Building an Enterprise-Wide Generative AI Platform for HR and Payroll Services

adp · HR · 2023 · source Agentic: no · promotional bias: med · source-backed · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Enterprise HR and payroll GenAI platform

ADP needed to make generative AI usable across a large HR/payroll business while protecting IP, ensuring answer quality, structuring enterprise data, and controlling the cost of model use.

Constraints:

The company operates at global payroll scale across many countries and workers.
Quality, IP protection, data structure, and ROI were presented as core blockers.
Different models have widely varying costs.
The platform had to connect governed enterprise data with model serving and RAG capabilities.

Architecture

ADP describes a One Data and One AI platform: governed data foundations using tools such as Delta Lake, plus AI platform capabilities such as model serving, LLMOps, vector search, and RAG. The platform supports ADP Assist and other generative AI services over HR/payroll data while focusing on quality, IP protection, structured access, and cost control.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  adp-users[ADP Users] --> adp-assist[ADP Assist]
  adp-assist --> one-ai-platform[One AI Platform]
  one-ai-platform --> one-data-platform[One Data Platform]
  one-data-platform --> delta-lake[Delta Lake]
  one-ai-platform --> model-serving[Model Serving]
  one-ai-platform --> vector-search[Vector Search]

Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component	Role	Interacts with
`ADP Users`	Employees or customers using HR and payroll AI experiences.	`ADP Assist`
`ADP Assist`	Generative AI assistant experience over HR and payroll workflows.	`One AI Platform`
`One AI Platform`	Model serving, LLMOps, vector search, and RAG layer.	`One Data Platform`, `Model Serving`, `Vector Search`
`One Data Platform`	Governed data foundation for HR/payroll data.	`Delta Lake`, `One AI Platform`
`Delta Lake`	Data lake foundation used in the data platform.	`One Data Platform`
`Model Serving`	Serves generative AI models behind applications.	`One AI Platform`
`Vector Search`	Retrieval layer for RAG over enterprise data.	`One AI Platform`

Data flow

Enterprise HR and payroll data flows into governed data foundations. The AI platform layers model serving, LLMOps, vector search, and RAG over that data so ADP Assist-style applications can generate answers while managing quality, IP protection, and cost.

“instance um tools like Delta Lake are broadly used in in we data um and of course there's a governance piece tied to this which has to do with catalog the catalog which helps us with other areas like observability and like all all the”transcript, stated

“capabilities like model serving lops uh Vector search and so on so forth right that what you're seeing here of course this is a this is a range of tools that that cover other technology firm as well but you know in in in data breaks we”transcript, stated

Why this architecture

ADP framed the architecture around a governed data platform plus an AI platform because generative AI quality depends on structured enterprise data, IP protection, model serving, RAG, and cost controls rather than model access alone.

“um how to achieve quality huge huge problem how to protect IP how do we access and structure data and how do we control the cost to achieve Roi these are like four topic areas that we are discussing uh across the board so for instance what is what does it mean to”transcript, stated

“pieces of this uh one data and onei platform now I can say that you know for instance um tools like Delta Lake are broadly used in in we data um and of course there's a governance piece tied to this which has to do with catalog the”transcript, stated

Models used: LLMs · generative AI models

Scale

ADP operates in 140 countries.
“in 140 countries um with 41 million over 41 million global workers that are paid through our platform and we're the number one HCM application HCM stands for human Capital Management so you can”transcript, stated
ADP pays over 41 million global workers through its platform.
“41 million global workers that are paid through our platform and we're the number one HCM application HCM stands for human Capital Management so you can imagine that you know when you work with ADP you're working with h a a pretty”transcript, stated
ADP reports over a million clients.
“our system uh we have over a million clients so pretty large scale uh we are in 140 countries um with 41 million over 41 million global workers that are paid through our platform and we're the number one HCM application HCM stands”transcript, stated

Alternatives considered (roads not taken)

Treat model access as the whole solution.

Rejected because: ADP identified quality, IP, data access, structure, and cost as core platform concerns.

Ignore model-cost variation.

Rejected because: The transcript emphasizes that generative models have a large cost range.

“llms not all generative created equal and there's a huge cost range there so how do we control cost right that's one of the the topics that really becomes one of the challenges that that bring to the table and in the conversation with”transcript, stated

Separate data and AI foundations.

Rejected because: The presented architecture pairs One Data with One AI as the operating platform.

Patterns

`rag-retrieval` (enterprise data)

The AI platform includes vector search and RAG over governed enterprise data.

“Vector search and so on so forth”transcript, stated

`cost-aware-routing` (operations)

Model and platform choices are evaluated against ROI and cost control.

“how do we control the cost to achieve Roi”transcript, stated

`privacy-preserving-processing` (governance)

The platform treats IP protection as a primary design concern.

“how to protect IP”transcript, stated

Tradeoffs

Platform foundation

Chose Governed One Data plus One AI platform over Ad hoc application-specific GenAI integrations.

Rationale: The architecture had to combine data access, governance, model serving, LLMOps, vector search, and RAG.

“capabilities like model serving lops uh Vector search and so on”transcript, stated

Model economics

Chose Cost-controlled model selection over Unbounded use of expensive generative models.

Rationale: Different generative options have large cost differences.

“there's a huge cost range there so how do we control cost”transcript, stated

Gotchas (what bit them)

Generative AI quality is a platform problem, not just a prompt problem.

Root cause: Reliable answers depend on governed, structured enterprise data and model controls.

Resolution: Pair One Data with One AI platform capabilities.

“answers that we get are high quality”transcript, stated

Cost can become a blocker as usage scales.

Root cause: LLM options vary widely in price and ROI depends on controlling usage.

Resolution: Treat cost control as a core platform challenge.

“cost becomes a a huge issue”transcript, stated

Evaluation

ADP frames evaluation around answer quality, IP protection, data access, cost control, and ROI rather than reporting a single benchmark.

Metrics: Answer quality · IP protection · Data access and structure · Cost control · ROI

“quality and how do you measure quality”transcript, stated

Outcomes

ADP framed a One Data plus One AI architecture for enterprise generative AI.
The platform centers quality, IP protection, structured data access, and cost control for HR/payroll AI services.

The transcript reports over one million clients, 140 countries, and over 41 million workers paid through ADP's platform.

“we have over a million clients so pretty large scale”transcript, stated