The Layer Between Your Business and the AI
Everyone is buying AI. Almost nobody can prove what it earned them. And the question serious buyers are finally asking out loud: can it steal my IP?
Two problems, one root
In July 2026, Palantir CEO Alex Karp sat down on CNBC and described what enterprise leaders tell him in private:
"I am paying for tokens that create no value, and these people are stealing the weights and alpha of my business."
Alex Karp, CNBC Squawk Box, July 2026
Karp runs a company that profits when you believe that, so discount accordingly. But strip the theatrics and he is describing two problems any leadership team that bought AI in the last two years will recognize.
The first problem is unprovable value. The subscriptions got signed, usage is high, and the CFO still cannot find the saved dollars. Here is a tell worth keeping for every vendor meeting: watch how the seller prices. A seller who can prove an outcome prices the outcome: a percentage, a success fee, a share of the savings. A seller who cannot prove an outcome meters usage, the way the electric company bills kilowatt-hours. AI is sold by the token, which is a meter. The pricing structure is the industry admitting it cannot yet trace its product to your results. That is not a reason to avoid AI. It is a reason to stop expecting the model alone to produce results.
The second problem is exposure. Every prompt your people send to a hosted AI service carries a piece of how your business works: pricing logic, cost data, client names, the judgment calls that make you better than the shop down the street. Karp's word for that accumulated edge is "alpha," and his warning is that it flows out of your building one question at a time. To be fair to the model companies, the major labs state in their enterprise contracts that they do not train on customer data. The honest version of the problem is not theft. It is verification. You cannot see what happens to what you send, and for some organizations, unverifiable and unacceptable are the same word.
The missing piece
A raw AI model is a brilliant hire on day one. No login, no files, no history with your customers, no idea what your business does or which of its numbers are sacred. Nobody would judge a new hire's value on day one, and nobody should be surprised that a raw model connected to nothing produces nothing measurable.
What makes the hire useful is everything you wrap around them: access to the right files and not the wrong ones, the rules of the company, the meaning of the words your industry uses, someone checking the work. For AI, that wrapper is software, and it has a name: the application layer. It sits between your data and the model. It decides what the model is allowed to see, in what structure, with what permissions. It turns the model's raw ability into your formats, your reports, your workflows. It keeps a record of what happened.
Palantir calls its version "ontology" and charges Fortune 500 prices for it. The name does not matter. The ownership does. Whoever builds the application layer makes the decisions that count: what the model sees of your business, where your data is cached, whether your cost book ever leaves the building, and what it costs you to switch models when a better one ships next quarter. If a vendor owns that layer, those are the vendor's decisions. If you own it, the model companies become what they should have been all along: suppliers. Interchangeable, replaceable, on your terms.
The four questions
Before your company signs anything with AI in it, ask the seller four questions:
Who owns the data?
Not who stores it. Who owns it, and what are they allowed to do with it?
Where is it cached?
Data you sent last quarter may still sit on someone's servers. Where, for how long, and who can read it?
Are the prompts secure?
The questions your people ask describe your business as precisely as the data does.
Can this vendor enter my business?
If they can see enough to serve you, they can see enough to compete with you. What stops them?
The four questions are Karp's, from the same interview, and his sharpest observation is that most buyers never ask them. Ask them of everyone, including us. A vendor who cannot answer in plain writing has answered.
How MPX approaches it
Three commitments.
We work with leadership, not around it.
Who sees your data, who owns your systems, and what your switching costs are: those are ownership and risk decisions, not software features. They belong to the CEO, the board, and counsel, not to a procurement checklist. An MPX engagement starts in the leadership room, and the first deliverable is a written data boundary: which parts of the operation may flow to hosted models, under what terms, and which parts never do.
You own the application layer.
We design and build the layer between your operations and the model so that it belongs to you: documented, inside your walls, independent of any one AI company. Models get swapped underneath it as better ones ship. The layer, and everything it knows about your business, stays.
We run our own operations on this.
MPX is building this practice the way we build everything, on companies we own first. A design firm, a construction business, real estate entities: their month-end close, invoicing, and project reporting run through application layers we built and have to live with. Nothing we would propose to you is theoretical, and where we have not proven something on our own books yet, we will say so in the first meeting.
The next step
One email: mike@mpxgroup.net. Describe the operation, the part of it that runs on too few people or too little information, and what you are afraid of exposing. Do not describe the AI.
The first conversation is about your business. The layer comes second.
That is the point.