quantity, with a behavioral record, a risk posture you can read, and an audit trail you can hand to someone paid to doubt you. Same agent. Different side of the balance sheet. The only variable is whether you're watching it.

The cheapest insurance against a seven-figure number is a zero-dollar one. Start on the free tier at veriswarm.ai, instrument one agent with the VeriSwarm SDK, and watch the thing you weren't tracking turn into the thing you can.