I gave a talk at Techweek this week on what it actually means to build an AI-native company. The thesis was deliberately ambitious: capture and curate every conversation and decision in your business, and you end up with something approaching a company brain - queryable by humans, and increasingly by the agents acting on their behalf. It is a stretch. It is also something we have been building for ourselves inside Airclerk - our AI Chief of Staff, Alex, sits on top of it and is how we run the business day-to-day. This is a short note on the gap between that destination and where most regulated businesses are starting from - and what to do about it.
The shape of the idea
An AI-native company treats its own institutional memory as a first-class asset. Not the wiki nobody updates. Not the shared drive nobody can find anything in. The actual substrate: meeting recordings and transcripts, decisions and the reasoning behind them, customer conversations, internal Slack and Microsoft Teams threads, design reviews, post-incident notes, regulatory correspondence, the renewal that almost got missed and why it didn't.
Curate that material - structure it, attribute it, make it searchable with the right permissions - and two things become possible. Humans get a memory that scales past any individual. Agents get the context they need to do useful work without being micromanaged. The "company brain" is shorthand for that combined surface.
It is a stretch goal. I said as much on stage. But the direction is real, and the firms that get there first will be operating on a different cost curve to the ones that don't.
The trap
The trap is treating the company brain as something you go and buy. You can't. There is no SKU. What gets sold under that banner today is mostly a search box over a document store, with a model bolted on. That is not the same thing.
The actual prerequisites are the unglamorous ones:
- Connected systems. Your policy admin, broking platform, CRM, claim system, ticketing, email and document store have to be reachable by the same agent under the same identity. Most aren't - and honestly, a lot of the industry software you depend on does not make this easy. Where a proper API exists, MCP servers and connectors are how it gets fixed. Where one doesn't, web agents and browser-driving techniques are a workable bridge until vendors catch up.
- Permissioned context. An advisor can see one client's file. The complaints team can see all of them. The board sees a redacted summary. The brain only works if it respects that lattice - and most knowledge bases simply don't.
- Evidence and audit. Every answer the brain produces needs to be traceable to a source, with a timestamp, with the identity of whoever asked. Without that, you can't defend a single output to a regulator.
- Governance discipline. Approval thresholds, exception queues, human sign-off for outbound work, supervision over high-risk conduct. Operating disciplines, not features.
- One workflow that actually runs. A company brain that has never been used to do a real piece of work is a slide deck. You earn the brain by shipping one workflow into production, then another, then another. The substrate is the by-product, not the project.
Why the foundations matter more for regulated firms
If you run a marketing agency, you can be sloppy about the foundations and get away with it for a while. If you run an insurer, a brokerage, a wealth manager, an accountancy or a bank, you cannot. The day a customer asks why they were declined, or a regulator asks how a piece of advice was generated, the evidence trail has to exist - and it has to be the same trail an agent would have used to do the work in the first place.
This is why the "company brain" conversation, for regulated services, has to start with implementation. The brain is downstream. The plumbing comes first.
Where Airclerk fits
Airclerk is the team that puts that plumbing in. We are not selling you the brain on day one - that is a destination we will help you grow into. We are doing the work that has to be done before the brain is anything more than a slide.
- A two-week Readiness Sprint that turns scattered AI experiments into a board-ready Claude implementation plan, with the first workflow named and scoped.
- Workflow implementation - governed Claude workflows connected to the systems your business actually runs on, via MCPs and connectors.
- AI governance - the controls, approvals, audit trails, cost controls and supervision that make the outputs defensible and the spend predictable.
- Managed operations for the firms that want us to run it day-to-day while their team grows into ownership.
Pick one workflow. Get it into production with the rails underneath it - identity, permissions, evidence, governance. The substrate accrues from there. Five workflows in, you will start to recognise the shape of the brain forming. Twenty in, your team will be operating in a different way than it does today.
If you came here from the talk
Thank you for coming. The company brain is the destination I think regulated services in NZ and AU should be planning toward. The first step is almost never glamorous - it is one well-chosen workflow, properly implemented, with the governance and evidence trails a board can defend. That is what we do. If you'd like to talk about what that looks like for your business, the Readiness Sprint is the easiest door in, and hello@airclerk.ai reaches me directly.