For accountancy and bookkeeping practices, payroll bureaux, IFAs and small wealth firms. We build AI agents that prepare the routine work from your own records and hand every decision to the person who owns it.
Owners, managing partners, practice managers, MLROs and operations leads. The people who carry the risk when something goes out wrong, and who want their team’s time back without giving up control.
Examples of jobs an agent can take. Where we have already built something like it, we say so; the rest are proposed examples we would scope with you in an audit. In every case the agent prepares, a person decides, and the log shows both.
| The job | What the agent does | What a person decides | Where it sits |
|---|---|---|---|
| Client document chasingProposed example | What the agent doesTracks what each client owes you and by when, sends polite reminders in your words, and files what arrives against the right request. | What a person decidesWho gets chased, and when a reminder becomes a phone call. | Where it sitsYour engagement letters and filing deadlines. |
| Payroll and pension declarationsWe have built thisPayroll requests and pension declaration tracking, in a client portal for an accountancy practice. | What the agent doesKeeps one live list of payroll requests and pension declaration dates for every client, and drafts the paperwork from the details on file. | What a person decidesEvery submission. Nothing is filed without a person. | Where it sitsRTI deadlines and The Pensions Regulator’s declaration of compliance. |
| Bank and supplier reconciliationProposed example | What the agent doesMatches statement lines to invoices and receipts, explains each match, and lists the ones it could not match with a reason. | What a person decidesThe exceptions, and anything above a value you set. | Where it sitsYour own controls and year-end evidence. |
| AML onboarding checksProposed example | What the agent doesReads ID and company documents, pulls ownership records from Companies House, flags possible sanctions matches and drafts the due diligence file and client risk assessment. | What a person decidesWhether to take the client on. The agent never passes or fails anyone. | Where it sitsMoney Laundering Regulations 2017: customer due diligence and risk assessment. |
| Month-end exception triageProposed example | What the agent doesRuns the checks your team runs every month and sorts what it finds into “fine”, “needs a look” and “stop”. | What a person decidesEverything in the last two piles. | Where it sitsYour month-end checklist. |
If a regulator, a supervisor or a partner asks how an answer was reached, you should be able to show them.
The model does the reading and the drafting. Fixed code does the routing and the sums. A person makes the decision.
Every input, draft, check and approval is recorded with who and when.
Answers cite the document or official page they came from. With no source, the agent says so and hands over to a person.
The agent prepares and your team decides. We map the approval points with you before we build.
Each agent has its own set of test cases, including ones designed to trip it up. It must pass them before it goes live and again before every change.
We agree where your data is stored and processed before the build starts, and write it down.
What this looks like in a real practice, and in a business that runs on supplier data.
Accountancy practice
Answers from official pages with every figure checked, a portal with an audit log that cannot be edited, and an honest account of launch day.
Read the case study: Accountancy practicea year in subscription fees replaced: £350 plus VAT a month, or £5,040 a year including VAT.
The fee replaced, before the new system’s own running costs. The retailer is VAT-registered, so £4,200 is the real cost.
IT retailer
Three supplier feeds, about 5,000 live products and a best-supplier dashboard that refreshes every 10 minutes, with broken prices held for a person.
Read the case study: IT retailerNo AI model is right every time, so we design around that. Figures come from your records or from fixed, tested code rather than the model’s memory. Written answers are checked against their sources, and when an answer cannot be supported the agent is built to refuse and hand the question to a person. We test every agent against questions written to trip it up, and publish what the tests found in our case studies.
We agree it with you before we build and write it into the plan. We tell you which services process your data, where, and for how long anything is kept.
Your firm stays responsible for its advice and filings, as it is today, which is why a person signs off every decision. We are responsible for the agent doing what we agreed and tested. If it falls short, we fix it, and the log shows exactly what happened.
Usually. We work from the exports, inboxes and portals you already use, and connect directly to a system only when it offers a proper way in. We will not ask you to change your practice software.
We agree the timeline in the audit, once we know the workflow and the data. A first pilot covers one workflow, not the whole practice, so it stays small enough to test properly.
No. We build tools for your team. We do not give tax, investment or legal advice, and we will not build an agent that gives buy, sell or hold recommendations or makes credit or onboarding decisions without a person.
In 30 minutes we will tell you whether an agent could take it on, what it would need and what should stay with your team.