The AI Risk Already Inside Your Business

There is a real argument happening about whether AI poses a danger to humanity. Serious people are on both sides of it, the evidence is contested, and I am not going to pretend I can settle it from a blog post about small business operations. It’s above my pay grade and probably above most people’s.

What I can tell you is what happened to actual companies in the last twelve months, because IBM and the Ponemon Institute went and counted.

They studied 602 breaches across 17 industries between March 2025 and February 2026. Among the organizations that suffered an AI-related breach, 92% had no proper access controls on their AI. Sixty-eight percent had no AI governance policy, which is worse than the 63% who had none the year before.

Read that twice. Almost every company that got hurt by AI this year had never set a boundary around it. The model didn’t outsmart anyone. Nobody told it what it was allowed to touch.

The thing that’s actually happening

The headline number in that report is the record global average breach cost, $4.99 million, up 12%. Let me be honest about that figure before I use it: your business is not going to lose $4.99 million. That’s an average across enterprises with legal departments and regulators and thousands of records per incident. If you run a 12-person firm, your bad day costs a great deal less than that in dollars.

The pattern is what transfers, and the pattern is specific.

Shadow AI incidents more than doubled this year, from 20% of breached organizations to 43%. Shadow AI means an employee using an AI tool the business never approved. Not a hacker. A person on your team, trying to finish something faster, pasting a client list or a contract or a spreadsheet of names into whatever tool was open in the other tab.

Those incidents cost more than the ones without shadow AI, $5.39 million against $4.63 million. More than one in five of them also drew a regulatory fine, which is the part that scales down to a small business cleanly. Privacy rules don’t have a small-company exemption. If client data leaves your control, the size of your company changes the size of the check, not whether you write one.

Meanwhile, AI-driven attacks rose 56%, and more than one in four organizations experienced one. The most common kind was deepfake impersonation, at 45% of AI attacks. That one scales down perfectly too. A convincing voice on the phone asking your bookkeeper to change wire instructions doesn’t cost an attacker much anymore.

What shadow AI looks like in a twelve-person company

Nobody thinks of it as shadow AI. It looks like this.

Your office manager pastes a client’s tax documents into a free chatbot to summarize them, because the alternative was reading forty pages.

Someone on your sales team runs your whole customer list through an AI tool to clean up formatting, and that tool’s terms say it may retain and train on submitted content.

A project manager connects an AI notetaker to every call on the calendar, including the one where you discussed a client’s financial trouble, and the transcripts live in an account nobody at your company administers.

A contractor you hired for six weeks uses their own AI subscription on your files, and when the contract ends, the data doesn’t come back.

Every one of those people was trying to do good work. None of them did anything malicious. And in most small businesses, none of them broke a rule, because there wasn’t one.

Four things to write down this month

None of this needs a consultant, a platform, or a budget. It needs about two hours and a willingness to be specific.

Name the tools that are approved. An actual list, by name, that everyone can see. Two or three sanctioned tools with the business paying for them beats a blanket ban every time. A ban doesn’t stop the work, it moves the work onto personal accounts you can’t see, which is how 43% of breached organizations ended up with a shadow AI incident in the first place.

Name what never goes into any of them. Be concrete or it won’t stick. Client financial records. Anything with a Social Security number. Signed contracts. Passwords and keys. Employee health or payroll information. Write the list for your business, not a generic one, and keep it short enough that people remember it.

Give people the sanctioned path before you close the unsanctioned one. This is the step everyone skips. If the approved tool is slower or harder to get to than the free one, you’ve written a policy that documents what people aren’t doing. Pay for the business version, turn off training on your data where the vendor offers it, and make it the easy choice.

Write down what each approved tool can reach, and check it twice a year. Which systems it connects to, what it can read, what it can change on its own, and who at your company can see the record of what it did. Access granted for one project has a way of quietly becoming permanent. The report’s phrase for the gap is access controls, and 92% of the companies that got hurt didn’t have them.

Two hours. It’s the cheapest risk work available to a small business right now, and almost nobody has done it.

The same argument, from a different direction

A few weeks ago I wrote about the three questions to ask before you let AI touch your business: what can it see, what can it change on its own, and how do you find out afterward what it did. That piece was about evaluating a vendor before you buy.

This is the same argument pointed inward. The AI already running in your business, that nobody selected and nobody scoped, deserves the same three questions. Usually it can see more than you’d guess, it can’t change much yet, and there is no record of what it did at all.

Governance is a heavy word for something this small. It’s a list of approved tools, a list of forbidden data, a sanctioned option that’s genuinely easier than the alternative, and a note of what each tool can reach. That’s the whole thing. It fits on one page.

That includes us. OwnerOS is early access, it drafts and never sends on its own, and we haven’t opened it to businesses handling regulated data because that work isn’t finished. Ask us the same three questions.

The part worth keeping

Whether AI turns out to be dangerous to humanity is a question I can’t answer and neither can most of the people answering it loudly.

Whether AI is currently leaking your client data through an account you don’t administer is a question you can answer this week, and the answer is written down in four bullet points or it isn’t written down anywhere.

Tell me what is slipping. I will tell you if we can help. A 30-minute call. We look at how you run today and whether this is worth it for you.

---

Sources

Cost of a Data Breach Report 2026, IBM and Ponemon Institute

Ten Takeaways from IBM’s 2026 Cost of a Data Breach Report, Baker Donelson, 30 July 2026