Most of what protects your company from AI-powered fraud isn't technology. It's how you decide to trust a request.

Picture your controller on a Thursday afternoon. A voicemail comes in, and it's you. Your voice, the way you always say her name, a little rushed because you're stuck between meetings. A vendor payment needs to go out today, the new bank details are in her inbox, and could she please just take care of it.
She's careful. She's been with you nine years. And nothing in that message would trip her up, because nothing in it sounds wrong.
We're not telling this to scare anyone. Versions of it show up in FBI warnings now, and most leaders we talk to around the Capital Region have stopped asking whether it could happen to them. They're asking a better question. What do we change?
That's what this post is about. The short answer surprises people: most of the fix isn't technology.
Same Old Scams, Much Lower Price
Here's the odd thing about all the AI attack headlines. The attacks are mostly old. Fake invoices. Phishing emails. Somebody pretending to be the boss who needs a wire sent before five. Business email compromise has been the expensive one for years.
What changed is the price. Writing a convincing, personal email used to take a criminal some effort, and most didn't bother, which is why so much phishing was full of typos. Now it takes seconds. Research got cheap too. An AI tool can read your website, your LinkedIn page and last year's press release and work out who approves payments faster than a new hire could.
Then there's voice. A short clip of someone talking, from a webinar or a podcast or a local news segment, can be enough these days to make a passable copy. On a phone line, passable is plenty.
So, the attacks didn't get smarter so much as cheaper, faster, and better dressed.
What the FBI's numbers say
The FBI's Internet Crime Complaint Center gave AI its own section in the annual report for the first time with the 2025 data. It counted 22,364 complaints with an AI connection and more than $893 million in adjusted losses. Business email compromise alone came to about $3 billion.
Keep in mind those figures only include people who reported, and who knew AI was involved. Most victims wouldn't know. Treat them as a floor.
What Changed for a 300-Person Company
If you run a company of 50 or 300 people around the US, the plain summary is that the old warning signs are gone.
We all spent years teaching staff to look for bad spelling and strange phrasing. That advice is close to useless now. The emails read fine. They mention real projects. Sometimes they come from a real vendor's account that was quietly taken over a month earlier, so even the thread history checks out.
Mid-sized companies are also more worth the trouble than they used to be. When the research and writing were manual, criminals went after a few big targets. With AI doing that work, a 75-person manufacturer in Clifton Park is suddenly a reasonable use of an afternoon. Your org chart is on LinkedIn. Your vendors are listed on your website. That's enough to start.
And attackers have more patience now, oddly enough. When the next email costs nothing to write, they can keep a conversation going for days, answer questions, wait. Fewer chances for anyone to get a funny feeling.
Where Leadership Teams Tend to Look in the Wrong Place
Two reactions come up a lot when we talk with leadership teams. Both are understandable. Neither does much on its own.
The first is buying something with AI on the box. There's a whole wave of AI-powered security products, and some of them are good. But if your team will still send money because a familiar voice asked, no filter fixes that. We'd rather say it plainly: you may not need a new product at all.
The second is one more round of phishing training. Training still matters. It just can't be the old version where the lesson is spot the typo. People need to learn something less comfortable, that a message can look perfect, sound perfect and still be fake. Sharper eyes won't save anyone here. A rule will.
The Fixes That Work Are Kind of Boring
Most of what protects a business from AI-assisted fraud costs almost nothing. It's process. This is roughly the order we'd go in.
Call back, always, on a number you already have. Any request to change bank details or send an unusual payment gets checked by calling that person on a number from your own records. Not the number in the email. Not the one the caller reads out. The FBI tells the public the same thing: hang up, look up the organization yourself, and call them directly.
Two people on every wire over a set amount. Pick the number with your CFO and make it non-negotiable, for everyone, you included. Especially you, since the CEO's voice is the one most likely to get faked.
A code word for the leadership team. The FBI suggests families agree on a secret word to confirm it's really them on the phone. Companies can borrow the idea. If you call asking for something urgent and can't give the word, the answer is no, and nobody gets in trouble for saying it.
Phishing-resistant MFA on email. Email is still where most of this starts. Somebody steals a password? With the right kind of multi-factor authentication, that password alone gets them nowhere. Text-message codes beat nothing. They're just the weaker version.
The last one sounds soft, but don't skip it. Your people need permission to slow down, because the urgency is how these scams work. Say it at the next all-hands: no request from leadership, ever, is too urgent for a callback. They need to hear that from you, not from IT.
Ask Your Broker This Before Your Next Renewal
Most leadership teams assume their cyber policy covers this kind of fraud. Sometimes it does. Often there's a catch.
When an employee is tricked into sending money, many policies treat that differently from a hack. It might be a separate coverage with a smaller limit. It might come with conditions, like proof you verified the request in a certain way. Policies vary a lot, so we won't pretend to know what yours says.
What we'd do is call your broker with a few plain questions. Is social engineering fraud covered? Up to how much? What do we have to have done, procedure-wise, for a claim to get paid? If the answer to that last one is a callback rule you don't have yet, congratulations, you just found your first project.
