Twelve months ago, this firm was running six people ragged across tax returns, client emails, and onboarding paperwork. Today, three AI agents — part of a bespoke AI operating system — handle the majority of that operational work, and the founding partner spends Monday mornings reviewing agent reports instead of digging through her inbox. Here's exactly what we built, how we built it, and what it cost.
The Starting Point: A Familiar Kind of Chaos
When Claire from Hartley Accounts contacted us in late 2025, her firm had six staff — three qualified accountants, one bookkeeper, and two administrators — serving around 140 small business clients across Yorkshire. Revenue was healthy. Growth was good. But the team was at capacity on what Claire called "the operational grind."
That grind looked like this:
- Client onboarding took 4–6 hours per new client: collecting documents, chasing missing information, setting up systems, drafting engagement letters
- Email management consumed 2–3 hours per person per day — most of it repetitive questions clients could have answered themselves
- Report preparation meant an accountant manually compiling data from Xero, formatting it into Word, and sending it — a 90-minute job done 30-plus times a month
- Tax return chasing had one administrator spending half her week calling and emailing clients for missing documents
Claire knew AI could help. She'd read the articles. But she didn't know where to start, and she couldn't afford to gamble on a system that might not work. So she booked a £500 AI Audit.
What the Audit Found
The audit covers every recurring process in the business. For Hartley Accounts, we scored each process on two dimensions: time consumption and AI suitability. The results were unusually clear.
Three processes were immediate targets:
- Client email triage — high volume, largely repetitive, rule-based responses. Automation score: 4.5/5
- Monthly report drafting — structured data in, structured document out. Automation score: 4.8/5
- Client onboarding — multi-step but templated process with clear decision points. Automation score: 4.2/5
One process we specifically recommended not automating in the first phase: anything requiring professional judgment on complex tax queries. AI can draft. It shouldn't decide on edge cases in tax advice — not yet.
"The audit was worth it just for the clarity. We'd been vaguely anxious about AI for months. Suddenly we had a prioritised list and a budget. That anxiety turned into a plan."
— Claire Hartley, Founder, Hartley Accounts
Total projected time savings from the three target processes: 19–24 hours per week. At a blended staff cost of £25 per hour, that's £475–£600 of recovered capacity every week.
Building the AI Operating System
We spent weeks one and two on infrastructure. The stack was the same one we use for all deployments: OpenClaw running on Google Cloud Run, connected to the firm's existing email provider, and linked to their Xero data via the accounting API. Total monthly infrastructure cost: £42.
Agent 1: Email Triage
This agent connects to the firm's general enquiries inbox and runs on every incoming email. It classifies each message into one of six categories: general enquiry, document request, payment query, meeting request, complaint, or complex query requiring a human response.
For the first four categories, it drafts a response and places it in a "pending review" queue. An administrator reviews and sends in one click. Complex queries get flagged to the relevant accountant with a one-paragraph summary of the email content.
Configuration took four days. In week one of live deployment, the agent correctly classified 89% of emails and drafted usable responses for 76% of those. By week four, after prompt refinement and adding the firm's FAQ document to the agent's knowledge base, accuracy reached 94%.
Agent 2: Monthly Report Drafting
Every month, Hartley Accounts sends 30–40 clients a management report: P&L summary, cash flow position, VAT status, and key observations. Previously this took a qualified accountant 60–90 minutes per report.
The agent connects to Xero via the accounting API, pulls the relevant figures for the reporting period, and generates a formatted report using a template built from the firm's existing report style. A qualified accountant reviews the output and adds any specific observations before sending.
The result: 90-minute jobs reduced to 10-minute review tasks. At 35 reports per month, that's 46 hours of qualified accountant time recovered per month — roughly one full working week.
Agent 3: Client Onboarding
New client onboarding involves collecting 8–12 documents (ID, business registration, bank details, previous accounts), creating the client file, sending the engagement letter for digital signature, and setting up the Xero connection. Previously: 4–6 hours of admin time spread across multiple staff.
The onboarding agent triggers when a new client record is created. It sends a customised welcome sequence over WhatsApp — the firm's preferred client communication channel — requests each document in sequence, chases automatically if documents aren't received within 48 hours, and notifies the relevant accountant when the file is complete.
Average onboarding time reduced from 12 days to 5 days. Admin time per client reduced from 4–6 hours to 45 minutes of oversight.
The Results at 60 Days
At the 60-day mark, we reviewed performance data across all three agents. The numbers exceeded the audit projections:
| Metric | Before | After 60 Days |
|---|---|---|
| Weekly hours on email triage | 14 hrs | 3 hrs |
| Monthly report preparation | 46 hrs | 6 hrs |
| Admin time per onboarding | 4–6 hrs | 45 mins |
| Average onboarding duration | 12 days | 5 days |
| Monthly infrastructure cost | £0 | £42 |
| Total weekly hours recovered | 0 | 22+ |
The full build cost — including audit, build time, and 60 days of running — came to £3,200. At £500 per week of recovered capacity, the system paid for itself in just over six weeks.
"The first Monday morning I spent reviewing the agent's reports instead of catching up on emails was the moment I believed it. We're not doing different work. We're doing the same work — just without the grind."
— Claire Hartley
What Didn't Work (And What We Fixed)
No case study is honest without the failures. Two things needed fixing in the first 30 days:
The email agent over-classified. In week one, it flagged 22% of emails as "complex queries requiring human response" — twice the actual rate. The agent was being cautious, which is better than the opposite, but it created unnecessary overhead. We added more category examples to the training data and the over-classification rate dropped to 8%.
The report agent struggled with non-standard periods. Two clients had non-April financial year-ends. The first time the agent ran their reports, it pulled the wrong date range. This was caught in the human review step — exactly as the oversight process is designed to catch it — and the template was corrected. It hasn't recurred.
Both issues are typical of early deployments. They're why we build with strict oversight guardrails and don't remove human review until an agent has run cleanly for 30 consecutive days. As we cover in the 90-Day AI Transformation Playbook, trust needs to be earned through verification, not assumed from the start.
What's Next for Hartley Accounts
The three agents are now part of the firm's daily operations. In phase two, we're adding a fourth: a tax deadline monitor that tracks upcoming filing dates across all 140 clients, sends automated reminders at 60, 30, and 7 days before each deadline, and escalates to the relevant accountant if a client hasn't responded.
The five-year vision Claire shared with us is to double her client base without adding to the current headcount. With an AI operating system handling the operational work, that goal is no longer ambitious. It's a plan.
If you run a professional services firm and you're spending more time on operational grind than on the work only you can do, the starting point is the same one Hartley Accounts used: understand what you're doing, score it, and pick the three best targets. Get in touch and we'll walk you through it.