Accounting Practice Data Audit: 6 Checks

Your reports answer the questions someone built a screen for. The questions that keep a firm owner up at night are different — which clients are we losing money on? and will next month’s invoice run actually produce anything? — and each one needs two things that live on separate screens.

That used to mean two CSV exports and an afternoon in a spreadsheet. Now it means pasting a prompt.

Below are six of them. Copy, paste, read the answer. We ran all six against a real account — 65 clients, 15,000 records — and the numbers under each one are what came back.

First, connect your AI to Uku

One-time setup. Create an API key in Uku under Settings & Apps → Public API, then paste this into Claude, ChatGPT, or whichever assistant you use. A Read key is enough for everything in this article.

Claude Cowork Uku integration setup

Connect to my Uku account through the Uku API and set it up so I can reuse it. Reference: https://help.getuku.com/en/integrations/public-api-v3/ Company UUID: [your company UUID] API key: [your uku_live_ key] Test the connection, then save it as a named connection I can use in future chats.

Fill in the two highlighted values Copy prompt

Live in a terminal instead? Install the Uku CLI with curl -fsSL https://getuku.com/install-cli | sh and sign it in once. After that Claude Code can answer the same questions by running it for you.

That is the whole setup. Everything below is a question you ask afterwards — in whichever of the two you prefer.

1. Which clients are we working for free?

Hours logged against money invoiced, per client, worst first.

Claude Cowork Unbilled work

In my Uku account, find every client where the billable hours we logged are worth more than what we actually invoiced them. For each client show: client name, billable hours, currency, total invoiced, and the effective hourly rate (invoiced ÷ billable hours). Two rules: only count invoices that have an invoice number and are not annulled, and never add up different currencies — group by currency instead. Sort worst first and tell me the five I should look at on Monday.

Copy prompt

What it found: 19 clients under water. The worst was 668 billable hours invoiced as 2,265 — an effective rate of 3.40 an hour.

The same question in a terminal — Claude Code reads your account through the Uku CLI and answers:

Read it before you act on it. A fixed-fee client whose hours ran long is a different conversation from one nobody billed. And one row here was not a client at all — it was the firm itself, booked as a client record.

2. Will next month’s invoice run bill anything?

A contract row can bill 0.00 for five reasons, and not one of them is an error. The product is inactive. There is no product on the row. The row’s own price is zero. An hours-based product has no time logged. A custom field the price depends on is empty.

Every one is a valid setup that quietly produces nothing — which is why nobody catches it.

Claude Cowork Invoice preflight

Before my next invoice run in Uku, check every row on every active billing contract and tell me which ones will bill 0.00 — and why. Check each row for these five causes: the product is inactive, there is no product on the row, the row's own price is set to 0.00, the product bills by hours but no billable time was logged this period, or a custom field the price depends on is empty. Show me client, contract and the reason in plain English. Then tell me how many of my active contracts are affected. Also list invoices that were calculated but never issued, grouped by currency, oldest first.

Copy prompt

What it found: 32 of 46 active contracts at risk — 53 rows would have billed zero.

Or from a terminal:

Then the part that stops being a neat trick. Checked against every invoice the account had ever produced, going back two years: 212 invoices had already gone out at 0.00. Fifty-five of them carried a real invoice number. Twelve were marked paid.

And separately: 496 invoices calculated and never sent at all — the oldest 909 days old. Together that is 266,745 USD, 47,996 GBP and 42,154 EUR of finished work that never left the building.

Before: you find out when a client asks about a blank invoice. After: you find out on the 31st, with the reason attached.

3. Where does the money actually come from?

Claude Cowork Revenue mix

Show me where my revenue actually comes from in Uku. Use only issued, non-annulled invoices, and report everything per currency — never add currencies together. For each currency: total invoiced, the biggest client's share, and the top three clients' share. Then for my biggest currency, break revenue down by service, and tell me which services earn money with no tracked hours behind them. Finally find price drift: the same product billed at very different prices to different clients. Show the lowest, the highest and the spread.

