Agentic CDPs: the claims have run ahead of the controls
More than half the vendors we track now claim at least one agentic capability. Only six document all three.
Most vendors claiming agentic capability document no way to govern it
Each square is one of the 213 active vendors. Documentation measures, not deployment performance (§12).
Products are now marketed as “agentic CDPs”, and one of the largest data platforms markets one.
The claims have run ahead of the controls. The agentic CDP is not yet a product class.
Agentic claims now span more than half the roster.
111 of the 213 active vendors claim at least one of the five agentic capabilities. Most claim only part of the set: most often a conversational assistant, least often agent governance.
Vendors claim the ability to act far more often than the ability to check.
Against those 111 claims, 33 vendors document agent governance, and nine, 4.2% of the roster, document any way to evaluate agent output. Six document all three kinds of control: what an agent may do, what record it keeps and how its output is judged. Size barely narrows the gap: among the 32 vendors whose controls could be scored, those with 1,000 or more employees document 2.00 of the three kinds of control on average, against 1.63 below 100.
Much of the agentic label is agent-washing.
Of the 111 vendors making an agentic claim, 42 document no way to build an agent. For 24, the claim rests on an assistant or a connection protocol alone, with nothing that builds, grounds or governs an agent. A label is cheaper than a control layer.
No independent mechanism verifies agentic claims.
RealCDP tests the data foundation but none of the five agentic capabilities. The Institute intends to extend RealCDP certification to agentic capabilities. Until it does, buyers are choosing on claims.
Value is shown in particular settings; durability is untested.
Published field studies show operational benefits in settings such as marketing messages and customer service, and members report positive satisfaction with the AI applications they use. Success has to survive reruns, hand-offs and model changes. Lasting value from customer-data agents has been neither demonstrated nor ruled out.
These counts record what vendors document about their offerings, not installed functionality or independently verified performance (§12).

