September 26, 2026
TechCrunch's Connie Loizos wrote about commissioning an interactive digital avatar of herself, trained on her own knowledge, including her reporting on venture fraud, so that other people could talk to it. That's not a novelty demo anymore. It's a preview of a real product category: a version of you that answers questions, pitches clients, or does customer intake while you're asleep. For a business audience, the appeal is obvious. A founder's avatar could field investor questions at 2 a.m. A support lead's avatar could triage tickets before a human ever sees them. But the piece is honest about the discomfort, and that discomfort is the actual story. Once a clone of your expertise exists, who owns it, who's liable when it says something wrong, and does it stay accurate as your own views change? Those aren't hypothetical governance questions -- they're the same ones companies are already fumbling with around ordinary customer-facing AI agents. If your business is even considering an AI stand-in for a real person, whether that's a support rep or a sales lead, the Helpdesk & Support layer around it -- who reviews what it said, where the transcript lives, how a bad answer gets corrected -- matters more than how lifelike the avatar looks.
Separately, TechCrunch reported that frontier models -- referred to as Astra and Opus -- have finished codebreaking work originally tackled by Alan Turing during World War II. It's a genuinely striking benchmark: reasoning that once took elite human cryptanalysts years is now closer, in some form, to a solved problem for the right model. I think the instinct to treat this as a pure capability win undersells what it actually signals. Codebreaking is pattern-matching against a fixed, adversarial structure with a clear right answer -- which is precisely the kind of problem frontier models are best at, and precisely unlike the messy, ambiguous decisions most businesses actually need help with. The lesson for a business reader isn't 'AI can now do intelligence work.' It's that the gap between benchmark headlines and deployable reliability is still wide, and it's worth being skeptical of any vendor pitch that leans on a historic-sounding achievement to imply their model is ready for your unstructured, everyday workflow. That skepticism doesn't mean dismissing the progress -- it means asking what the model actually gets wrong on your specific, messier data before you trust it with anything that isn't graded like a puzzle.
Put the avatar story and the codebreaking story side by side and a pattern emerges. Both are examples of AI getting extremely good at producing something that looks like a correct, human-quality output -- a voice that sounds like you, a cipher that's actually solved -- while the harder questions of accountability and applicability lag behind. That's the same tension we've flagged before around agentic tools that act on a company's behalf without clear ownership of the outcome, and it's why any team evaluating AI beyond a chatbot window should be asking not 'can it do this' but 'who's responsible when it doesn't.' For internal workflows, that means building AI into processes with an audit trail rather than bolting it onto whatever's already fragile -- something Internal Tools & Admin is built around, precisely because the risk isn't the AI failing loudly, it's the AI failing quietly and convincingly.
So here's the question worth sitting with: if you could commission an AI avatar of yourself for your business today, would you actually trust it to talk to a customer unsupervised -- and if not, what exactly is missing?
Sources