We'll go through this in more depth at our November 17 event, along with the rest of what's on the agenda.
Who Shouldn't Lose Sleep Over This
Not every business needs a big security project because of AI. Say you've got 15 people, you pay vendors through a platform with its own verification, and nobody can move money alone. You're in decent shape. Get MFA right, agree on the callback rule, and you've covered most of it.
It gets more serious in the 25 to 300 employee range when serious money moves every week. Regular wires, lots of vendors, a busy finance team, leaders whose voices are all over the internet. That describes a lot of the Capital Region companies we talk to. Same for anyone in a regulated industry, where fraud can turn into a reporting obligation on top of the loss.
AI Turned Cybersecurity Into a Leadership Decision
For a long time, security was something the IT budget handled and leadership signed off on once a year. Buy the firewall, renew the antivirus, check the backups. Technical problems, technical fixes.
AI-assisted fraud is different. The weak spot is how your company decides to trust a request, and that was never a server problem. Who can approve what. How fast. Whether anyone feels allowed to push back on the boss.
Those are leadership calls. An IT provider can set up the tools, and should. But only the people running the company can decide that the CEO's voice, by itself, is no longer enough to move money. Once a leadership team says that out loud, most of the AI risk shrinks to something you can manage.
How TAG Solutions Approaches AI-Era Cybersecurity
TAG has helped businesses across Albany and the Capital Region with IT and security for more than 30 years. We've watched phishing go from laughable to very good, and the advice we give has changed with it.
Every client has an assigned vCIO. Companies that need more security leadership can also work with a vCISO, someone who owns the security program and sits with your leadership team when decisions like these get made. Behind them is the team that runs the technical side every day.
With most Capital Region clients we start with an assessment. How is email protected? How do money and access get approved? What's public about your people, and where would a convincing fake get through? You get back a short list in plain English, cheap fixes first.
We back everything with service guarantees.
Join us on November 17
We're hosting a working session for CEOs, CFOs, COOs, Presidents and owners of companies with 20 to 500 employees. No technical background needed. An FBI cybersecurity expert will talk about how AI has changed what they're seeing. We'll run a live mock cyberattack on a fictional business so you can watch it unfold, dig into the insurance gap, and you'll build a 90-day action plan before you leave. No vendor pitches. Lunch is on us, and registrants get a complimentary cyber risk assessment within five business days.
________________________________________________________________________
If your team would still send money because a familiar voice asked, that's the gap to close first.
You'll leave with a 90-day action plan built for your company, plus a complimentary cyber risk assessment within five business days.
________________________________________________________________________
Questions Business Leaders Ask Most Often
Can AI really copy my voice from a few clips online?
Probably, if there's audio of you out there. Webinars, podcasts, a TV interview, even a long voicemail greeting. The copies aren't perfect. On a phone line, with a rushed tone and a believable story, they don't need to be. Which is why telling people to listen more carefully doesn't help much. Better to just assume any voice can be faked and set things up so a phone call, by itself, never moves a dollar. A callback to a known number and a second approver take care of most of it.
Do we need to buy an AI security tool?
Maybe eventually. Almost never first. Start with the cheap stuff. Get the callback rule going. Add a second approver on payments. Email gets phishing-resistant MFA. And the leadership team agrees on a code word. Those cover the attacks we see most. After that, sure, better email filtering that uses AI to catch impersonation can earn its keep, especially if you process a lot of vendor invoices. Anyone who leads with a product before asking how your payments get approved is selling you something.
Is our cyber insurance going to cover a deepfake scam?
You'd have to check, and most people haven't. Fraud where an employee is tricked into paying often gets treated differently from a hack. Sometimes the limit is lower. Sometimes there are conditions about how you verified the request. Ask your broker straight out whether social engineering fraud is covered, for how much, and what procedures the policy expects. Write the answers down. If the policy expects a callback process, get one in place before you ever need to file.
How do I train employees when the fakes look perfect?
Mostly, stop training them to spot fakes. Train them on the rules instead. Nobody changes bank details without a callback. Nobody sends a wire alone. If the boss calls in a hurry, saying no is allowed, and it won't get anyone in trouble. Then test it a couple of times a year with a realistic fake. People remember what they've practiced way better than what they read in a policy. You want a team that doesn't have to win a guessing game, because the process catches what eyes can't.
Is the November 17 event worth the time for a CEO?
We think so, but we're biased, so here's the fair version. It's meant for owners and the C-suite at companies of 20 to 300 people. You don't need any tech background, and nobody's going to pitch you anything. An FBI cybersecurity expert talks through what they're seeing. There's a mock attack you get to watch happen to a fictional company, then insurance, then you sit down and write a 90-day plan. If you already have a callback rule, two-person approvals and a recent assessment, you might not need it. Most companies don't have all three.
________________________________________________________________________
Not ready for an event yet? Get short, practical AI advice every week. [ Get Weekly AI Tips ]



