Duane Jackson's Sodium has moved to general availability. It is API-first, built around a single client record and designed with AI integrated from the start rather than bolted on later. Leigh frames the real challenge not as product quality but as firm inertia: practice management is the Lego wall nobody wants to dismantle, and a golden brick is only useful if someone is prepared to pull the old ones out. Lara and Robbie both know from experience how painful that process is, and the conversation turns quickly to whether AI-assisted migration might eventually lower the barrier.
Apron's William AI is now generally available, having launched in beta in March. Lara walks through what it actually does: connecting to client email inboxes, extracting and categorising documents, publishing to Xero or QuickBooks, and flagging anything it is not confident about. The auto-publish toggle defaults off and needs firm-specific guidance rules to reach its potential. The beta hit 50% autonomous publish rate; the GA pitch is 90% for firms that put the configuration work in.
Robbie leads on the Xero and Claude integration, which went live globally on 12 May. Early practitioner testing found it read-only, limited to account-level data and prone to missing transactions. Leigh and Lara discuss what it means that the major general ledgers are simultaneously embedding AI inside their own products and surfacing their data inside the large language models.
Intuit's intelligence layer across QuickBooks covers AI-powered chat, portfolio benchmarking and a capability that appears pointed directly at end clients. Lara raises the concern that clients with incomplete books could get confident-sounding answers to questions they should be asking their accountant instead.
Lara covers the QuickBooks AI-powered MTD checker, which flags duplicates, missing transactions and non-trading income sources before submission. QuickBooks claims the highest cumulative MTD pilot sign-ups during HMRC's testing period. Robbie welcomes it as a live use case rather than a theoretical one.
Robbie covers Combinely, the browser-based AI co-worker backed by YC and OpenAI. Early UK adopters include Burgess Hodgson, where it handled 2,600+ tasks across December and January with a reported 75% reduction in income and expenditure creation time. Leigh raises the structural tension: a tool that sits on the periphery of a workflow is easy to adopt and equally easy to quietly drop.
Lara covers FreeAgent's integration with Equali, pulling e-commerce data from Shopify into FreeAgent for reconciliation and categorisation, and notes FreeAgent's incoming Apron partnership as part of a broader push beyond its freelancer roots.
The episode closes on the Accountex panel from ICAEW, ACCA and IFAC. The fat middle concern runs through the final section: AI handling transactional work, hollowing out the junior pipeline and, with it, the intuition that comes from years of doing the basics. Robbie's view is that AI will eventually learn the human stuff too. The question is what accountants do with the time that creates.
Chapter list
00:00 Introduction and Overview of Topics
04:00 The Launch of Sodium: A New Practice Management Tool
08:25 Apron's William AI: Enhancing Document Management
12:44 General Ledgers and AI Strategies
23:31 QuickBooks AI-Powered MTD Checker Flags Errors Before Submission, Not After
27:18 Combinely's AI Co-Worker Handles 2,600+ Tasks in a Month for Early UK Firm Adopters
30:45 FreeAgent's E-Commerce Expansion
36:09 The Future of Accountancy: AI and Brand Perception