Peregian Digital HubAI Adopted

Building your company's brain

Building your organisation's brain

How SMEs capture, organise, and make their knowledge accessible so AI systems can use it reliably and improve from experience.

Building your organisation's brain means making the knowledge your people carry—how they solve problems, where they find information, which edge cases trip them up—accessible to AI systems so those systems can work reliably and improve from experience. Without this, AI becomes a generic tool that reads what is written but misses the shape of a decision. With it, agents learn your business and amplify the judgment of your best people.

Documented methods: start with skill libraries

Hiten Shah, a product strategist, argues that organisations should prioritise building a skill library—a documented repository of proven ways of working—as their foundational AI strategy. A skill is a reusable packet of procedure, judgment, and edge-case awareness: it captures how experienced people approach repeated work, bundled as instructions, examples, templates, checklists, and rules of thumb.

Why start here rather than with data access? Shah contends that agents need to understand how a company does its work, not just have access to data; without that understanding, agents can read all available information but still miss the shape of a decision. Two firms using the same frontier model will diverge dramatically if one has a skill library and the other does not, because a company's AI advantage comes from the work it teaches the model to do well rather than from the model it chooses.

The most valuable skills are private and company-specific, since the most valuable methods are specific to each organisation. Generic marketplace skills will exist but offer little competitive advantage. Shah makes a key observation: "Your company already has skills. They are sitting in old docs, Slack threads, customer calls, review rituals, onboarding notes, and the heads of the people who know how the work really gets done." The work is mapping repeated tasks where experienced people consistently outperform others (particularly tasks involving judgment rather than effort alone) and packaging those approaches as reusable skills.

Centralising context and connecting systems

YC's Peter Koomen, who built the firm's internal AI infrastructure over the past 18 months, describes the infrastructure as resting on one core principle: centralising all critical organisational context into one database—every company funded, founder, financial transaction, and internal note—so agents can answer arbitrary business questions. When context is fragmented across email, Slack, spreadsheets, and individual machines, agents cannot reliably construct the information they need. Koomen reports that YC moved to a centralised architecture and found that removing the back-and-forth between teams to get a thing done prompted staff to ask far more complex questions, unlocking use cases that would have been invisible in a fragmented system.

But centralised context alone is not enough. Agents also need the ability to operate across the systems where your organisation already keeps its data and work. Aaron Levie, CEO of Box, demonstrates this pattern: Claude using Box and Linear (a work-tracking tool) to turn product roadmap documents into trackable tasks, bridging a gap most SMEs have—knowledge stored in one tool and delivery tracked in another. This is not integration for its own sake; it is the operational difference between an agent that can fetch information and an agent that can execute work. Background and workflow-triggered agents will consume far more tokens than chat-based interactions because they run continuously while doing work unattended. To make this scale, agents need safe code execution, tool use, compute sandboxing, and the ability to connect across enterprise systems.

Learning loops: continuous refinement

Satya Nadella, Microsoft's chief executive, argues that competitive advantage comes from building a proprietary learning loop that compounds over time. Organisations must build human capital (people's knowledge, judgment, relationships, pattern recognition) and token capital (owned AI capability) as complementary, compounding assets; human capital does not lose value as token capital grows, but becomes more valuable through directing AI systems toward meaningful goals.

The mechanism is concrete. At YC, Koomen describes self-improving skill loops that run autonomously each night, reading past agent conversations to identify missed opportunities and context, then refining skills. A skill for writing two-sentence pitches has become better than Koomen at the task after learning from partner feedback. Each interaction, each correction, each refinement feeds back into the system. Nadella notes that organisations that encode their accumulated expertise into learning systems early will build advantages hard to replicate, because the loop compounds with every improved workflow and generates better training data unique to the firm.

Nadella frames this as the central competitive stakes: a world where all value accrues to a few foundation models risks hollowing out industries through commoditisation of organisational knowledge, repeating the economic displacement pattern of earlier waves of technological change. The alternative is a frontier ecosystem where every firm owns its learning loop.

Public work and distributed learning

Tobi Lütke, Shopify's chief executive, describes an unintentional learning engine at Shopify built around an internal AI agent called River. River lives in public Slack channels rather than private windows, and the result is that staff absorb patterns by watching other engineers work with the agent, without formal training or curriculum. In one month, 5,938 Shopify employees used River across 4,450 channels; roughly one in eight pull requests merged that week came from River. Lütke reports that River's merge rate climbed from 36% to 77% over two months without model retraining: teams noticed where it got stuck, fed back what it should have known, and the agent improved by absorbing each team's accumulated knowledge.

This matters because of what it reveals about how teams actually learn. Private AI windows (ChatGPT, Claude, Cursor) lock apprentices out; whoever is at the keyboard learns, everyone else is locked out of the model's improvement loop. By contrast, Lütke writes, organisational speed is set by the slowest communication channel; searchable, public work with an agent spreads information faster than meetings, email, or private DMs and creates a record the next person finds.

Lütke calls this a "Lehrwerkstatt"—a teaching workshop where learning spreads by proximity rather than formal training. The agent does not replace the mentor or apprentice; it makes the whole company an apprentice by putting experienced people's work on display so others can watch both human and AI learn. But this requires deliberate design: public-by-default channels, searchable history, and a norm that work is documented as it happens.

Building in practice

Pedro Franceschi, co-founder and CEO of Brex, frames building an organisational brain as a matter of redesigning core processes end-to-end rather than automating existing ones. Redesigning entire processes (not bolting AI onto legacy work) yields the largest gains: Brex redesigned K.Y.C. onboarding end-to-end, discovering that free K.Y.C. at lead stage enables risk-based funnel qualification.

This requires the CEO to personally lead AI strategy. Franceschi argues that the CEO must be the chief AI officer: only they have organisation-wide context and authority to redesign core processes, so delegation to engineering teams leaves structural opportunities unexplored. He emphasises that hands-on engagement with AI tools is essential for company-wide adoption, since it is the only way to understand the technology's boundaries and possibilities at real scale.

Building the infrastructure also bridges an organisational gap. Most organisations neglect at least one of three tiers of AI adoption: product AI (customer-facing), operational AI (serving at scale), and corporate AI (how staff work). A company brain connects all three: it powers agents that simplify how staff work, and it surfaces organisational knowledge to both internal operations and customer-facing AI.

One concrete example: Brex built CrabTrap, an HTTP proxy that audits all agent requests using LLMs as judges for ambiguous ones, achieving 98% auto-approval and 2% human review. This is both a security architecture and a feedback mechanism—every decision improves the next one.

Sources