Nearly seven in ten companies say at least some of their AI projects ran over budget during the past year. One-third say overruns happened mostly or always.

Those figures come from CFO Dive's reporting on a new WitnessAI study of 300 U.S. executives at organizations with at least 1,000 employees. The study was released on July 22. It is vendor-sponsored research, so the percentages deserve the usual caution. The pattern is still hard to dismiss.

Only 9% of respondents said more than three-quarters of their AI initiatives had produced a measurable financial return. Yet 70% said their companies were already using or piloting AI agents capable of autonomous action.

Adoption is moving faster than accountability.

The underlying WitnessAI release found that fewer than one in five respondents — 18% — had formally inventoried all agents and secured approval from the security team. Nearly half of organizations with deployed agents had continuous monitoring. Twelve percent reported little or no oversight.

The financial consequences were not theoretical. Eighty-six percent said they had investigated at least one AI-related security or operational incident in the previous 12 months. Forty-three percent put the combined cost of those incidents at $2 million or more. Seventeen percent estimated costs between $10 million and $24.9 million.

The governance bill was substantial too. More than half of respondents allocated 21% to 45% of their AI budgets to risk management and governance. Even then, ownership remained blurred. Only 46% said the CFO actively modeled AI-specific risk and return. Another 38% said finance reviewed AI spending without separately modeling the exposure.

This is what happens when AI is bought as a category rather than managed as a workflow. Bills sit with procurement. Usage sits with IT. Time savings appear in departmental reports. Risk sits with security. No single person owns the complete economic result.

The practical response is smaller than an enterprise transformation program. Give each AI workflow one owner, one baseline and one outcome. Measure the cost per completed task, the human time actually removed, the error rate and the actions that still need approval. If those figures cannot be seen in one place, the return cannot be managed.

That is also the right way to use Viktor. Viktor lives inside Slack and Microsoft Teams. You @mention it in a thread the same way you would ask a colleague. The output — a PDF, a report, a task created in your CRM, an email drafted in Gmail — lands where it should land.

A finance manager could ask Viktor to collect AI vendor invoices, summarize departmental usage and produce a weekly exception report. An operations manager could give it one repetitive workflow, require a human approval before any sensitive action, and compare the completed work against the old baseline. Both uses create evidence. Neither depends on a vague promise of productivity.

The survey's warning is not that businesses should stop using AI agents. It is that speed without ownership creates invisible cost. The companies that make AI pay will know who owns each workflow, what it costs and which result justifies keeping it.

Viktor is SOC 2 certified, GDPR aligned, CCPA compliant, and CASA Tier 3 certified. Your credentials never touch the AI — they are stored in an encrypted vault and injected at runtime. Your data never trains a model (contractual agreements with OpenAI, Anthropic, and Google). Every sensitive action waits for your approval in Slack before it executes. Full security details: viktor.com/security

You get $100 of free credits to begin. No time limit, no commitment. That's enough to do real work and see what Viktor can actually do before you spend a penny. There's also $50 off your first bill. You must use this exact link to receive both benefits.

Get started with Viktor at https://AIThatDelivers.com.

Disclosure: Some links in this article are affiliate links. If you choose to get started with Viktor using the links provided, I may receive a commission — at no additional cost to you. I only recommend tools I use and believe in.