Generic AI is a commodity. Building on top of it is not.
Since GPT-4 shipped, thousands of companies have wrapped a foundation model in a product UI and called it an AI platform. Most of them share a common structure: a prompt, a chat interface, a payment funnel. The technical moat is zero. The margin is whatever OpenAI and Anthropic don't charge.
The law firm in Manchester that subscribes to "Law GPT" is paying for a prompt template. The medical clinic in Warsaw that buys "Clinical AI" is paying for a chatbot with a specialised system prompt. When the underlying model improves, the wrapper stays the same. When the underlying model changes hands, the wrapper breaks.
Why wrappers fail in regulated work
When a solicitor advises on a contract, she is not running an autocomplete. She is integrating the clause, the case law, the client history, the jurisdictional context, and her own risk posture into a single judgment. The output is load-bearing — it determines a real-world outcome, and she carries the professional liability.
A wrapper cannot deliver that. A wrapper is a single prompt, in a single conversation, with a single model. It has no memory of the matter. It has no verification layer. It cannot tell you why it said what it said. If it hallucinates a case citation, you only find out in court.
"A wrapper is a prompt. A system is a pipeline. The two look similar for about ten seconds — and then the difference is the rest of your career."
What Blackflake is, structurally
Blackflake is a multi-agent system. Instead of one prompt, one model, one answer, a Blackflake engagement runs through a composed pipeline of specialised agents:
- An extraction agent identifies the matter type, jurisdiction, and key clauses.
- A domain agent applies the vertical reasoning — contract analysis for legal, ICD-10 coding for medical, eligibility matching for procurement.
- A retrieval agent pulls matter-specific context from the relevant source of truth.
- A verification agent audits the output against rules, prior matter history, and known failure modes.
- A QA gate flags anything below confidence threshold for human review.
Each agent has a defined role, a documented toolset, and a declared failure mode. The system is inspectable — which is the part wrapper products structurally cannot replicate.
The Human-AI Collaborative Development methodology
How Blackflake is built matters as much as what it is. One human architect defines the structure. Autonomous agents build components. A QA layer verifies. This compresses the time from "idea" to "production vertical" from quarters to weeks.
By the numbers: ~55 autonomous agents, 64 workflows, 31 machine-readable sources of truth, 225+ production surfaces across six brands. Built by one architect using a system of agents that audit each other. Not a hypothetical roadmap — the operating surface today. See the ecosystem schematic on the homepage for the shape of it.
The moat is architecture, not cleverness
A wrapper's moat is its distribution — who saw the Facebook ad. Blackflake's moat is its composition — the specific way agents, skills, sources of truth, and verification gates are wired together for a regulated vertical.
A competitor with a foundation-model API key cannot clone that overnight. They would need to build a working ecosystem of ~60 agents, a verification discipline, and the vertical-specific reasoning. That is a year of engineering, minimum. During that year, the Blackflake ecosystem compounds further.
The business-model shift
Wrappers bill per seat, per month, per token. The token price floor drops every quarter. Wrappers compete on interface polish — a brutally expensive place to compete when your interface is a chat window and the model is someone else's.
Blackflake engagements are services-as-software: you license the deployed system, run it forever, compound on the same engine as you expand verticals or jurisdictions. It is the same shift Workday forced on enterprise HR software, what Palantir forced on government analytics. It is now ready for regulated professional services — legal, medical, procurement, compliance.
When you should buy a wrapper anyway
Some problems don't need infrastructure. If you want a chatbot for internal FAQs, a wrapper is fine. If you want to draft a cover letter faster, ChatGPT is excellent. If your output isn't load-bearing — if no one gets fired or sued when it is wrong — you don't need verification, inspection, or matter memory.
For everything else, you are not really looking for a tool. You are looking for a system you can own, audit, and trust in production. That is what we build.
— Bartek Kubas · Founder-architect · Blackflake Łódź, Poland · 22 April 2026