LedgerPlus: from Claude to Combinely

LedgerPlus is a growing accountancy practice with ambitions to match. They were already taking advantage of Claude, and doing it well - tax return reviews, Monday reporting drafts, even a Python script Kyle had built himself to run one client's fortnightly payroll.

It just wasn't enough. Every workflow still needed a human to summon it, feed it and act on it, and none of it could move at the pace the firm wanted to grow. One week after going live on Combinely, a client payroll ran start to finish on a schedule with no errors.

The Opportunity: creating the first AI accountant - a real member of the team.

LedgerPlus were not AI beginners. They had been running Claude thoughtfully for months. Whilst it was proving very helpful, it never broke through.

• No system access. Claude could cross-check a payslip, but a human still had to key hours into Xero one employee at a time.
• No context, no schedule, nothing shared. Kyle's best prompts lived in Kyle's account, cold-started every conversation, ran only when a human remembered.
• DIY didn't scale. Kyle's Python payroll script worked, until it didn't.

Kyle could reason with Claude. He couldn't run a firm on it.

How they use Combinely

1. Payroll, end to end

The old cycle: managers key hours into a wage sheet, or the team pulls them from Deputy or Square by hand. Tronc gets allocated by seniority. Somebody opens the Xero pay run and copies hours, holiday and tronc across, employee by employee, unticking anyone not being paid that period.

That entire cycle is now one agent. It reads the wage sheet, maps each employee to the right Xero pay calendar, applies tronc and net deductions, unticks anyone not being paid that fortnight, and cross-checks its own work three ways before it stops. Every run lands in draft in front of the accountant, with a full record of what it did. The firm removed the typing, not the control.

For one client the agent goes a step further: it signs into Deputy, exports both weeks of timesheets, and keys the hours into the wage sheet before it ever opens Xero.

Firm-wide there are 20+ live triggers - some running on a clock, some waiting on an inbound email - so the work happens whether the team's laptops are open or not. Around the payroll agent the team built a Nest pension deadline monitor, a leave-balance reconciliation, and an agent that logs into HMRC and reads what a “you have new messages” notification actually says.

Within the past 2 months, the firm has created agents to automate the payroll process for 90% of their clients who have more than 3 employees.

2. The VAT review Claude used to sense-check

The job that used to mean exporting a return and pasting it into a chat window is now an agent that opens Xero itself, generates the return, exports the source documents and works through a QC checklist. The reviewer starts from a checked, annotated first pass - closer to a partner's marked-up version than a raw draft.

The same pattern spread across tax and reconciliation reviews, plus a bank rec comment sweep that runs Monday, Wednesday and Friday at 7am - so clients wake up to fewer questions in their inbox and cleaner books by lunchtime.

3. A colleague, not a tool

Claude was a tool the accountant had to operate - open the chat, paste the data, ask the right question, read the answer back into the work. Coco does the work. It signs into Xero and Deputy in a real browser. It knows the difference between salaried and hourly staff, tronc subject to PAYE but not NI, reverse charge on software costs.

Kyle stopped operating the AI and started reviewing it. So did the rest of the team.

The Result

• 1 week from go-live to payroll running end to end
• 40+ agents live across payroll, VAT, tax, bank rec, invoicing, reporting and client comms
• 2,000+ tasks run in the first two months
• 90% of clients with more than 3 employees now have their payroll on the agent
• Around 20 hours saved per month in payroll processes, reported by the firm
• 100% daily usage across the firm

The bottleneck was never the reasoning. Once the reasoning didn't need to be summoned, fed and acted on by a human, payroll stopped being a job and became a review - and every hour that used to go into it went back to the client. Kyle is no longer asking whether payroll can be automated. He is asking how to bring the next twenty clients under the same standard of service.