No vendor owns the entire intelligence layer. They all own components. The question is which components, how well, and what's missing. A practical evaluation maps each vendor's claims against the four core components of an intelligence layer (semantic foundation, agent orchestration, trust and evaluation, operating model integration), then asks five specific questions that separate vendors with substance from vendors with positioning.
The four vendor categories, and what each actually offers
Data platform vendors (Snowflake, Databricks, Microsoft Fabric, and similar). They own the data storage and processing layer underneath your intelligence layer. They are adding semantic and agent capabilities on top of that strength. What they really offer is integration with your existing data infrastructure. What they don't own is agent orchestration at the level a real intelligence layer requires, or evaluation infrastructure, or operating model design.
Orchestration startups (LangChain, CrewAI, AutoGen, and the many smaller players). They own the agent coordination layer, sometimes well. What they really offer is the engine room of agent reasoning and tool use. What they don't own is your data, your semantic foundation, or any of the governance and evaluation work that has to sit above their layer for the system to be reliable.
BI incumbents pivoting toward AI (Tableau, Power BI, Looker, ThoughtSpot, and similar). They own the human consumption surface for analytics. They are bolting chat and agent features onto that strength. What they really offer is a familiar interface for existing users. What they don't own is the semantic foundation an agent consumer needs, the orchestration tier, or anything resembling an operating model for agent reliability.
Agent-specific platforms (the newer wave of vendors selling "your AI agents, deployed and managed"). They own a slice of orchestration plus a slice of evaluation, usually with strong demos. What they really offer is a faster path to deploying agents than building from scratch. What they don't own is the semantic foundation underneath those agents, or integration with your existing operating model.
Every vendor in every category is currently claiming "we are your intelligence layer." Roughly none of them are. They are each one component of it, dressed up to look like the whole.
Five questions to ask in every evaluation
The first question. Which components of an intelligence layer do you own end-to-end, and which do you integrate with from external systems? This is a yes-or-no question, and any vendor unable to answer it cleanly is a vendor obfuscating which components they actually own. A good answer names the components owned (with specifics) and the components integrated.
The second question. What does our specific case (name a real production case you care about) look like running on your platform, including the parts that touch components you don't own? This forces the vendor to describe the architecture honestly, including the gaps. Vendors with substance answer this concretely. Vendors with positioning give you a demo.
The third question. Which customers in our industry have deployed this in production, and what specifically did they have to build or integrate themselves to make it work? Reference customers exist for almost every vendor at this stage. The vendors with real production deployments can answer this in specifics. The ones without will give you logos of customers who ran pilots, which is not the same as production.
The fourth question. What is your story for [the component you don't own]? A good vendor has a clear position on the components outside their core. They have partnerships, they have recommended patterns, they have integration documentation. A weak vendor handwaves at this, or worse, claims they own it when their architecture clearly doesn't.
The fifth question. How do we end up owning the intelligence layer, rather than renting it from you? This is the question that separates a vendor relationship from a dependency. The answer should be specific. Architecture diagrams that show your team owning the design decisions. Documentation patterns that survive the vendor going out of business. Data exports that work. If the answer is "you build on us forever," you are evaluating a lock-in product, not an intelligence layer.
The pattern that emerges
A Gartner analysis on Agentic AI vendor landscape projects that through 2027, vendor positioning in agentic AI will outpace vendor capability by a factor of three to four. Translation: vendors are claiming roughly four times what they can deliver. Your evaluation framework has to assume this gap exists, and the five questions above are how you find it.
The data leaders who navigate this well are the ones who refuse to let any single vendor define their architecture. The ones who get this wrong end up with an intelligence layer that is, structurally, just one vendor's product wearing a strategic label.
If this resonates, subscribe. And share it with the colleague currently negotiating a contract right now.
— Kyle