How William AI is leading the way in AI bookkeeping

When we set out to build William AI, Apron’s bookkeeping agent, we quickly ran into a unique challenge: there was virtually no market standard to follow (or to beat). 

In the emerging world of agentic bookkeeping, no clear North Star existed. As Tallulah Watson, Senior Product Designer at Apron, notes:

"There wasn't, and there still isn't, a huge amount of market standard for us to go off."

Defining what’s possible in AI bookkeeping required us to build rapidly while working directly alongside accountants and bookkeepers. Rather than copying existing software, we had to pioneer a whole new way of working.

Learning through iteration

Designing for autonomy meant solving the trust puzzle. Early in the process, our team actually over-corrected. In an effort to provide maximum transparency, we overloaded the interface with information.

We quickly realised that throwing endless data points at users isn't the same as giving them clarity. 

True confidence comes from radical simplicity. Through constant testing and iteration, we stripped back the excess to create an interface where every piece of information serves a clear purpose.

Take William AI’s reasoning, for example. Instead of jargon-ese error messages, we built William to offer explanations in plain English whenever it wasn’t sure about a document. 

"What others have implemented is more like processing rules for suppliers, whereas we are approaching it as adding context, teaching us what's happening in the business."

Static supplier rules break down when a single vendor covers multiple purchase types. William AI, by contrast, dynamically learns from natural language guidance and historical patterns. 

In other words, bookkeepers and accountants tell William how to handle a specific document type or supplier using normal language, and William learns and remembers.

Built with guardrails, and with bookkeepers

Automation shouldn't mean losing control. To ensure William acts as a reliable digital team member, we engineered over 20 validation checks to catch errors at the source. 

We are also constantly refining the experience for both bookkeepers and business owners. For example, a recent update allows William to automatically detect paid invoices and collapse them into a single, unobtrusive line, keeping client dashboards clear of unnecessary noise while also keeping every record accessible.

By pairing machine learning experts, designers, and engineers directly with accounting and bookkeeping professionals, Apron is setting the new standard for agentic bookkeeping.