The frontier doesn’t matter

If you spend too much time on the internet you might not realize that the frontier doesn’t actually matter. At least not to enterprises, who are driving the majority of the spend for the big labs…and yet still struggle to get even a fraction of the productivity boost from LLMs that elite software engineers regularly achieve.

Yes, the models are very good now: good enough to earn gold at the International Mathematical Olympiad, perform autonomous software coding at scale, and uncover zero-day vulnerabilities before attackers can exploit them. By some estimates, AI has reduced task completion times by 80%. But for who?

Enterprises are getting horribly poor yield from models

Every company – at least in theory – has access to the same models, whether its a hot Silicon Valley startup or a large enterprise or anything in between. But if you look at the concrete business results companies are able to achieve with these identical models, things very much don’t look the same across companies. In fact, AI adoption inside companies is bimodal today:

1. Engineering teams using coding agents are moving faster than ever, ripping PRs from their phones, shipping (good) software at an unprecedented rate.
2. Yet knowledge workers outside of technical teams are not seeing the same type of leverage. Not even close.

The AI-natives, companies who have meaningfully embraced and adopted AI at every level of the technological and sociological stack (if you will), leverage each model improvement into a competitive advantage on how their business runs. Everyone else is left holding the same models, but not the same outcomes.

More practically, AI rollout at large companies typically looks something like this. You get Claude or ChatGPT licenses for everyone, teach people to use markdown, and hold lunch-and-learn sessions. A handful of power users do remarkable things with all this model access. Of course, the reason you rolled out all this AI was to increase revenue, or reduce costs, but neither has happened. Quota attainment is still flat, the support queue is just as long, and the month-end close still takes over a week. The only number that has changed is the increasing token bill.

This difference between the AI-natives and everyone else rests mostly on the ability to use intelligence in ways that produce outcomes for your specific business.

We saw this firsthand as R&D leaders at Cockroach Labs. The models we had access to were more than good enough to get the job done. What they couldn’t do was:

• Connect securely (with the right permissions and governance structure) to all of the tools where they needed to actually do work like Jira, Snowflake, Gong, Zendesk, etc.
• Abstract the repeatable skills and workflows users were creating into shared, cloneable artifacts across the company – so that everyone benefited from what power users were pioneering
• Use these models to build (and securely host) apps that have the built-in RBAC, shareability, and durability that enterprises need by default

Practically, this meant that the 100+ of people that reported to us (and everyone else) were getting a fraction of the productivity gains that they could given the model’s underlying capabilities.

So we built something to change that.

You, too, can be AI-native

Allow us to sketch for you the holy grail of being AI-native. Throughout our conversations with some of the world’s largest enterprises, we’ve heard them describe a similar vision.

It starts off simply: with people doing things they could not do before. A Finance person builds their own profit and loss margin analyzer app, carefully permissioned to the right people on the leadership team. The VP of RevOps has a pipeline analyzer app that updates every time a new piece of information enters the organization in any system. Fewer people spend their weeks chasing status updates.

Next, these personal wins start accumulating and compounding to a system that is shared and built upon. Power users come up with creative and impactful solutions that get shared and distributed to the rest of your team. A renewal agent that one enterprising CSM built to save themselves time can now escape the sandbox and help the entire team, including the new hire that started this week. There will always be power users in your organization; but are they leveling up everyone around them too?

If you continue on this path, the org starts to function differently. An agent that reviews support tickets notices that three customers hit the same edge case bug this week, kicks off a process to draft a fix, and flags it to the right PM and Engineering team for review. There are multiple core loops such as these across the company that can diagnose and spot issues, draw patterns, and suggest fixes with human approvals.

It’s about everything but the model

No amount of frontier intelligence alone will take you to this promised land of an AI-native organization.

The first step is empowering the subject matter expert – every knowledge worker – with the ability to do anything they could conceivably imagine with AI. But this is an enterprise, not the Wild West, and so they need to be able to do this in a governed, properly permissioned, auditable, and secure way. Your team shouldn’t file a ticket to build something and wait for a consultant who doesn’t fully understand the job or an Engineering team that has a long roadmap. People are tremendously creative when given the right opportunity and platform; if the first response to an excited idea is always “maybe in a few weeks” they will quickly stop having these ideas.

The second gap is the shift from personal AI power users to collaborative, shared organizational AI that works inside the messy realities of an enterprise. Whatever works for the AI early adopters needs to make the overall organization better, not just them. This is the difference between one AE being a power user with a skills library, markdown files, and custom dashboards they build for every customer meeting, and an AI-native operating system that the entire GTM runs on. Raise the floor to your highest performer.

In this fantasy world, the model is simply the commodity layer. They are getting better and will continue to do so for the foreseeable future. What matters is the stuff on top, your organizational processes, context, security, governance and exceptions. These are the competitive advantages that compound over time, even as models underneath get swapped.

Introducing Tiles

Back to our time at Cockroach Labs. Faced with this increasing organizational morass and a future we wanted where our colleagues could achieve their AI dreams, we built something called Mica. It simply (well, not so simply under the hood) did these things for our teammates.

We did not expect what happened next, which is that 85% of the company began using Mica every day within a week of launching it. They created thousands of apps, many of which now are core operational pieces of how Cockroach runs, within the subsequent 12 weeks. Mica turned Cockroach from a place where individuals used AI, to a place where the organization ran on AI.

In fact, it went so well that it was clear more companies than just Cockroach needed a system like this. So today we’re launching Tiles, the organizational AI platform for enterprises. Tiles enables companies to move from personal AI to a shared organizational AI system. We believe it is the missing link (plus a lot of hard work) for enterprises to actually become AI-native.

• Build production-ready apps without filing an Engineering ticket. Describe the app you want (a margin analyzer for Finance, a live pipeline tracker for RevOps) and get a working one. Vibe coding, minus the Wild West: the database, hosting and RBAC come already built in, and every app runs with the viewer's own permissions, so nobody has to wait on a security review.
• Put agents to work on a schedule. Anyone, not just engineers, can describe an agent in plain English: a Monday pipeline digest, a weekly renewal-risk check. It shows up wherever you want it to: Slack, email, or in the Tiles product. Each agent only touches the tools it was built for, and follows the governance that is defined at the enterprise level.
• Discover and propagate the best work. Every team gets a kit of its official skills, agents and apps, along with its context and how it works. When someone figures out something good, it becomes everyone's baseline in one click, and the new hire can inherit the entire playbook on day one. There will always be power users; this is how they level up everyone around them.
• Enterprise-grade security, whatever model is underneath. Pre-built integrations, from Salesforce to HiBob, run with each person's own permissions or Service Accounts. Admins decide what requires a human sign-off by the team, down to the tool call level. Tiles runs wherever you want it: in your VPC, encrypted with your own keys, or on AWS Bedrock, Vertex, or using your own API keys. The model is now the commodity layer, so swap in any frontier model or use a combination of open-source and frontier models based on the task at hand.

The next set of model improvements are coming either way, and everyone will have it. What's still up for grabs is whether your company can turn it into a different way of working. The frontier was never the hard part, come see the rest.

If you are interested in transforming your company’s operations, we’d love to hear from you and share what we’ve built. Sign up for a demo here!