Custom where off-the-shelf stops.
For commodity flows, we almost always recommend the standard offering. Klippa or Basecone for pure receipt scanning is faster and cheaper than anything we can build — and we'll tell you so. We only come into the picture when more than three systems need to work together, or when the AI layer needs its own risk framework or industry-specific flow that standard packages don't cover.
Think of healthcare accountancy with automatic DBC checks, housing association audits with a DAEB/non-DAEB split, or a cooperative of independent bookkeepers that wants a shared white-label AI tool. For these kinds of challenges we build the custom solution and integrate it tightly with your existing stack, often drawing on our experience with custom LLM integrations and custom accountancy software.
The architecture always follows the same pattern. We place a thin abstraction layer between the AI model and the accounting package, so that both the model and the underlying stack can be replaced without rewriting the business logic. Prompts and model choice are version-controlled, tested with representative client data, and automatically evaluated for accuracy before a new version reaches production. That's deliberate: the LLM landscape changes every few months, and a firm running on GPT-4 today should be able to switch to Claude 5 or an open-source alternative within a year without hassle, if that suits it better on price, speed or compliance.