A data platform for AI is the database, the data model and the loading jobs that give an agent records it can trust. An agent needs four things from your data: one source of truth, history, permissions and a known freshness.
What a data platform for AI is
A data platform receives records from the systems a business runs on, gives them a shape and keeps them. It has a database that persists, a model with keys and constraints, jobs that load and update the data, and rules that say who may read what. Dashboards read it. Agents read it too, through retrieval or through tools, which are often exposed with MCP (see what MCP is).
It is not one product. In the two platforms described in this guide the database is PostgreSQL, and most of the work is in the model and in the loading jobs around it.
Cotizera, a quoting and deal-tracking system for small and medium businesses in Latin America, shows why this layer comes before the agent. Its case study states that customer, product, pricing, inventory and pipeline data matter more than the chat interface. The agent touches real records and creates documents that a customer will read. If the records are wrong, the quote is wrong.
What an AI agent needs from your data
An agent answers from the records it is given. It cannot tell a duplicate from a customer, or a placeholder from a date, unless the data makes the difference visible.
- One source of truth. Each customer, order or contract has one valid record, with a key that other tables point to. For a business in the warranty management industry, I found the data spread across MongoDB collections with no relational integrity. I moved it to a relational PostgreSQL schema where dealers link to agreements and agreements link to claims, and I deduplicated dealers with a normalized payee identifier. The case study calls the result the single source of truth of the organization.
- History. People ask an agent what changed, not only what is true now. A system that overwrites or deletes its data cannot answer that. In the warehouse example below, work orders keep their earlier versions.
- Permissions. The Cotizera design states that an agent that touches real records needs permissions. As general advice, enforce them in the database and in the service in front of it, outside the model. The guide on connecting a database through MCP covers that boundary.
- Freshness. An agent should be able to say how old its data is. The analytics pipeline loads incrementally. It detects changed agreements by hash and changed claims by timestamp, so each run processes only what changed.
The same platform keeps its calculations in the database. PostgreSQL functions compute figures such as revenue growth for the dashboard. As general advice, an agent should call those same definitions. Then a person and an agent who ask the same question get the same number.
Data sync or data warehouse: how to decide
These rules are general advice drawn from the two platforms. They are not measured thresholds.
A sync between two systems is enough when all of these hold:
- Each system owns its own records, and the sync only copies them across.
- Only the current state matters. Nobody asks what a record looked like last month.
- A failed run can be repeated without losing anything.
- Every question can be answered inside one of the two systems.
You need a warehouse when one of these appears:
- Someone needs history, or a record of what each sync did.
- A question crosses systems, such as work orders by building and by tenant.
- The same customer or unit has a different identifier in each system and has to be reconciled.
- Duplicates reach the destination.
- Failures have to be seen and explained after the fact.
In the integration I inherited, the choice of PostgreSQL over DynamoDB came from the need for complex queries, historical tracking and dimensional modeling. The warehouse did not replace the sync. It sits between the two systems, and the original sync logic was preserved.
How to prepare company data for AI agents step by step
- Write the questions the agent must answer and the records each question needs.
- Decide which system owns each record: customers, prices, orders.
- Model keys and constraints. Give each entity one identifier, add foreign keys, and add mapping tables where two systems name the same thing differently.
- Clean the data at load. In the analytics pipeline, placeholder timestamps become empty values and duplicate dealers are removed.
- Keep history. Store the earlier versions of rows that change, and log each sync run and each error.
- Load incrementally and in batches, with change detection by hash or by timestamp.
- Add operating controls: a dry-run mode, a switch that turns a destination off without a deployment, and retries with backoff.
- Expose the data to the agent through narrow operations with permissions, and test them with real questions. The guide on automated LLM evaluation describes that testing.
Steps 1, 2 and 8 are general advice. Steps 3 to 7 come from the case studies of the two platforms.
What goes wrong when agents read company data
Each item below comes from one of the two systems.
- The sync deletes what it moves. The inherited integration deleted its data after every sync. There was no history and nothing to analyze.
- Two identifiers get confused. In 40 work orders the unit identifier held the building identifier, and technicians were routed to the wrong units. Any reader of those rows, a person or an agent, gets the wrong unit.
- Duplicates and placeholders pass as facts. The warranty data had duplicate records and placeholder timestamps such as
0001-01-01. A model reads a placeholder date as a date. - Nobody sees the failures. The inherited integration had an error rate of 15 to 20 percent in production and no visibility into which syncs failed.
- Every run reloads everything. The warranty data was growing and there was no incremental processing strategy.
A worked example: from a sync bridge to a data warehouse
I inherited an integration between PropertyWare and ServiceFusion. Its original design had served production for years, and its limits appeared as the business grew. It moved work orders through temporary DynamoDB tables, in sequence, and deleted the data after each sync. Recovery from an error was manual.
I kept the sync logic and placed a persistent PostgreSQL database on Supabase between the two systems. The model is a Kimball dimensional model. Fact tables hold work orders and leases. Dimension tables hold portfolios, buildings, units and tenants. Mapping tables reconcile records across the two systems. Work orders keep their history: when one changes, the earlier version stays (SCD Type 2).
Around the model I added the controls. A deduplication service checks several criteria before it creates a job in ServiceFusion. Processing runs in parallel through SNS fan-out. Errors are retried with exponential backoff. Feature flags turn the ServiceFusion sync off without a deployment, and a dry-run mode allows safe tests in production. A SQL correction with validation logic fixed the 40 work orders that pointed to the wrong unit.
The case study reports an error rate under 1 percent, a sync time reduced by 70 percent and historical analysis that was not possible before. The system stayed in production during the change.
For an agent, the useful part is the shape of the data. A work order points to one unit. Its earlier versions exist. Each sync and each error has a row. An agent can be given narrow tools over tables like these, and its answers can be checked against them.
What these case studies show and what they do not
- Neither case study describes an agent reading the platform. The analytics platform feeds a dashboard that people read. The warehouse serves the sync between two systems and its reporting. That an agent needs the same foundation is my engineering judgment, supported by the Cotizera design. It is not a measured result.
- The Cotizera case study describes the design of an agent and its layers. It reports no production measurements.
- The figures quoted here are the ones the case studies reported in August 2025. They are not an agent load test, and I did not measure them again for this guide.
- The warehouse analysis records open work at its date. The automatic schedule was disabled and syncs were started manually. The customer mappings between the two systems were not complete.
- No source here measures the quality of answers. Whether an agent answers better on a given platform has to be tested with real questions.