Introducing Research Lab

Research Lab is an AI-native organisation for building new companies, products, and public research from the same operating system.

The definition matters. We are not a traditional consultancy, not a passive research institute, and not a product studio that simply adds a chat interface to existing workflows. Research Lab is designed as an AI team: a small human staff working with a larger layer of AI agents across research, engineering, operations, analysis, finance, documentation, and strategy.

The company itself is the first experiment.

Our work starts with a practical question: if AI systems can now reason over documents, write and debug code, inspect data, use tools, compare options, and support long-running workflows, what kind of organisation becomes possible? OpenAI's work on GPT-5.5 and Google's work on Gemini for Science show the same broad direction: AI is moving from answering questions to helping teams execute complex work. Research Lab is our attempt to turn that shift into useful products and durable companies.

A research-driven company builder

Research Lab will develop products directly, but it will also create spin-offs when a product becomes strong enough to stand as its own company.

That model changes how we work. Each product begins as a focused research and implementation loop: understand the domain, map the existing process, build a prototype, measure failures, improve the workflow, and publish the useful learnings where we can. The output may be a commercial product, a new company, an open-source component, a technical note, a benchmark, or a decision not to continue.

This is research-driven product development, not research as decoration. The company will produce research about the markets it enters, the AI systems it uses, the evaluations it builds, and the organisational model it is testing. We want our own approach to be inspectable: where agents reduce cost, where they increase risk, where human review is still necessary, and where smaller teams can do work that previously required much larger organisations.

Anthropic's research on how AI is transforming work inside Anthropic, estimating productivity gains, and labor-market impacts supports the same lesson: the value is not only task automation. The value is in redesigning how teams are organised around the new capability.

Better products at lower cost

Our first focus is business processes, financial technology, transactions, banking, and retail and business finance.

These are areas where cost and experience are still badly misaligned. Retail users and small businesses often pay through time, confusion, hidden fees, slow processes, poor interfaces, manual reconciliation, weak comparison tools, and systems that were built for larger markets. The opportunity is not only to make software cheaper. It is to make the experience materially better while reducing the cost to deliver it.

AI makes this possible in several ways:

• Agents can handle repetitive operational work that keeps services expensive;
• Better data extraction and comparison can make financial options clearer;
• Automated checks can reduce manual review without removing accountability;
• Product teams can test and iterate faster with smaller headcount;
• Open components can lower the cost for other builders in the same ecosystem.

The goal is not automation for its own sake. The goal is better financial and business services for people who have been underserved by expensive, slow-moving, or poorly localised software.

This is also why we are starting several products in Kosova - it is a real test of whether AI-enabled products can work where infrastructure is fragmented, trust has to be earned, budgets are constrained, and users still deserve a modern experience. If a product can reduce cost and improve quality there, it is more likely to be robust elsewhere.

Safety is part of the product

AI-native does not mean uncritical. More capable agents create new failure modes, and those failures matter most in finance, operations, and institutional work.

Research on chain-of-thought monitorability, reasoning models that do not always say what they think, monitoring internal coding agents for misalignment, and Anthropic's model diffing all point to the same operating principle: agents need evaluation, monitoring, comparison, and audit trails. Interpretability work such as Natural Language Autoencoders and research on emotion concepts in language models also makes clear that model behaviour can be useful without being simple.

For Research Lab, those are not abstract research topics. They are product requirements. If an agent helps with a transaction, a business-finance decision, or an operational workflow, we need to know what it saw, what it did, what it changed, when it should escalate, and how its behaviour shifts after model or prompt updates.

What we will publish

Research Lab will publish what is useful and responsible to publish:

• Research notes on AI-agent operations and product development;
• Evaluations and benchmarks from our own systems;
• Implementation documents for financial and business workflows;
• Open-source code and reusable components;
• Postmortems on experiments that did not work;
• and announcements for products and company spin-offs.

The health and science examples from Google, including Gemma work on cancer therapy discovery, show how AI can help research move faster when it is connected to real experimental loops. The economic research from Anthropic's Economic Index shows that access and adoption are uneven. Research Lab sits between those two facts: AI can accelerate useful work, but the benefits will not distribute themselves automatically.

What comes next

Over the coming months, we will launch products in financial data, business processes, transactions, and AI-agent infrastructure. Some will remain inside Research Lab. Some may become independent spin-offs. All of them will be used to test the same thesis: smaller AI-native teams can build lower-cost, better products if they combine research, product engineering, evaluation, and real deployment from the beginning.

Research Lab exists to make that thesis concrete.

Sources

• Introducing GPT-5.5
• How we monitor internal coding agents for misalignment
• Reasoning models struggle to control their chains of thought, and that's good
• Evaluating chain-of-thought monitorability
• Appendix: Project Deal
• Natural Language Autoencoders Produce Unsupervised Explanations of LLM Activations
• Emotion Concepts and their Function in a Large Language Model
• A diff tool for AI: Finding behavioral differences in new models
• Labor market impacts of AI: A new measure and early evidence
• How AI Is Transforming Work at Anthropic
• Estimating AI productivity gains
• Preparing for AI's economic impact: exploring policy responses
• Anthropic Economic Index: Tracking AI's role in the US and global economy
• Reasoning models don't always say what they think
• How a Gemma model helped discover a new potential cancer therapy pathway
• Gemini for Science: AI experiments and tools for a new era of discovery