In 1964, a US Supreme Court judge was asked to define obscenity. Potter Stewart said he couldn't. "But I know it when I see it."

The same applies to the term 'AI native' right now.

I literally cannot go on Twitter without seeing those two words. Strange though that if you asked 5 different people what it means, they'd tell you 5 different things.

Maybe in five or ten years we'll have a clear definition. Today there isn't one. There's no described way of "this is how you use AI, this is how everyone does it". Every company is working it out as they go.

It's also not a fixed state. AI isn't fixed. If anything it's completely fluid, so the companies built on it have to keep moving too. What was best practice in March is a bad habit by September.

A friend put it well the other day: if you tell me you're an AI engineering expert, no you're not, because it changed this morning.

Nobody's really an expert. Which is the exciting part. We're in a window where companies get to define this for themselves, and we get to put our mark on what AI native means and how it should work.

We've spent the last year running an AI native company and helping other companies become one. Every client is a new data point. We've now seen enough to know what works and what doesn't.

And most of what works isn't the hype. A lot of AI is cool demos. The stuff that actually matters just functions behind the scenes. The best of it works without you really noticing it's there. Your baseline resets, and you're doing so much less manual work than before that you forget what the old way felt like.

So here's our recipe. The actual pieces that make a company AI native, in the order we'd build them.

1. Clean Context First

Step one is context. Clean data, and a company that AI can actually read.

AI is an accelerant. Bill Gates said it about automation back in 1995: apply it to an efficient operation and you magnify the efficiency, apply it to an inefficient one and you magnify the mess.

AI compounds chaos. It also compounds order. You get to pick which one.

Most companies' knowledge lives in people's heads, Slack threads, old Drive folders and documents nobody has opened in two years. Every AI session starts from zero because none of that is readable. So the first job is mapping where the work and the knowledge actually live, then cleaning it up.

We deploy our Company AI Brain first for this reason. It's where the knowledge, decisions and ways of working live, so AI stops giving generic answers and starts giving yours. But the brain isn't where the work happens. It provides the context for work to be done everywhere else.

AGI is here, and every mistake AI makes nowadays is most likely from a lack of context.

2. Foundation Before Agents

Most companies want to dive straight into an automation. Someone sees a demo, picks a process, and builds a one-off agent for it. Six months later they have five isolated automations, none of them talk to each other, and every one needs maintaining.

Build the foundation first: the shared systems everything else plugs into. For us that's things like email triage, which is usually the first thing a team feels because it gives back hours in the first week, and our skills platform, where everything the team builds lives.

Build them once and every team uses them. That's how it compounds, instead of building a bunch of custom stuff that only one team ever touches.

3. Everyone Works in a Harness

An AI native company is everyone doing their work in a harness. Not a chat window they open twice a week. The place where their work actually happens. For most companies, this is Claude Cowork, Codex, or something custom.

A friend told me yesterday he doesn't even look at UIs any more. He hasn't seen one in ages. He doesn't browse the web. He gets an agent to go and research and do things for him, and he reviews what comes back.

That's where this goes. The interface stops being the website or the app. It becomes the conversation with something that already knows your context.

4. Skills, Context and Models Do the Work

Once the context is clean and people are in a harness, you realise most of the work can be solved with three things: skills, good context and the newest models.

A skill is just a way of working, written down well enough that AI can do it. The best ones come from your best people's real prompts, not from someone guessing what a team might need.

Even the models from the last six months are enough. For most of what a business does day to day, we already have AGI. What holds most companies back now is context and process.

In our experience only the last five or ten percent of the work really needs to live outside of skills.

5. Educate Everyone, Use Champions

You want the whole team educated, not one AI hero. One person who "does the AI stuff" is fragile. When they leave, or get busy, it all stops.

So everyone gets the baseline: they work in the harness, they use the shared skills, and they learn by doing their actual work, not a generic course.

But you also want champions: people who are constantly trying new tools and new processes, because AI keeps changing under you. The team holds the baseline; the champions keep pushing it forward and bring back what works.

And the best part is when it clicks for someone. Teach a man to fish. Once people can build for themselves, they start making things you'd never have thought to build for them. Someone makes a useful skill, someone else finds it and uses it, and the whole company gets a bit better without anyone writing a project plan.

6. Custom Agents Come Last

Then there are the fully custom workflows: the ones that run with no human in the loop. This is what the cool Twitter demo shows you.

They can be a skill on a timer, a proper agent or a plain automation. Whatever shape they take, they come last, once the context, the foundation and the people are in place. Build them first and you're building on sand.

Eventually, if you can set the north star, the goals and the outcomes clearly enough, agents can do a lot of it. But you only get there by doing the earlier steps.

7. Keep Trying Things

AI native isn't somewhere you arrive. The models change, the tools change, the market changes. Something that took a custom build six months ago is now one skill and a good prompt.

The companies that stay AI native are the ones that keep trying new stuff, keep killing what stopped working, and don't get attached to anything they built last quarter.

The Agent Myth

"Agent" can mean anything. It can be a simple automation. It can be Claude Code running a task.

People get caught up in the idea of agents as human-like employees: a digital worker with a job title, doing someone's job end to end. That hasn't really stuck yet, and I think it's worth asking why.

My guess: an agent can never have full, perfect context. There's always something in someone's head, a relationship, a judgement call, a thing that changed this morning. So there's always still a human in the loop.

The human sandwich is still there. A human sets it up, AI does the middle, a human checks it. What changes is that the human layers get thinner as the trust and the context build.

What Didn't Work

  • Pre-building skills for teams based on what we assumed they needed. Almost nobody used them. The skills that stick come from people's real work.
  • Trying to give people a new system to learn. Instead, work around the systems they have, don't fight behaviour change.
  • Asking people to report their context manually. Every attempt failed the same way: people are busy, and nobody fills in the form.

Where This Leaves Us

This is our working model, not a finished playbook. In 6 months it'll look different.

That's kind of the point. Nobody knows exactly what an AI native company looks like yet. But after a year of building one, we're getting a pretty good feel for it when we see it.

If you want all of our learnings on building AI native companies: thehourglass.ai/sauce