Satya Nadella published an essay on July 12 that dominated tech conversation for the rest of the week: the “Reverse Information Paradox.” His argument, in short: when your company uses AI, you pay twice. Once with money, for the subscription or the API usage. Again with something more valuable, the proprietary knowledge you hand over every time you feed a model your data, your corrections, your way of doing things, just to get it to work well. Nadella put it directly: “You essentially pay for intelligence twice, once with money, and again with something even more valuable: the proprietary knowledge you must reveal to make that intelligence useful.”

His fix is a “trust boundary”: a wall inside each company’s own environment that nothing crosses without explicit consent. Not prompts, not evals, not organizational memory, not the model itself once it’s been adapted to that company’s data. In practice, he’s arguing every enterprise should own and train its own model.

Two credible voices in enterprise tech pushed back within days. Aaron Levie, CEO of Box, quoting Sourcegraph co-founder Jesse Zhang. And Zhang himself, in a separate post. Both make the same uncomfortable point about Nadella’s framework: training a model per enterprise sounds clean in an essay, and it’s a lot harder, and more fragile, in practice.

This looks like an infrastructure debate with no relevance to marketing. It maps directly onto a decision your own team is making right now.

What your marketing team actually needs to protect isn’t a model

Here’s the argument stated plainly, because we made this case before this debate existed: the real asset in your marketing operation was never a model trained on your data. It’s the criteria.

What makes a campaign brief actually complete at your company, not in the abstract, at yours specifically? What criteria does your team use to decide which creative variant gets the media budget before a single dollar goes to paid spend? What defines a valid audience segment for your business, beyond the generic demographic buckets any ad platform hands you? What, in concrete terms, counts as a qualified lead for you, not the sales-playbook definition?

Those four answers almost never get written down. They live in the head of whoever has been on the team longest, or in a couple of campaigns that “worked” that nobody documented the reason for. That criteria, once it’s written and turned into something measurable, is your advantage. Not a model that only your data team knows how to maintain and that has to be retrained every time a new version of Claude or GPT ships.

The disagreement, in the participants’ own words

Levie, quoting Zhang: “The biggest challenge right now with the topic of every enterprise having their own model is that your most valuable information and insights are not only always changing, but they’re often your most sensitive information… you can’t keep your security layer inside the model or an agent… training a model per enterprise is going to be a lot harder than it looks.”

Levie isn’t dismissing custom-trained models outright. He expects 100x more use cases for them, mostly inside domain-focused products built by vendors, not inside every buyer’s own IT department. What he’s questioning is the generalization: that training your own model should become standard practice for every enterprise, including yours.

Zhang, in his own post: “Your best bet to make that IP useful is to turn it into skills or artifacts that models can use in-context… it is irreversible… when things change, you cannot untrain what you did.”

That’s the line that lands directly on marketing. Your most valuable information (what actually worked in last quarter’s campaign, how your real audience responds, which subject line moves open rates) changes every quarter. Baking it into a trained model’s weights is an expensive bet that’s also hard to reverse. When the criteria changes, or a better model ships next year, you can’t untrain what you already put in. You start over.

We already answered this, before the debate existed

This isn’t a new position for us. It’s the same thesis we’ve made twice on this blog, before Nadella’s essay was published.

In evals as strategic IP, we quoted this same Aaron Levie saying that nearly all progress in AI agents comes down to evals, not the model underneath them. It’s the identical argument he’s now making against Nadella: the criteria that defines what a good output looks like for your business belongs to you, travels across any model you choose to run, and compounds with use.

And in the prompt is temporary, the eval is permanent, the argument was structurally the same one Zhang is making now. The wiring around a model (the prompts, the way you glued a workflow together) depreciates with every model update. What compounds is the standard that defines “good” for your process. Zhang calls training a model irreversible. We had been calling it depreciation. Same failure mode, viewed from two different angles.

What this looks like inside a marketing team

Translated into your actual operation, the criteria worth owning and compounding takes this shape.

The campaign brief. Not the blank template you fill out each time, but the written criteria for what makes a brief complete at your company: what information can never be missing, what questions need an answer before anything moves into creative production. That’s what a new model, or a new hire, needs to read to produce work at the level of your best person.

Creative variant scoring. Before media spend goes out, what criteria does your team actually use to decide which version of an ad launches first? If the answer is “whichever looks best” or “whichever the creative director liked,” that’s an opinion, not a criterion. A written standard, what elements predict performance in your category, with your audience, on your channel, is something you can apply consistently and sharpen with every campaign.

Audience segment definition. Not the generic segmentation any ad platform ships by default. How your company validates, with your own data, that one segment actually behaves differently from another. That doesn’t live inside any trained model. It lives in your CRM, your campaign history, and the criteria your team applied to draw the line.

What counts as a qualified lead. Not the sales-playbook definition. The specific combination of signals that, in your sales cycle, separates someone ready to talk to sales from someone who’s just browsing. Written down and applied consistently, that definition is what stops marketing and sales from arguing about lead quality every single month.

None of those four require training a model. They require your team sitting down to write, for the first time, something it probably already knows how to do but never put on paper.

What Nadella gets right, and what his fix doesn’t solve

Give Nadella his due: he’s right that your company’s proprietary knowledge is valuable, and that giving it away without noticing is a real risk. Where the argument breaks down is the mechanism. His proposal assumes your company has the team, the budget, and the patience to build and maintain model-training infrastructure. Most marketing teams we work with have none of that, and don’t need it.

Written criteria doesn’t require that infrastructure. It runs the same way against Claude, GPT, or whatever ships next year, because it isn’t locked inside anyone’s weights. And unlike a trained model, you can edit it the same day your strategy changes, with nothing to untrain.

This connects to something we wrote earlier about why your AI marketing needs a verifier, not a smarter model: the model is a commodity, the layer of criteria that reviews and decides around it is yours.

If your team ships AI-assisted content and campaigns every day but couldn’t hand you those four criteria in writing right now, that’s the work worth doing before any other AI investment for marketing. This isn’t a pitch. It’s the question that has to get answered before any other tool purchase is worth making.

Write the criteria before you scale AI marketing