AI Watermarks and the End of Document Trust

[00:00:02] Your clients have been wondering whether a person actually wrote what you sent them. They were probably right to wonder. What's different is that they can check, and the check will tell them nothing useful at all. This is the Business of Tech. I'm Dave Sobel. That check applies to almost every document that crosses your desk. We'll start with Anthropic. The company announced that text and images produced by Claude will carry invisible watermarks,

[00:00:30] patterns embedded directly into the words themselves. The rollout begins with new models. The stated reason is the European Union's AI Act, which requires that machine-generated content be identifiable as machine-generated. This is not a research preview or an opt-in setting. It is the default behavior of the product, and Google already does the same in Gemini. OpenAI hasn't shipped it yet, so not universal, but the obligation driving it covers all of them.

[00:01:00] Now the two details that matter more than the announcement. Anthropic intends to release detection tools to third parties, so the ability to test a document doesn't stay inside the company that made the model. And the mark is durable, but not indestructible. Hold on to that a second. It comes back. Then look at what's already being produced under no human supervision at all. A company called BlazeHive put out a release, their own announcement about their own product, so weigh it accordingly,

[00:01:29] claiming that their AI agent has autonomously written and published content that now holds more than 500 keywords in Google's top three results, and 1500 on page one, inside of five months. The number worth noticing is not the ranking. It is the own description of the pipeline. Researched, written, humanized, and published by the agent, with, their words, no writers, editors, or prompts.

[00:01:56] Think about that third verb, humanized. Their pipeline includes an automated pass that checks the output against 30,000 documented AI writing patterns before it goes live. Stripping the machine fingerprint is not a workaround somebody invented. It's a product feature. It costs $99 a month. So, a durable mark arriving by default, detection tools going public, and content already being published with no human in the loop by design.

[00:02:24] Everything about that sounds like it finally turns a suspicion into a fact. It does something closer to the opposite, and the reason is arithmetic. If you're listening to this and haven't hit follow yet, on Apple Podcasts, search business of tech. It takes five seconds, and you'll get the next episode automatically. Managing security across a dozen client tools is eating your margins and your team's time.

[00:02:53] OpenText secure cloud bundles let you consolidate protection across endpoint, cloud, and identity into a single conversation with your clients. Simpler renewals, better coverage, healthier margins. Learn more at cybersecurity.opentext.com A test is only worth running if it can come back negative. And the economics of the last six months have made sure this one almost never will.

[00:03:22] Start with what business is actually willing to pay for the best available AI. The spend tracking firm Ramp looked at where corporate AI dollars actually go, and found Anthropics premium model, Fable 5, taking about 6% of token purchases at those companies, while accounting for roughly 11% of the spending, because it costs about twice what the tier below it does. Ramp's economist reads that as a ceiling on what companies will pay.

[00:03:51] Their sample skews tech, so the real number is likely lower. That is the market answering a question out loud. Offered the frontier at a premium, most buyers decline. Capability is not what they are short of. Now watch what they bought instead. Alibaba released a 27 billion parameter open model that runs frontier class reasoning on a high-end laptop, with no cloud service involved at all. Nothing metered, nothing logged, nothing sent anywhere.

[00:04:20] That's how the collapse happens. The cost of producing plausible professional writing fell to roughly nothing. And it fell everywhere at once. Which means the volume of AI-touched documents did not rise. It went universal. And the mark attaches where the model chooses the words. Handed a document you wrote and asked for light edits, and there is almost nothing for it to attach to. So it does not detect AI assistants.

[00:04:46] It detects who produced the first draft, which is now the free part of the job. You've not learned whether the work is any good. You've learned who typed first. Now go back to that durability detail. Anthropics own guidance is specific. Light editing probably won't strip the mark, but a complete rewrite, where every word gets replaced, will. That sounds like a high bar. It is for a person.

[00:05:13] Hand the draft to a different model and ask it to rewrite, and every word gets replaced by definition. So the mark still sorts documents into two piles, just not the ones anyone thought of. Marked, the person who used the tool and handed over what it made. Clean, the person who ran it through a second pass. And the rewriting pass caused the same nothing as the first draft did, where it never touches a server because it ran on a laptop. The test punishes the honest and clears the careful.

[00:05:41] It is a metal detector at a door that everybody now walks through carrying nothing because the people worth catching learned to leave the metal at home. It cannot see the only thing that actually varies. Which brings us to the number underneath it all. Enrock Security, a firm that sells AI governance, so weigh it accordingly, watched 4800 business users and 139,000 real AI interactions.

[00:06:06] Active use is up to 31% of eligible employees, nearly double last quarter. And of the people using these tools, about 5% are getting meaningful productivity out of them. 10 times their figure from 3 months ago, climbing fast. But it still means 19 out of every 20 people using AI work are producing output that isn't doing much. The watermark registers the thing that is now universal. It cannot register the thing that stayed rare.

