What Microsoft’s diagnostic study admitted
In October, Mustafa Suleyman, who runs Microsoft’s AI division, said the company’s diagnostic system now beats doctors on hard cases, and called it a step towards “medical superintelligence”.[1] The more useful detail was further down: most of the performance came from the underlying model rather than the orchestration layer Microsoft built around it.[1]
That’s a vendor admitting in public that the clever wrapper mattered less than the engine inside it. If your firm buys “AI-powered” paraplanning, compliance monitoring or client onboarding tools, it tells you where the value in those tools sits.
Ask which part of what you’re paying for still has value after the model underneath gets better.
A demo is built to show that the tool works. Whether it keeps its value is the question that protects your money.
The wrapper economics the demo hides
Most AI products resell someone else’s model. The vendor licenses it from an outside lab, wraps it in an interface and sells you the result. That’s how the market works, but it means part of what you pay for is model capability the vendor rents, and anything rented can be offered to everyone else next quarter.
It has already happened. The model labs built document-style workspaces and persistent canvases straight into their own products, and a batch of startups that had charged for exactly that feature lost their edge overnight. The capability was still there, but nobody needed to buy it separately any more.
Palantir’s Alex Karp has argued against what he calls the “token industrial complex”, meaning the habit of handing control to model labs rather than keeping it in-house. As one AI founder, Dylan Arie, put it, the bar used to be subject-matter expertise, and now the models have the subject-matter expertise. The thing you used to pay a wrapper for is increasingly sitting in the base model itself.
What keeps its value when the base model improves
Some of what a wrapper adds lasts and some disappears, and the job is telling them apart.
In fairness, orchestration, evaluation and guardrails are real engineering. When an AI project fails, the model is rarely what let it down. More often those layers were never built properly. NVIDIA made the same point from the other direction, showing that tuning the scaffolding around a model, the data pipeline and the orchestration, beat simply swapping in a stronger one.[2] Good scaffolding is worth paying for.
Durable value is the part that doesn’t care which model sits underneath: integration with your CRM and back office, guardrails tuned to your compliance process, your own firm’s data, a workflow your team actually uses, a conductor layer you can point at a better model when one ships. That lasts. What disappears is raw capability dressed up as product, the bit that is really just the model doing what the model does.
The one question that beats the demo
The test you can run live, with no engineering background: ask the vendor to name every capability in the product that exists only because a particular model can do it, and to tell you what remains if that capability ships natively in the base model next quarter.
A good vendor will have an answer, because they’ve already thought about their own defensibility. A weak one will get uncomfortable, and that tells you what you need to know.
You can check their answer against four things, none of which needs technical skill. The model pages tell you which model they run. The changelogs tell you how often the thing underneath changes. The subprocessor list tells you whose model is actually processing your clients’ data. The data processing agreement tells you what happens to it. Forty-five minutes is enough to ask for all four and read the vendor’s face while they answer.
How to price an edge that might disappear
Price the lasting part and the rented part separately, and pay for each accordingly.
- Separate the wrapper value from the commodity capability. Work out what you’d still be paying for if the model vendor gave the raw capability away for free. That number is what the tool is worth. The rest is rented, so price it as a rental.
- Favour model-agnostic tools. A tool whose conductor layer can be pointed at a different model protects you when the market moves. One welded to a single model hands its roadmap to a lab you don’t control.
- Keep the term short. Lock-in is something you negotiate. Enterprise buyers already switch between models freely as new ones arrive,[3] so a vendor who insists on a three-year term is asking you to carry a risk they won’t. Treat switching cost as a line item to minimise at signing. Don’t let it become the reason you overpay now.
- Prefer outcome-based pricing to per-seat or per-credit. IDC expects 70% of software vendors to move away from per-seat pricing within four years.[4] Credit-metered models make your bill unpredictable and punish you for using the thing. Pay for output you can measure against the cost of the work it replaces. (Our list of questions to ask in an AI demo has more to put to a vendor.)
Start with your next renewal
You don’t need to rip anything out this week. Pull the contract that renews next, and before you re-sign, ask the one question: which part of this still has value after the model underneath gets better? Price the lasting part, treat the rest as rented, and shorten the term so you’re not paying this year’s premium for something that will be standard next year.
This article is for informational purposes only and does not constitute regulated financial advice or a compliance opinion. Consult a qualified compliance professional for advice specific to your firm.
Sources
[1] TechCrunch, ‘Microsoft’s Mustafa Suleyman says its AI outperforms doctors on hard cases as he chases medical superintelligence’, 22 October 2026. Available at: https://techcrunch.com/2026/10/22/microsofts-mustafa-suleyman-says-its-ai-outperforms-doctors-on-hard-cases-as-he-chases-medical-superintelligence/ [Cited for the base-model concession in Microsoft’s diagnostic study.]
[2] TechCrunch, ‘NVIDIA just showed that the harness, not the AI model, is now the real hero’, 21 August 2026. Available at: https://techcrunch.com/2026/08/21/nvidia-just-showed-that-the-harness-not-the-ai-model-is-now-the-real-hero/ [Cited for the value of the orchestration harness over raw model gains.]
[3] TechCrunch, ‘OpenAI is gaining on Anthropic with business users, new data indicates’, 20 August 2026. Available at: https://techcrunch.com/2026/08/20/openai-is-gaining-on-anthropic-with-business-users-new-data-indicates/ [Cited for enterprise buyers switching between models freely.]
[4] IDC (2026) market prediction on software pricing models. [Cited for the forecast that most software vendors will move away from per-seat pricing within four years.]