Sep 18, 2026 · 7 min read
How to build an AI-native company: capture everything, close the loop
AI-native companies are not defined by how many AI tools they buy. They are defined by whether their work leaves a record that AI can learn from, and whether decisions get checked against results.
An AI-native company is one where AI is not a tool people occasionally use but the layer the company runs on: meetings, conversations, decisions and outcomes are captured as a matter of course, so that AI agents can read that context, do real work with it, and improve over time. The practical recipe has two parts. Capture everything important, and close the loop by comparing what you decided with what actually happened.
What is an AI-native company?
An AI-native company treats AI as its operating system rather than as an add-on. Every important workflow produces an artifact an AI system can read, and the organization improves because that system keeps learning from its own outcomes.
The contrast is with companies that bolt a copilot onto existing workflows. The AI-native question is not which tasks AI can speed up, but whether the company is legible to AI and learns from what it does. That leads to different habits:
• Conversations are recorded and transcribed by default, with consent.
• Decisions are written down where agents can find them, not buried in DMs.
• Tickets, docs, metrics and customer feedback live in systems agents can query.
• Outcomes are measured and fed back into the next decision.
We have written more about the organizational side in what an AI-native company looks like.
What does Y Combinator say about building AI-native companies?
YC’s core advice is to make your company queryable and run every important process as a closed loop. Partners have made this case in a Startup School talk and on the Lightcone podcast in 2026.
In “The Playbook For Building An AI Native Company” (April 2026), YC Partner Diana Hu argues that AI should be the operating system a company runs on, not just a tool it uses. Her central claim is that “every important process in your company should be captured by an intelligent closed loop.” She defines one plainly: “A closed loop captures information, feeds it back into an intelligent system, and improves the process over time.” Companies in the old world, she says, ran as open loops: they made a decision, executed it, and often never systematically measured the result.
To build those loops, Hu says, the whole organization should be legible to AI. Her concrete list starts with “recording your meetings with an AI note taker, minimizing DMs and emails,” plus embedding agents in communication channels and building dashboards across revenue, sales, engineering, hiring and operations. Her example is sprint planning: an agent with access to tickets, Slack, customer feedback, docs and call recordings can analyze what shipped last sprint and propose the next plan. The principle underneath, in her words: “you need to provide models with as much context as you would provide an employee.”
In a Lightcone episode, “Inside YC’s AI Playbook” (May 2026), YC General Partner Pete Koomen described building YC’s own internal agent infrastructure. His opening advice was that “you need to start recording all the artifacts. It’s like a shared organizational brain.” He also described how agents working over a single database, with everything in one schema, showed him “there’s a lot of value, at least, in getting all of the context into one place.”
Two ideas repeat across both: capture context comprehensively, and put it somewhere an intelligent system can use it.
What does Brex do to operate as an AI-native company?
Brex’s CEO says the company is rebuilding itself as AI-native, and three parts of that are publicly documented: company-wide AI meeting notes, a personal autopilot that gives the CEO context from across the company, and a “customer world model” that ingests customer touchpoints.
On the Lightcone episode “The CEO Must Be the Chief AI Officer” (June 2026), Brex co-founder and CEO Pedro Franceschi described building a customer world model at Brex, where “we’re trying to get every single touch point that the customer has of us.” He listed examples ranging from dashboard clicks to what customers say in emails and on calls, all ingested and consolidated to anticipate what each customer will need next. He said it is already in use in Brex’s sales work, including to get up to speed on an account before meeting a customer.
On the meeting side, Granola’s published case study on Brex quotes Franceschi: “As we rebuild Brex into an AI-native company, we need tools that move fast without ever compromising accuracy.” The case study, which is vendor-published, describes Brex rolling out Granola’s AI notepad for internal and external meetings with summaries posted to Slack and shared folders of past conversations that teams can query.
The most striking part is how Franceschi uses that context himself. On the Core Memory podcast with Ashlee Vance (April 2026), he described an autopilot he built that screens his email, Slack, Google Docs and WhatsApp, plus notes from Granola, which he said runs on every meeting he has. He organizes it around the programs he cares about, such as financial performance or internal AI strategy, and roughly 25 key people. It produces a summary of what is happening across the company that he should know about, along with action items, and can draft follow-ups using the context from each meeting. In the same conversation he said Brex had taken out two layers of management as part of its turnaround.
Put those together and you can see what captured context changes about leadership. When the discussion behind a decision is written down, a leader no longer depends on it being filtered up through layers of status updates. They can see the actual reasoning, including a strong argument from someone junior that might never have made it into a summary, and back the best idea on its merits rather than by seniority. That is the organizational version of what Diana Hu calls removing human middleware: when a company is queryable, information does not need to be routed up a hierarchy to reach the people making decisions.
The pattern is the one YC describes: conversations become text, text lands somewhere searchable, and agents and people build on it.
