Become AI Native in less than 60 mins

Theo Tabba defines an AI native organization as one where people manage agents, agents can read and write to the company, and the company gets smarter over time. He draws a sharp distinction between using ChatGPT and being truly AI native, comparing the gap to the difference between having a website and calling yourself a tech company. The system has three elements: people, who handle strategy, taste, judgment, and trust; agents, who interface with context on behalf of people; and context, which requires the company to be AI readable. The output of a well-functioning system is speed and market signal, both of which feed back to make the system smarter.

Tabba frames everyone in an organization as now effectively a manager, citing Andy Grove's principle that a manager is judged by the output of their team. AI handles the middle of work, meaning execution, while humans focus on the beginning, meaning strategy and research, and the end, meaning review and communication. Deploying agents without people who understand how to manage them will not produce results. For an agent to operate autonomously it needs four things: a clear goal, the right skills, the right tools, and the right context. Tabba argues that impatience with AI typically stems from providing simple prompts without supplying those four elements.

A skill is a markdown file encoding a specific capability. An eval is visibility into agent output, measuring quality against a desired result such as a rating of eight, nine, or ten out of ten. Skill chains run multiple skills sequentially, where one skill fires and calls the next. Organizations not using skill chains are described as managing on hard mode, treating every agent as an ultra junior employee requiring constant direction. As a live example, a proposal skill chain fires three skills in sequence: one creates a proposal microsite, a second ensures the output sounds human rather than like AI, and a third QA skill prevents hallucinations and over-promising by ensuring nothing is included that was not pulled directly from transcripts or data. The chain fires automatically on a trigger such as a proposal request detected in a meeting transcript or inbox, deploys live on a link, and sends a Slack notification. The proposal was generated and the notification received within approximately two minutes. Tabba states that personalized AI-generated proposals have resulted in millions of dollars of revenue for LCA, and that without this automation getting a proposal to a prospect could take days, risking the prospect cooling off or going elsewhere.

The context layer, referred to as the brain, is structured as folders containing markdown files that guide agents to relevant information. A routine runs approximately every one to two hours to collect information from Slack, meeting recordings, emails, and project boards into a brain inbox. Not all captured information is stored; a curation step acts as a librarian to filter what belongs in each folder. Human judgment must remain in the loop to evaluate what context is good or bad before it flows back into the system, preventing bad agent output from polluting the brain layer. Tabba describes the context layer as giving agents what he calls 2020 vision on a company, enabling access to organizational information that even employees cannot readily recall. Platforms like Notion AI offer search plus context plus an agent layer but risk provider lock-in and operate as black boxes, which the speakers position as a drawback compared to building your own system.

A concept called traces or exhaust refers to decisions, documents, and explorations produced during work that typically end up in a graveyard of files no one revisits. Agents can act on these traces to create new artifacts, store lessons, and capture how decisions were reached. When humans manage agents and provide feedback, the agents remember it, update skills, update memory, and package information into lessons. Market signals such as whether customers buy more after a new feature or churn faster after a landing page change also flow back through tools into the brain to keep it current.

A proposal that previously took up to three days now takes minutes using this system. A functional prototype that previously took one to two weeks to build and test now takes minutes, including feedback collection and synthesis into a second version. LCA built a working playlist-style music prototype live during the session using Claude, demonstrating that stakeholders can feel and react to something rather than waiting weeks for a product requirements document. LCA has approximately 50 to 55 people and roughly six years of accumulated context from work with Fortune 2000 companies, which is cited as a driver of prototype quality.

The recommended startup strategy is to niche down across three vectors: industry, function, and company size. Restaurants are described as a particularly hot niche because the industry is fragmented, though restaurants too small will not have the budget to afford the service. The prioritization framework moves from niche to general and from high-frequency to low-frequency workflows, with low-frequency high-value niche workflows described as potentially having higher ROI. Tabba argues that becoming AI native does not require being a technical expert and that the key mental shift is thinking through the lens of managing agents and what those agents need to succeed, predicting that companies adopting that mindset will be ahead of most companies in the world.