Bottom line: On August 6, 2026, four AI startups announced funding on the same day, each targeting a single narrow business function: customer support routing (Omilia, $67M), recursive AI research (Mirendil, $100M-plus in Google Cloud compute), e-commerce personalization (Malachyte, $10M), and back-office company setup (Naive, $28.5M). None of them are building a general-purpose assistant. That’s the actual signal — investors are backing teams that picked one expensive, repetitive task and went deep, not teams promising to automate everything with one model.
Omilia’s $67M Bet: Most Customer Support Queries Don’t Need an LLM
Omilia, an Athens-based company that has worked on automating voice calls and customer support since 2002, raised a $67 million Series B led by Expedition Growth Capital, according to TechCrunch. It’s the company’s second raise since a $20 million round from Grafton Capital in 2020 — in that stretch, Omilia grew annual recurring revenue 10x to $60 million.
CEO Dimitris Vassos drew a direct contrast with generative-AI-first competitors like Sierra, Decagon, and Parloa: a large share of incoming customer support queries — account balances, order status, basic lookups — don’t need a large language model at all, and routing every query through one wastes money at scale.
“We will use any available weapon to win the battle for customer service. Companies like Sierra, Decagon, etc. identify as generative AI companies. Their sole purpose is to deploy generative AI and limit themselves. You may have a bazooka, but if your enemy is near you, you need a knife,” Vassos told TechCrunch.
Omilia has since expanded into self-learning agents that work across multiple customer contact points, but the company’s stated edge is unit economics rather than model sophistication — Vassos said Omilia hasn’t needed the large cash infusions its competitors have because both Omilia and its customers see better margins under its approach.
Mirendil’s $100M+ Google Cloud Deal: Compute as the New Funding Round
Mirendil, an AI lab working on self-improving AI, signed a multiyear partnership with Google Cloud worth more than $100 million to source compute capacity, TechCrunch reported exclusively. Co-founder and CEO Behnam Neyshabur — who, like several of Mirendil’s founders, previously worked at Anthropic — told TechCrunch the deal is worth roughly half of what Mirendil raised in seed funding at a $1 billion valuation in late June 2026.
The deal gives Mirendil access to both Google TPUs and Nvidia GPUs plus managed training clusters, which the company will use to pursue recursive self-improvement — AI systems that iteratively improve their own capabilities. Neyshabur’s framing: point a problem at the system and it keeps getting better over time, the way a human scientist accumulates domain expertise. Mirendil’s stated ambition is for its AI to eventually take on the work of an entire frontier AI lab, accelerating research in fields like medicine, biology, and materials science. A handful of other startups — Recursive Superintelligence and Ricursive Intelligence among them — have recently formed around the same goal.
Why a Compute Deal, Not Just a Funding Round, Matters
A $100 million cloud commitment functions like a second funding round denominated in GPU-hours instead of cash — and it signals that Google is willing to underwrite a research bet on recursive self-improvement specifically, not just generic model training. For any team competing with Mirendil on research-automation claims, the practical bar just moved: raising cash isn’t enough if a rival has locked in guaranteed TPU and GPU access at scale.
Malachyte’s $10M Seed: Spotify’s Recommendation Engine, Rebuilt for Shopping
Malachyte, founded by three former Spotify employees — Sidd Motwani, Ian Anderson, and Shivaditya Sinha — raised $10 million in seed funding, according to TechCrunch. At Spotify, the trio built Vector AI, the behavioral-intelligence system that powers roughly 90% of recommendations served to Spotify’s 800 million users by predicting listener intent rather than relying purely on past behavior.
Malachyte applies a “two-headed Vector AI” model to e-commerce: it starts forming a read on a shopper’s intent from the moment a page loads, before the first click, and continuously re-ranks what a store shows based on real-time signals rather than login history or past purchases alone. CEO Motwani gave a concrete example to TechCrunch: a search for “heavy-duty boot” followed by two clicks on steel-toed boots is enough for the system to move work pants and gloves up the page and push dress shoes down — with no account or purchase history required.