[00:06:33] Which stops being a point about detection the moment you remember what your shop hands clients every week. So here's where it shows up in your business. Your client has wondered. When the security assessment came back, or the documentation, or the quarterly review deck, somewhere in there a client looked at a paragraph and had a thought about where it came from. They were probably right to have it. And they let it go because raising it meant accusing you of something they couldn't demonstrate over a document they'd paid for.

[00:07:02] That question has been sitting in the room for a little while. What changes now is that it comes with a test attached, and asking it stops being an accusation and starts being due diligence. And you can't answer it with the document. That's the part of thinking about. Whatever comes back, marked, clean, inconclusive, none of it tells your client what they actually want to know. Which is whether anyone competent looked at this before it reached them. The artifacts stop being able to carry that answer. Only a person can.

[00:07:31] Which is why the two things happening in the channel right now are the same thing. Channel Dive reported that IT service providers are struggling to deploy the AI they are already selling. Not because the tools are unavailable, but because the engineering talent and the specific expertise are not there. And the vendors who spent the last two years funding partner enablement are stepping back from it. Meanwhile, the providers who can actually integrate this work are capturing AI services revenue that is growing fast. Read those all together and the picture is not a skills gap.

[00:08:01] It's a widening one. With the subsidy that used to close it being withdrawn. Then the second one. Information Week, working from Gartner's research, reports that agent projects fail on process. Organizations deploying agents into workflows nobody ever wrote down. And the failure showing up as a broken business process rather than a broken model. The agent did what it was told. Nobody could say what it should have been told.

[00:08:29] Both of those land on the same asset and it isn't the tooling. It's a provider who can state what happens, in what order, under whose name. And who is still there, it doesn't. It sounds like a documentation project right up until you notice the only part of your delivery a competitor can't download. The MSPs getting ahead in security aren't adding more tools. They're getting the work off their plate.

[00:08:55] Guards consolidates the stack, endpoint, email, identity. And then puts an autonomous analyst on top of it. Triaging the alerts, correlating the signals, drafting the client reporting. Automatically. It's purpose built for MSPs protecting S&B clients, month to month. Real SecOps without hiring a SecOps team. Start at guards.com. That's G-U-A-R-D-Z dot com.

[00:09:27] Why do we care? Because the thing that used to separate providers, being able to produce a good looking assessment, a clean run book, a thorough review, just stopped separating anybody. Every competitor you have can generate that document tonight for nothing. And the market has been handed a test that proves it. What's left to compete on is a written review standard and somebody's name on it, which is the one thing in your delivery that can't be downloaded, matched in a quarter, or undercut by a firm that bought the same tooling you did.

[00:09:58] So what to consider? Write the standard for one deliverable, not for the shop. Pick the document you produce most often, the security assessment, the monthly report, whatever ships on the highest volume. And write down what actually happens to it before it goes out. What generates the first draft, who reads it, what they check against, and what they're specifically looking for. One deliverable, one page. Doing this for the whole business is a project nobody finishes. Doing it for one is an afternoon.

[00:10:26] And it tells you immediately whether you have a review process or just a habit. Find out whether your review would survive being described to a client. Read the page you just wrote, as though a prospect is holding it. If the honest version is, a senior tech looks it over when there's time, you've found the actual gap. And it's not a tooling gap. It's that the step exists in someone's head and varies within the week. That's fixable, but only once it's visible.

[00:10:56] The providers capturing AI services revenue right now are the ones who can describe their process without improvising. Put the standard in the sales conversation before a client puts the test in the room. This is a competitive instrument, not a compliance document. Bringing it into a renewal or a prospect meeting is the reason to choose you. Here's what happens to your deliverables. Here is who signs off. Here's what they catch. If you've defined the criteria the next provider gets measured against.

[00:11:25] The competitor who's never written it down then has to answer a question you wrote. If this trend continues within 12 to 18 months, the differentiator in a competitive services bid stops being what you produce and becomes whether you can hand it over a written account of who reviews it and against what. And the providers who never wrote one will be pitching artifacts into a room that's already stopped believing artifacts mean anything. This is the business of tech.

[00:11:55] What would you fix in your business with the right playbook? Small Biz Thoughts members get a library of templates and operational resources, recorded member calls, and classes through IT Service Provider University. Operational education built specifically for independent MSPs. Start at smallbizthoughts.org Interested in advertising? Head to mspradio.com slash engage.

[00:12:23] The Business of Tech is written and produced by me, Dave Sobel, under ethics guidelines posted at businessof.tech. Thanks for listening. I'll see you on the next episode. Proud member of the MSP Radio Network.