Economy

Add as a preferred source on Google

H6.ai uses database snapshots to test AI invoice processing

H6.ai preserves customer database snapshots to test changes to its AI invoice processing without altering live financial records, according to a Databricks blog published Nov. 12.

Updated 2 min read
Data centre cooling plant.
Data centre cooling plant.

H6.ai used frozen copies of customer databases to test changes to its artificial intelligence invoice processing without altering live financial records, Databricks said in a blog published Nov. 12.

The approach helps distinguish software errors from changes in customer records, H6.ai said in “How H6.ai builds AI-powered financial automation with Lakebase Postgres.” Databricks, which supplies the database product, published the account.

H6.ai said changing purchase orders can make the same invoice produce different results even when its software stays unchanged. Preserving the underlying records gives engineers a consistent starting point when checking software changes.

The company described its current testing system for accounts payable, the money businesses owe suppliers, rather than results measured over a specified study period. It supplied no figures for processing accuracy, testing speed or cost savings.

H6.ai said its software works with customers’ existing enterprise resource planning systems, which hold business records. It reads invoices, matches them to purchase orders, assigns invoice lines to accounts and checks compliance with customer policies.

Each customer has a dedicated Lakebase Postgres database containing purchase orders, supplier details, account information, past transactions and policies, the company said. Keeping separate databases isolates each customer’s financial information.

When an invoice produces an incorrect result, H6.ai said engineers preserve the customer’s database state in a separate branch. That copy retains purchase order quantities, prices and other records even after the live system changes.

A new branch initially shares storage with its parent database and stores changes separately, the company said. This lets engineers preserve the conditions behind an error without rebuilding those conditions later.

Each test pairs an invoice with a preserved database snapshot and an expected result, H6.ai said. Those expectations cover the purchase order match, account assignments and whether the invoice should go for review.

Tests can combine invoices from different customers and business systems while retaining a separate starting point for each case, the company said. Each run writes its results to its own branch, which is then deleted.

H6.ai said tests use the same processing routes as its live service, while controls prevent changes to customers’ active databases or postings to their business systems. Dedicated service accounts have restricted access to manage branches.

The company said completing an invoice-processing task does not establish that the result is correct. Its checks also assess whether the invoice matched the right order, reached the right account and followed customer policies.

H6.ai said keeping each corrected error as a permanent test case helps engineers check new models and changes to matching or policy rules before release. It also plans to query Lakebase records for analysis without copying the data.

Related articles

View all