OWNED AI

Owned Intelligence

Own the layer in between

By Jason NeillRevised (with Claude Opus 4.6 and Opus 5.5)

Every AI vendor wants to be your platform. Send us your documents. Connect your CRM. Pipe in your support tickets. In return you get the best model available this month.

For any business with something worth protecting, that's a bad trade. Your most proprietary data ends up shaping someone else's product. Your workflows depend on a model you can't inspect, version or keep. Pricing changes when they say so. If they deprecate the model your process runs on, you start over. And if your workflow turns out to be valuable, the lab can ship it as a feature. Then you're the next plugin.

I don't think that path is sustainable. I also don't think the answer is avoiding frontier models. They're remarkable and some work needs them.

The answer is to own the layer in between.

We call it the proxy. It sits between your people, your systems and every model you use. It's software your business owns and runs. We help build it. It does four jobs.

Privacy. Everything leaving the building passes through it. Sensitive fields get masked or stripped. Some data never leaves at all. Policy lives in one place instead of in a hundred individual prompts.

Routing. Most work doesn't need the biggest model. Classifying a document, pulling fields off an invoice, drafting a routine reply. A small open-weight model fine-tuned on your data does that well, fast and cheap, on hardware you control. The proxy sends hard reasoning to a frontier model when the task calls for it, along with only the context that task needs.

Evals. Every model gets scored on your work. Real tasks with known answers. When a new model ships you find out in a day whether it's better for you, instead of trusting someone's benchmark. Swapping models becomes a config change.

Connections. Frontier labs, SaaS tools and data vendors each become a contract the proxy manages. Keys, limits, logging and cost all live there. You can see what went where and why.

Under the proxy sits the data work. People skip this part, and it's most of the value. Operating data is messy. It lives in an ERP from 2011, a CRM nobody trusts and spreadsheets someone emails around on Fridays. AI-native data engineering means building pipelines for training and retrieval from day one. Clean, versioned and governed. That's what lets you fine-tune a model that actually knows your business.

Put those together and it compounds. Every task that runs through the proxy produces data. Every eval makes routing smarter. Every fine-tune makes your own models better at your work. The frontier keeps improving too, and you get to use it on your terms.

This works at any size. A forty-person distributor can run it. The open models are good enough, the hardware is cheap enough and the tooling has caught up.

Skip the vendor metrics when you measure it. Seats, tokens and pilots launched tell you about activity. Look at the edge where the work meets a customer or a teammate. How long from a problem to a fix. How many people outside engineering shipped something this month. What share of your AI work runs on models you own. What you'd lose if your biggest vendor disappeared tomorrow.

If the answer to that last one is everything, you're renting.

Renting is fine for a while. It isn't a moat.

Or just chat with our Director of Engineering.