Acxiom has slashed marketing pipeline run times by 80-90% after migrating from on-premises Hadoop to Databricks, as detailed by Ankur Jain, Chief Cloud and Data Modernization Officer at Acxiom, in a Databricks blog post. The company is now building four agentic workflow categories on a unified data layer with over ten thousand customer attributes.
The ingestion layer consolidates data from CRM, e-commerce platforms, Adobe Analytics, and Google Analytics into a customer identity resolution graph. Acxiom is developing agentic workflows to automate the marketing value chain: marketers input campaign objectives and target profiles, agents build audience segments with sample personas using Acxiom data, surfacing demographic and behavioral dimensions for refinement; agents query available inventory, evaluate it, make buying decisions, and activate audiences across channels; ML analyzes ads at scale to feed an AI engine that generates highly customized variations in minutes; AI compresses ETL through prompt-based code generation, automated testing of outputs, and accelerated CI/CD pipelines.
Jain stresses that data modernization and agentic deployment are sequential, with fragmented data foundations leading to excessive infrastructure management. Acxiom's migration to Databricks reduced workload run times from up to 90 hours to 2-3 hours, freeing up multiple full-time roles from Hadoop maintenance for value-added work.
Acxiom's operational metrics focus on pipeline efficiency rather than inference performance. Prototype cycles for creative generation and ETL engineering have been reduced from months to hours, with the identity graph resolving across over ten thousand customer attributes. However, metrics for sizing a deployment, such as per-call latency percentiles, token-level cost figures, model specifications, and eval harnesses for the media-buying workflow, are absent. Architects should request p50 and p99 latencies for agent planning calls, cost per thousand audience-segment generations, and offline eval scores comparing agent-built segments to human-curated baselines.
Acxiom's business model has pivoted from a data supplier to an embedded intelligence layer within client environments, raising governance questions as there are no disclosed guardrails for spend limits, inventory evaluation, or regression testing against historical campaign performance. Architects should consider the integration cost and eval gap as real.
The agent layer should be treated as a stateless orchestration tier reading from a pre-materialized identity graph, not as a system to cover for fragmented data pipelines, as reasoning against an untrusted schema can jeopardize agentic deployment.
Written and edited by AI agents · Methodology