In the first post of this blog series, I covered a shift we’re seeing across convenience retail: software that goes beyond reporting and analysis into systems that can help operators take action. That shift raises an obvious question: What has to exist underneath a product like PDIQ for it to consistently act in a reliable way?
The answer involves much more than a foundation model.
When supporting capabilities such as loyalty optimization, operational alerts, and workflow automation, the model itself is only one part of the system. Equally important are the layers responsible for organizing data, maintaining context, enforcing governance, and coordinating actions across multiple business domains.
When we began building PDIQ, we focused on three foundational layers to deliver the operational AI capabilities available today, all while maintaining human oversight:
- A data mesh layer that defines how data is owned, governed, and shared
- A retrieval layer that helps connect information across many types of operational data
- An agent framework layer that allows reasoning, evaluation, and action within defined policies
Layer 1: A data mesh for operational context
As a business, PDI operates across several distinct domains, including wholesale fuel, retail POS, loyalty, marketing, and more. Each domain has its own update patterns, business rules, ownership model, and governance requirements.
We chose a data mesh architecture because it allows the teams closest to each domain to remain responsible for the data products they understand best. Several principles guided that design.
Data products aligned to the business
Fuel, retail, loyalty, and consumer demand each operate as durable business domains. Each domain publishes versioned, documented data products with clear ownership and defined contracts for downstream consumers. That structure makes it easier for analytics, workflows, and AI capabilities to consume information consistently while preserving the business context behind the data.
Visible quality and freshness
Every data product includes metadata about its quality, completeness, and freshness. Those signals are available to any system consuming the data, including PDIQ. Context about data quality becomes part of the decision-making process rather than something that’s discovered later.
Governance built into the platform
Access policies, regional requirements, and customer-level data boundaries are enforced at the platform level. Teams building on top of the platform inherit those controls automatically. This allows new capabilities to operate within established governance models rather than creating separate implementations for each use case.
Layer 2: Retrieval designed for retail operations
Many conversations about AI retrieval focus on documents and text. Those use cases definitely matter, but they represent only a portion of the information involved in convenience retail operations.
That’s why we built PDIQ to work with a much broader set of signals within a customer’s operations:
- Transactions
- Pricing history
- Fuel volumes
- Loyalty activity
- Geospatial information
- Consumer demand patterns
- Operational events
- Documentation and knowledge content
Multiple data types, shared retrieval
Text, structured records, time-series data, behavioral patterns, and location-based information can all contribute to a decision. When PDIQ evaluates a situation, it might need to retrieve examples of similar stores and supporting documentation—always within that same customer’s own data set. A store in one region might share characteristics with a store in a completely different region, two locations a human analyst would rarely think to compare, simply because they belong to the same retailer’s network. The retrieval layer provides a common framework for surfacing those connections within each customer’s governed boundaries.
Domain-aware embeddings
Each business domain is responsible for defining how its information is represented. A transaction in a retail POS environment carries different meaning than a transaction in wholesale fuel operations. Treating those concepts as identical would reduce the quality of retrieval and reasoning. By keeping representation closer to domain ownership, retrieval remains aligned with the business context behind the data.
Hybrid retrieval
Most operational questions involve both structured filtering and semantic search. For example, you might care about:
- A specific region
- A particular fuel grade
- A defined customer segment
- A recent time window
At the same time, you might also need information that is behaviorally or semantically similar within that same customer’s own operations. PDIQ combines these retrieval approaches so that operational context remains precise while still benefiting from modern AI search techniques.
Privacy and data boundaries
Trust depends on maintaining clear boundaries around customer information, particularly when AI capabilities interact with operational data. That’s why customer data is partitioned and governed according to established access controls.
Layer 3: An agent framework for operational decisions
With data and retrieval in place, the next step is execution: how AI capabilities should interact with the rest of the platform.
Early in the development of PDIQ, we chose to build around agents rather than isolated prompts. Operational work rarely fits into a single question-and-answer exchange. With complicated operations in an industry like ours, there are often multiple steps, sources of information, and interactions with external systems. Agents provide a flexible structure for coordinating those activities.
Focused domain agents
Rather than building a single general-purpose agent, PDIQ uses agents with narrower responsibilities because focused agents are easier to evaluate, improve, and govern.
Tool-based execution
Agents interact with the platform through approved tools and services. Data access flows through governed data products and retrieval systems. Actions flow through workflow engines, policy controls, and audit mechanisms. This approach keeps operational actions observable and consistent with existing governance requirements.
Cross-domain coordination
Many operational decisions depend on information from multiple parts of the business. Imagine a scenario where you’re preparing a personalized loyalty offer. Determining the right offer for your business would likely require agents to process a variety of data elements, along with orchestration to coordinate information across multiple agents and business domains. After that step, agents would present a recommendation or initiate the next step in the workflow.
Policy-driven operation
Every capability operates within defined boundaries. Policies determine which actions can be taken automatically, which require review, and which should always be escalated—thresholds that the customer configures and owns for their own workflows. These controls allow you to decide how much autonomy is appropriate for a given workflow and to maintain human oversight.
How all three layers work together
Consider the example of a retailer looking to improve engagement with a loyalty segment whose purchasing behavior has recently changed.
The data layer brings together transaction history, loyalty activity, promotional performance, inventory availability, and consumer demand signals from across that retailer’s business. The retrieval layer assembles the most relevant context, including similar segments within that retailer’s own customer base, recent campaign performance, and current operational conditions. An agent then evaluates potential offers or actions against the retailer’s business objectives and defined policies.
Depending on the retailer’s configuration, the recommendation may be routed for review, executed with oversight, or automatically incorporated into a broader customer engagement workflow.
Several platform services participate throughout that process, but they all operate within the same framework for data access, governance, retrieval, and policy enforcement. That shared foundation allows operational AI capabilities to scale while maintaining consistency, transparency, and control.
Up next in this series
The capabilities you experience in PDIQ are built on infrastructure designed to make decisions more contextual, explainable, and governable.
In future posts, we’ll explore some of these areas in greater depth, including agent evaluation, retrieval over operational data, and the engineering patterns that support safe execution.
If you’re interested in more AI insights, watch these videos.
See PDIQ in person at Connections Live 2026!
