When you hire a management consultancy, you assume the advice is impartial. You are paying for independent judgement, not for someone to sell you a product. Most buyers of AI consultancy assume the same thing. A significant number of them are wrong.
The incentive structure most firms don't disclose
The major cloud providers — Microsoft Azure, Google Cloud, Amazon Web Services — have built extensive partner ecosystems around their AI products. Consulting firms that operate inside those ecosystems earn revenue in several ways: referral fees when a client signs up for a cloud service, margin on consumption when they resell cloud credits, or co-sell agreements that tie the consultancy's growth targets to the platform's growth targets.
None of this is secret. The partnership tiers are published. The commercial arrangements are legal. The problem is that they are almost never disclosed to the client being advised — and they create a structural bias toward recommending the platform the consultancy profits from, regardless of whether it is the right fit for the client's problem.
This matters more now than it did two years ago. The AI tool landscape in 2026 is genuinely fragmented. Claude, GPT-4o, Gemini, Mistral, and a growing range of open-source models all have meaningfully different strengths. The right architecture for a KYC document processing pipeline is not the same as the right architecture for a conversational agent or a real-time analytics system. A firm with a financial interest in one platform will almost always find a reason to use that platform, even when a better option exists.
Three questions that reveal where a consultant sits
You do not need to audit a firm's commercial relationships to understand whether they have one. Three direct questions will tell you almost everything.
The first: do you have a preferred AI provider? An honest answer is something like "we use different models for different jobs." A consultant who answers with a specific platform without qualification — "we are an Azure AI partner" or "we build everything on Google Cloud" — has told you exactly where their recommendation will land before they have heard a single detail about your business.
The second: do you earn revenue from AI model subscriptions, cloud consumption, or vendor referrals? Most firms will not volunteer this. Asking directly forces a direct answer. A firm that earns margin on the tools they recommend has a conflict of interest that should be disclosed before they advise you on tool selection. If the answer is yes, ask them to separate the advisory fee from the product revenue so you can evaluate the recommendation independently.
The third: what would you recommend if my cloud contracts are already committed to a specific provider? Watch what happens to the recommendation. A genuinely model-agnostic firm will give you the same answer regardless of your existing cloud commitments, because their advice is based on your problem, not on which platform they need to bill against. A firm that immediately reorients toward your committed cloud has revealed that the recommendation follows the commercial relationship, not the technical requirements.
What model-agnostic looks like in actual practice
Model-agnostic is not a commitment to indecision or religious neutrality about tools. In practice, it means that the model is treated as a component — one that can be selected, configured, and replaced based on what the job requires — rather than as a fixed constraint that shapes every other architectural decision.
In a production system this looks like: Claude Sonnet for multi-step reasoning and document analysis where accuracy matters most, GPT-4o for structured extraction tasks where speed and cost efficiency are the priority, and open-source models running locally for workloads involving sensitive data that cannot leave your infrastructure. The choice is made per task, per risk profile, and per cost constraint. Not per vendor relationship.
The architecture that supports this is designed so changing the model underneath a workflow does not require rebuilding the workflow itself. Your business logic, your data pipelines, and your integration layer all sit above the model selection. When Anthropic repriced its agent usage tier in May 2026 and disrupted workflows across the industry, clients whose infrastructure was model-agnostic simply adjusted the routing. Clients whose infrastructure was built around a single provider had to rebuild.
Why this matters for what you pay over time
The commercial argument for model-agnostic architecture is straightforward. AI model pricing is not stable. Capabilities are not stable. The model that offers the best price-to-performance ratio today will be undercut or outperformed within 12 to 18 months, based on the rate of change we have seen since 2024.
If your AI infrastructure is locked to a single provider, every one of those changes becomes your problem. Price increase: absorbed. Capability gap: worked around. Provider outage: downtime. If your infrastructure is model-agnostic, those changes become routing decisions. The business logic stays the same. The provider underneath gets swapped.
The consulting firm that builds your AI system will not be the one absorbing the cost of that lock-in. You will. Which is a reasonable argument for asking, before they start, whose interests the architecture is actually designed to serve.
A note on where Kelriva sits
We do not have commercial partnerships with AI providers. We do not earn margin on model subscriptions or cloud consumption. Our revenue comes entirely from the fixed fee we charge to design, build, and hand over AI systems. That structure means the only thing that influences our tool recommendations is what works best for your problem.
We use Claude, GPT-4o, Mistral, and open-source models across different client systems depending on what each task requires. We build on AWS, but we are not an AWS partner and do not earn from that relationship. When a better infrastructure option exists for a specific client, we use it.
We mention this not to market ourselves but because it is the direct answer to the question this article raises. If you are evaluating AI consultancies, apply the same questions to us. The answers should be verifiable.