Why does capturing every conversation matter?
Because most of the context behind product and business decisions is spoken, not written. If it is never captured, no agent can use it, and no one can later check whether a decision matched what customers actually said.
A customer explains a workaround on a call. A salesperson hears the same objection for the third time. An engineer flags a risk in standup. None of that reaches a ticket unless someone writes it down, and whoever writes it down filters it.
Capturing conversations fixes three problems:
• Agents get the same context as employees, which is Hu’s core principle.
• Evidence becomes traceable. A request can point back to the calls where it came up.
• Patterns become visible. One request is an anecdote; the same issue across many calls is a signal.
We walk through the downstream half of this in turning customer calls into product work.
How do in-person teams capture context?
In-person teams can use wearable or phone-attached AI recorders, or a phone app, for meetings that do not happen on Zoom or Google Meet, and then bring the transcripts into the same place as their other context. Consent comes first.
Video calls are easy: most teams already use a notetaker. The gap is office conversations, customer site visits and field work. Two devices that target it, described in their makers’ terms:
• Pocket, from Open Vision Engineering, is a MagSafe-compatible device that attaches to the back of a phone and records in-person conversations and phone calls, with transcription and summaries in its app.
• Plaud NotePin is a small wearable recorder that can be clipped or pinned on, and turns recordings into transcripts and AI summaries.
Whatever you use, move the transcript into the system where your team’s context lives, so it is not stranded in a personal app.
A note on consent. Tell people when you are recording, every time, and make it easy for them to say no. Laws differ by location. According to the Reporters Committee for Freedom of the Press, US federal law requires the consent of at least one party, and about 11 states primarily require the consent of all parties, with other states applying all-party rules to some kinds of conversations. Rules outside the US vary too. This is not legal advice; check the rules where you and the people you record are located, and set a company policy before rolling devices out.
What does a closed loop look like in practice?
A closed loop connects four steps: capture context, decide with evidence, ship, and compare the result with what you expected. The last step is the one most companies skip.
Here is a product example:
1. Capture. Calls, support threads, Slack discussions and usage data flow into one shared memory.
2. Decide. A pattern emerges, say onboarding drop-off mentioned across many calls and visible in analytics. The team writes down a decision and the impact they expect.
3. Ship. The work is broken into tickets and built.
4. Measure. After release, the team checks the actual change in the metric against the prediction, and records what they learned.
The fourth step makes the loop self-correcting. Without it, you are running what Hu calls an open loop. With it, every release sharpens the next prediction. We cover the measurement step in more depth in how to measure what you shipped, and the full model in what closed-loop product management is.
How do you start?
Start small: pick one loop, capture everything that feeds it, and measure its outcome. You do not need to reorganize the company first.
1. Pick one process with measurable outcomes, such as product planning or onboarding.
2. Record the meetings that feed it, on video calls and in person. Announce it and get consent.
3. Move decisions out of DMs and into shared channels or docs.
4. Put transcripts, tickets, docs and metrics in one place where they can be queried together.
5. For every meaningful decision, write down what you expect to happen and how you will measure it.
6. After shipping, compare results with the prediction and feed what you learned into the next decision.
7. Once one loop runs reliably, apply the pattern to the next process.
Where Fijord fits
Capturing context is half the job. The other half is turning it into decisions and checking them against reality. Fijord, a product-management platform in early access, is built for that half.
Fijord connects to the tools teams already use, including meeting notetakers and recordings (Granola, Gong, Zoom, Google Meet), Slack, Linear, Jira, Notion, GitHub and Mixpanel, and works alongside them. It builds one shared product memory from calls, docs, tickets and metrics. Signals surface as evidence builds, and every insight links back to its source. Transcripts from other tools, including in-person recorders, can be added as manual inputs. From that evidence, Fijord drafts briefs, epics and tickets and syncs tickets to Jira or Linear, while Pulse surfaces blockers, approvals, risks and opportunities.
Each recommendation carries a predicted impact. After release, Fijord compares that prediction with actual results in your analytics, and the outcome feeds the next decision. That is the closed loop, applied to product work.
Frequently asked questions
What is the difference between using AI and being AI-native?
Using AI means adding tools to existing workflows. Being AI-native means the company’s work is captured, queryable by AI, and improved through feedback loops.
Do I need to record every meeting to be AI-native?
No, but record the conversations that feed important decisions, with consent.
How do I capture in-person conversations?
Use a wearable or phone-attached recorder or a phone app, then bring the transcript in with your other context. Tell people you are recording and check local consent laws.
What is a closed loop in product management?
It is a cycle where evidence informs a decision, the decision ships, and the actual result is compared with the expected result, so the next decision is better informed.
Where does Fijord fit in an AI-native stack?
Fijord sits after capture. It turns calls, docs, tickets and metrics into evidence-linked product decisions, then compares predicted impact with actual results after release.