Why Aiva exists, the case for the clone layer.

There's a question, in every field, that only one or two people in your network would have a really good answer to. The kind of question where you don't want a framework, you want someone's judgment. You want them to tell you not what the textbook says, but what they did the last time they were in this exact situation, and why they'd do it differently this time.

Those people are scarce. They're expensive. They're already booked. Their wisdom doesn't scale, and most of it dies with them.

The current wave of AI doesn't solve this. It produces fluent, confident answers, but opinion-free, voice-free, accountability-free ones. The "AI" answer is everyone's average. The expert answer is one specific person's read of a specific situation, drawn from years of being right and being wrong about it.

That's the gap Aiva is built to close.

The category is resolving into three layers

Look at the AI products that have actually shipped in the last 18 months. They cluster into two well-funded categories, and one wide-open one.

Agents (Cursor, Cognition, Replit, the new wave of coding and ops agents) are about operational speed. They execute. They write code, write copy, push tasks across a pipeline. Their value is reliability at 100× the speed of the most experienced practitioner. Mature category, lots of capital, getting better fast.

Company brains (Glean, Notion AI, Deel's recently-announced brain, Paperclip) are about context. They consolidate your org's institutional knowledge so the company stops re-asking itself the same questions. Also mature, also well-funded, also already useful.

The third layer, what we'd call clones, is where you put the named, accountable expert judgment that the other two can't fake. It's the thing your CFO asks for sign-off on. The thing your design partner expects when she invites a clone into her workspace. The thing a procurement team requires before they let a tool make a $200K-affecting recommendation.

That layer doesn't exist yet. It's the unfilled gap. That's what we're building.

What LLMs structurally can't fake

The case for clones isn't "LLMs aren't good enough yet." LLMs are very good and getting better. The case is that there are four shapes of judgment that LLMs can't access from their training data, no matter how much of it there is.

1. Conviction without data

The strongest take you have on a topic in your field is the one you can't fully justify with citations. It's the pattern you've seen forty times and now recognise in five seconds. The reason you say "no, that won't work" without being able to write a deck-friendly paragraph explaining why. LLMs interpolate within the training corpus; they can't have conviction that isn't grounded in something in the corpus. Your conviction is.

2. Failure-pattern recognition

You know what doesn't work because you've watched it not work, sometimes in your own company, sometimes when you've been hired to clean up. That's a different kind of memory than the public web has. Most failure modes never get written up at all, and the ones that do get hagiographied into "lessons learned" that wash out the actual texture of the failure.

3. Trust networks

When you make a recommendation, part of what you're really saying is "I'd stake my reputation on this, and you know enough about me to weight that appropriately." An anonymous AI can't stake anything. A named clone (built off a person who has a public professional record) can. That's what makes it referrable in a way pure LLM output never will be.

4. Trade-off explicit-ness

The hardest thing about good advice is naming what you're willing to give up. "Go aggressive on outbound. Accept the churn." "Hire a generalist now. You'll regret missing the window. Accept the messy onboarding." Generic AI hedges all the trade-offs out of every answer. A clone can hold an opinion about what to sacrifice because the person it was built on has actually had to sacrifice things.

What we're shipping

The product has three surfaces. They share the same underlying clone, the same compiled system prompt, the same knowledge document, the same voice. What changes is who's interacting with it and how.

1. Clone Studio

Where creators build. Forty-five minutes of voice interview, a few PDFs and podcast appearances dropped in, a benchmark calibration pass, and you've got a working clone of your judgment in an afternoon. The full walkthrough is its own post.

2. The marketplace

Where consumers come to find expertise. Solopreneurs and growing teams browse clones, add them to their account, chat for one-off questions, or build cross-functional "teams" of clones who collaborate on a project. Royalties accrue to the creator whenever a paying subscriber engages.

The marketplace launches soon. Every clone built in Studio is automatically in it unless the creator opts out, so the supply that's accumulating now is the supply that's there on launch day.

3. MCP + integrations

Where the clone meets the work. Once your clone is built, you can plug it into the tools you already use: Claude Code, Cursor, Claude Desktop, Hermes Agent, anything that speaks MCP. Ask for your own judgment from inside the work you're already doing. Push back when it's wrong; the correction folds back into the next compile.

4. Enterprise (in design partner)

For organisations who want their best people's judgment available across the whole company. A Series C might have one Chief of Staff with a calibrated judgment on every cross-functional sign-off. Cloning her means the rest of the org can move at her pace without the bottleneck, and the sign-off is still hers, with her rubric, her name on it. Currently with a handful of design partners; productisation lands at Phase 2.

The clone doesn't die after week one

A clone built once is a snapshot. The thing that's interesting, what makes this a real product rather than a one-shot art piece, is the four feedback loops that keep it sharpening:

• Daily prompts. The platform surfaces one question a day that targets a gap in your current clone. Sixty seconds to answer.
• Supplemental deep-dives. A coverage matrix shows which of your stated expertise areas are thin. You top up one area at a time.
• In-tool corrections. When your clone says something wrong in Claude Code or in a marketplace chat, you push back. The correction lands in your revisions ledger.
• Auto-ingest. Connect your RSS, podcast, Substack. New content reads into the clone automatically.

All four channels write to the same pending-changes queue. Click "update your clone" and the next version recompiles with everything new folded in.

This is the under-appreciated mechanic. Export gives you a snapshot artifact. BYOK gives you free inference on that snapshot. Neither gives you a clone that keeps getting better. That part requires the platform, and it's the part most creators tell us they didn't realise they wanted until they had it.

Why now

Three things converged in the last 18 months that made Aiva possible. None of them on their own was enough.

1. Voice cloning got good enough that the clone can sound like you, not like a slightly-uncanny TTS. The interview is voice-based because the same audio that captures your thinking also trains the voice model. Two birds.
2. LLMs got good enough at instruction-following that a long, layered system prompt + a structured knowledge document can hold a stable persona across substantive conversations. Five years ago this wasn't a product; the model would drift within a few turns.
3. MCP shipped. The Model Context Protocol means clones can live inside the tools creators already use, they're not stuck on our domain. That's the thing that makes the personal-use moment work, and it's only ~6 months old.

The product wasn't possible to ship cleanly in 2023. It is now. The marketplace launches soon. The bench fills now.

What we're betting on

A few things, explicitly, because we'd rather you push back on the thesis than smile politely at the pitch.

1. Named accountability becomes a procurement requirement. Enterprise buyers will, within 18 months, require that any AI making material recommendations is traceable to a named human. Anonymous agents won't pass the legal/compliance gate. The clones layer is what passes it.
2. The "Aiva marketplace" beats "I built my own AI agent" for most experts. Most creators don't want to ship a product company. They want to scale their advice beyond the dozen people they can talk to in a week. We make that infrastructure invisible.
3. The four feedback loops compound. Every push-back, every daily prompt, every supplemental deep-dive makes the clone marginally more valuable. Two months in, three months in, six months in, the cumulative effect is more than the sum of the updates. We've watched this start to happen on Amy's clone already.
4. "Trust is a named product." Procurement and board reporting won't tolerate anonymous decision support no matter how capable. The clones layer turns AI-derived advice into something with a signature on it.

Where we are right now

Studio is open. Sign-up is no longer invite-only, you can build a clone today. The marketplace launches soon; every clone built between now and then is included by default on launch day, no migration. MCP is live for personal use across all the major editor clients.

If you're a consultant, operator, strategist, engineer, designer, marketer, anyone whose 1:1 advice gets the same shape of questions over and over, we'd love to have you on the bench. The interview is forty-five minutes. The first build is one afternoon. The marketplace earns from launch day.