Satya Nadella put a name to an AI cost that never appears on the invoice. In a July 12 post, he called it the “reverse information paradox”: you pay for access to a model, then reveal how your company works so the model can become useful.
Most vendor reviews stop at one question: will you train on our data? That question matters. It misses the knowledge created after the work begins—the corrections, accepted outputs, edge cases, and private tests that teach a generic model what “good” means inside your company.
That learning is an asset. Make sure you can walk out with it.
The expensive part is the correction
Take a 40-person agency using AI to qualify inbound leads. On day one, the model knows the obvious signals: budget, company size, timeline, and service fit.
Then the sales lead starts correcting it. This prospect looks small but came through a trusted partner. That one has budget but always stalls at procurement. A third is worth pursuing because the brief matches a capability the agency wants to build this quarter.
After 200 reviews, those corrections contain more value than the original prompt. They encode the agency's real sales judgment, including exceptions that were never written into the CRM or sales handbook.
If those decisions live only inside a vendor's chat history, the company has taught the system without building anything it can carry forward.
The model is the generalist
Nadella makes a useful distinction. The external model is the generalist: broad capability, rented on demand, replaceable when a better option arrives. Your accumulated context, evals, memory, and corrections form the company veteran.
The veteran knows why an output passed review last month. It remembers the customer exception, the risk threshold, and the tone your team will approve without three rewrites. That is the part your company paid to create through staff time and real work.
If changing the generalist wipes the veteran's memory, the durable value ended up in the wrong layer.
Privacy is one check. Portability is another.
A no-training or zero-retention agreement can protect today's request. Good. It does not automatically make the learning around that request reusable elsewhere.
The contract and the architecture need to answer a second set of questions:
- Can you export the memory, traces, feedback, and eval results?
- Can you reuse outputs from your own work to improve a system you control?
- Are accepted and rejected examples stored in your tenant or trapped in a vendor console?
- When you switch models, what company-specific capability moves with you?
Terms differ by provider, product tier, and deployment. Check the product you are buying rather than relying on a broad claim about the vendor.
Build the veteran outside the model
This does not require a research lab. For most SMBs, an owned learning loop is four ordinary pieces connected deliberately.
Keep the source of truth in your systems. Customer context belongs in the CRM. Policies belong in the document store. Workflow state belongs in the application that runs the work. The model receives the slice it needs for the current task.
Capture the reason for a correction. An edited email tells you what changed. “Removed the discount because this is a renewal” tells you why. The reason is what turns a one-off edit into a reusable decision rule.
Turn disagreements into evals. Save 30–50 reviewed examples that represent normal work, awkward edge cases, and expensive failures. Run the same set when a provider ships a new model or when you consider switching. A demo measures surprise. An eval measures fitness for your work.
Put a thin adapter in front of the provider. Keep your prompts, schemas, tool definitions, and routing logic in code or configuration you control. The adapter does not need to be a grand platform. Its job is to make the model replaceable without moving the rest of the workflow.
Run the Monday test
Imagine your current model disappears on Friday. On Monday, what can you hand its replacement?
A useful handover includes the operating instructions, relevant customer context, approved and rejected examples, the eval set, tool definitions, and current workflow state. The new model may behave differently, but it should inherit the company's judgment instead of beginning as a new hire.
If the honest handover is “the team can scroll through the old chats,” you do not own a learning loop. You own a browser history.
Run this test on one workflow before renewing the next annual AI contract. Lead qualification, customer-support triage, proposal drafting, or invoice review is enough. Pick the one where your team makes the most corrections. That is probably where the hidden asset is forming.
Don't turn this into an infrastructure hobby
TechCrunch's report connects Nadella's warning to open models and on-premise deployments. Those can help, especially when sensitive data cannot leave the business. They are not the only answer.
A private cloud deployment with clear contractual rights can preserve the loop. A local model with no feedback capture can still waste every correction. The useful boundary is defined by what leaves, what you can export, and what survives a model change—not by where the GPU sits.
The rule
Models will keep improving, repricing, and disappearing. Your team's learning curve should survive all three.
Rent the model when it makes sense. Own the veteran.