Customer data: the hard part hasn’t moved

1H 2026 · Finding 3 of 5

An agent is only as good as the customer data it works from, and it carries that data’s mistakes to every customer it touches.

The leading obstacles are organizational and semantic, and they recur in every survey wave

42.6%name poor cooperation across theorganization as an obstacle41.2%lack a data model sharedacross systemsMembers reporting organizationaltrouble, by survey wave85.7%202189.5%202281.5%202542.6%name poor cooperation acrossthe organization as an obstacle41.2%lack a data model sharedacross systemsMembers reporting organizationaltrouble, by survey wave85.7%202189.5%202281.5%2025

Left: the two leading obstacles in the current Member Survey (Table 22). Right: members reporting any organizational trouble in the 2021, 2022 and 2025 waves (§13).

What has been said

AI agents, or a move to the warehouse, will make the work of resolving and defining customer data unnecessary.

What the evidence shows

The hard part hasn’t moved. Moving the data into a warehouse doesn’t solve either, and neither does adding an agent on top.

The leading obstacles are organizational and semantic, and they recur in every survey wave.

Getting teams to work together leads at 42.6% of members, with the absence of agreed data definitions close behind at 41.2%; technology staff time or skill is third at 32.9%, and executive backing comes last of ten. Data and organizational difficulties appear in all three survey waves, 2021, 2022 and 2025.

Few vendors document the capabilities that would address them.

Twenty-six of the 213 vendors document a semantic layer, the software that holds shared data definitions, and the definitions it holds still have to be agreed and owned inside the organization. Under a third of own-store and hybrid vendors document consent and data rights.

An agent inherits the data it works from.

An agent that works directly from unresolved, undefined or unpermitted data inherits those problems. Members’ concerns about AI run to security and privacy (63.0%), the quality of training data (50.0%) and whether results can be trusted (48.6%).

63.0%security and privacy50.0%quality of the training data?48.6%whether results can be trusted63.0%security and privacy50.0%quality of the training data48.6%whether results can be trusted

That supports the CDP’s purpose, and cautions against expecting AI to remove the work.

That supports the CDP’s purpose, which is to resolve, define and govern customer data. It is also a caution for anyone who expects AI to make that work unnecessary.

What this finding does not establish

Answer options and respondents changed between survey waves, so individual obstacles cannot be compared directly across waves; that difficulties remain widespread holds (§13, Appendix B).