Deepfakes Are Coming for Your Invoices and Reviews — Provenance Is Now a Small-Business Survival Skill

For two years, the AI conversation for small businesses has been about production — AI helps you write, design, and market faster. The next phase is the mirror image: AI lets anyone produce convincing fake content about your business. An invoice that looks exactly like your supplier's. A phone call in your bookkeeper's voice. A one-star review describing a rude technician who was never in your employ.

In early October, Google DeepMind released a widely-watched explainer on the science of watermarking AI-generated content — covering everything from deepfakes to synthetic DNA data. The same week, security researchers were documenting voice-cloning scams and fabricated personas aimed at real organizations. For a small business, the takeaway isn't the technology. It's that "seeing is believing" is quietly expiring, and the defenses that actually matter cost close to nothing to set up.

The Four Fraud Patterns to Expect

Fake vendor invoices. A realistic PDF invoice from a supplier you genuinely use, with "updated remittance details." It arrives during a busy week, references a real PO number scraped from an email thread, and gets paid. Small businesses are the ideal target — no AP department, no segregation of duties, owner approvals done at 11pm on a phone.

Cloned voice wire requests. Voice cloning needs a short sample — a webinar recording, a podcast appearance, even a voicemail greeting. Then comes the urgent call or voice note: "It's me, I'm boarding a flight, wire the deposit before end of day." The voice is right. The urgency is manufactured.

Fabricated reviews. A one-star review describing shoddy work your crew never did, written in fluent, specific, believable prose. AI has removed the cost of writing dozens of them. Platforms move slowly on fake-review disputes, and disgruntled ex-employees or shady competitors now have a cheap weapon.

Fake personas. A "customer" with a polished LinkedIn profile and an AI-generated headshot, demanding refunds or threatening public complaints. Personas used to take months to build; they now take minutes.

The Zero-Budget Defense Playbook

1. Money movement runs on codewords and callbacks

Any request to move money that arrives by email, text, or voice message gets verified by calling back a number already on file — never a number included in the request itself.

Add a second layer: a verification codeword known only to people authorized to request payments. Any "urgent" request without it is treated as hostile until verified. No exceptions for deadlines — the urgency is the scam.

2. Invoice verification is a checklist, not a vibe

Keep a vendor master list with the bank details you have on file for each supplier. Any invoice asking to change remittance details gets verified through a known channel — the phone number from your own records, not the one in the invoice. A legitimate vendor will not mind the call. A fraudster cannot survive it.

3. Build your review-dispute file before you need it

After every job, keep dated records — photos, signed work orders, confirmation emails. When a fabricated review appears, you're assembling evidence instead of reconstructing memory from six months ago.

Respond publicly, briefly, and factually. A heated owner reply is the screenshot that travels. "We have no record of this service, and we've disputed the review with the platform" is stronger than a paragraph of outrage.

4. Treat inbound media as a claim, not evidence

A "customer" sends a video of damage your crew allegedly caused? Before the refund offer, ask for corroborating details an AI can't generate: the invoice number, the crew lead's name, the exact appointment time. A real customer with a real complaint answers those questions immediately. A generated complaint stumbles on specifics.

Where stakes are high, ask for the original file and check its metadata. This isn't bulletproof, and it doesn't need to be — the goal is to raise the cost of the scam past what it's worth.

5. Watermark and document your own genuine assets

The flip side of provenance is making your real content provably yours. Turn on content credentials or watermarking where your tools support it. Keep an internal asset log: photos of completed work, official documents, your team's actual bylines. When a fake appears, you can produce the genuine item and its history instead of arguing against a forgery from nothing.

Trust Is Becoming a Technical Property

The instinct is to file all this under paranoia. Resist it. This is infrastructure, the same way double-entry bookkeeping is infrastructure. The businesses that absorb this shift won't be the ones with enterprise fraud-detection budgets — they'll be the ones with boring verification habits written down in advance and practiced by everyone, including the owner.

Set it up before the incident, not after. The codeword takes five minutes. The vendor callback rule is one sentence in your AP process. The asset log is a folder. Provenance used to be a nice-to-have; it's now part of staying in business.

If AI agents are handling parts of your operations, verification habits are part of ownership — The AI Agent Owner's Playbook covers how to keep automated systems accountable to the business they serve.