The bet here is narrower than “personalize the whole store” — it’s specifically real-time intent inference for anonymous, first-time visitors, which is the segment most existing recommendation systems handle worst because they have no historical data to work from.
Naive’s $28.5M: Automating the Paperwork of Starting a Company
Naive raised $28.5 million to build infrastructure that lets AI agents handle the bulk of setting up and running a business, TechCrunch reported, and says it has signed up more than 30,000 developer customers within months of launch. The company packages U.S. LLC incorporation, payments setup, email accounts, phone numbers, cloud infrastructure, and storage provisioning behind a single API.
Developers use a prompt Naive supplies to agentic coding tools — Cursor, Claude Code, or Codex — which then call Naive’s API to provision the business infrastructure. Some steps still require a human: KYC/KYB identity verification and any payments have to be completed by the founder, not the agent. A governance layer lets users set budgets, restrict what the agent is allowed to do, and require human approval before sensitive actions execute. Naive also ships templates for common business types — AI SEO, full-stack SaaS, recruiting, accounting, customer support — plus a mobile emulator agents can use to complete tasks that require a phone interface.
Pro Tip: If you’re evaluating agent-driven back-office tooling like Naive, check the governance layer’s approval gates before you check the feature list. The dangerous failure mode for autonomous company-setup infrastructure isn’t a missing feature — it’s an agent that completes an irreversible action (a payment, an incorporation filing) without a human checkpoint.
The Pattern Across All Four Deals
Every one of these four companies picked a single, well-defined, expensive-to-do-manually task and built specifically for it, rather than positioning as a general AI platform. Omilia explicitly rejects the “generative AI for everything” framing. Mirendil is narrow by design — one research capability, pursued with dedicated compute. Malachyte solves one problem (anonymous-visitor personalization) that generic recommendation engines handle badly. Naive automates one operational category (company setup and back-office provisioning) rather than claiming to run the whole business.
Investors backing four narrow bets on the same day, across four different verticals, is a more useful signal than any single round: the money is following defensible, measurable unit economics in a specific task, not breadth of capability. This mirrors what we found evaluating the build-your-own-agent route with n8n and Claude — the automations that actually hold up in production are the ones scoped to one workflow, not an attempt to replace an entire job function in one system.
Add up the fresh equity alone — Omilia’s $67 million, Malachyte’s $10 million, and Naive’s $28.5 million — and it’s roughly $105.5 million committed in a single day to three companies with no overlap in what they actually automate. Layer Mirendil’s $100 million-plus compute commitment on top and the total capital in motion that day tops $200 million, none of it aimed at a single “do everything” model. That’s the detail worth sitting with: at the exact moment frontier labs are spending billions to make one model handle more tasks, a separate and just-as-well-funded market is betting the opposite — that the money is in doing one task so well a generalist model can’t catch up on cost or precision.
What This Means for You
If you’re evaluating an AI vendor for a specific function (support, personalization, research automation, or ops): favor a vendor whose pitch is “we solved this one task well” over one promising a general platform. Omilia’s own argument against its generative-AI-first competitors is the evaluation question to ask any vendor — does every request actually need a full LLM call, or are they charging LLM-scale prices for lookup-scale problems?
If you’re a developer considering Naive-style agentic infrastructure to bootstrap a company or internal tool: the governance and approval-gate design matters more than the list of integrations. Test what happens when the agent is denied an action, not just what happens when it succeeds. For a narrower, more controllable starting point, our Cursor vs. GitHub Copilot vs. Claude Code comparison is a useful baseline for the coding-agent layer these tools plug into.
If you’re building your own vertical AI product: the funding pattern here argues against a broad-platform pitch. Pick the one task in your domain that’s expensive, repetitive, and measurable, and build the unit economics story before the feature list. For teams exploring the research-automation end of this spectrum specifically, see our guide on building a personal AI research assistant with Perplexity and Claude as a lower-cost starting point than a Mirendil-scale compute commitment.
Sources: TechCrunch. Last updated: 2026-08-07.