Copy prompt

What it found: the work you spend time on and the work that pays are not the same work. Mailbox and address services were 55% of revenue with zero tracked hours behind them.

Or from a terminal:

The same run caught price drift — one service billed anywhere from 10.00 to 400.00 depending on the client. Nobody decided that. It accumulated, one exception at a time.

4. Do the people who own the work actually do it?

Hours per person is a report. Open tasks per person is a view. What matters is the two of them in the same row.

Claude Cowork Capacity vs ownership

In my Uku account, put open task ownership next to hours actually logged, per team member. For each member show: number of open tasks (status new or in progress), their share of all open tasks, and hours logged in the last 30, 90 and 365 days. Flag anyone who owns open work but has logged nothing in the last 90 days. Don't try to calculate utilisation — there is no capacity field on the member record, so say what you can and stop there.

Copy prompt

What it found: two people owned 68% of all open work and had logged nothing in three months.

Or from a terminal:

A leaver whose queue was never handed over, a role that stopped tracking time, or a real bottleneck — whichever it is, every task in those two queues has a due date nobody is working towards.

5. What small things are quietly costing us?

Claude Cowork Data hygiene

Run a data-quality pass on my Uku account and report five things. Keep it to short lists, no commentary. 1. Clients whose names are near-duplicates of each other 2. Every different seller name that appears on our issued invoices 3. Time entries longer than 12 hours, and whether each is marked billable 4. Clients with work logged in the last 90 days but no active contract 5. Clients with no activity for over 90 days that are still on an active contract

Copy prompt

What it found, in one pass:

Or from a terminal:

• Five different seller names on issued invoices. Most were correct. Seven carried a one-letter typo of the firm’s own name — seven clients got an invoice from a company that does not exist.
• 63 time entries longer than 12 hours, 1,179 hours in total, every one marked billable. The largest was a single 66-hour entry. Each is a forgotten timer or a typo, and all of them sit in the billing base.
• Duplicate clients, and 10 clients with live work but no active contract — including one record that is not a client at all, just a broken reference nothing in the interface can show you.
• 3 clients dormant for 90+ days but still on an active contract. Still billed, no longer served.

Before: you find each of these by accident, usually because a client points it out. After: one question, once a month, before month-end close.

6. Bulk work that survives being interrupted

The five above only read. This one writes, and it is the one place where a terminal beats a chat window.

We killed a run mid-flight to see what would happen, then re-ran the same file. Nine lines were skipped as already done, ten were created, and one was refused — the process had died between sending it and hearing back, so the tool did not know whether it had landed. Zero duplicates.

That refusal is the point. The honest answer to “did this write go through?” is sometimes “I don’t know”, and a tool that guesses will either double-bill a client or skip them.

Trying this yourself

Everything above runs on Uku API v3, available on the Elite plan. Two ways in:

• Your AI assistant — the connect prompt at the top, then any of the six. This is the path for most people; there is nothing to install. More on connecting your AI to Uku.
• The Uku CLI — for terminals, scripts and coding agents. One line to install: curl -fsSL https://getuku.com/install-cli | sh

Read-only questions need a Read-scoped key and nothing more. Anything that touches money needs an All-scoped key you grant on purpose.

Frequently asked questions

What is a practice data audit?

A pass over your own practice data looking for money and work that quietly went missing: hours you never invoiced, invoices about to bill 0.00, work calculated but never sent, tasks owned by people who are not working, and duplicate or broken records. Each question joins two things that live on different screens, which is why no standard report answers them.

Do I need a developer to run these?

No. Every check here is a prompt you copy into Claude, ChatGPT or another assistant that can call an API. You connect it to Uku once, then ask in plain English. A developer is only needed if you want to build something on top of it.

Why would an invoice bill 0.00?

Five reasons, all invisible until the invoice exists: the product on the row is inactive, the row has no product, the row’s own price is set to 0.00, an hours-based product has no time logged this period, or a custom field the price depends on is empty. None of them is an error, which is exactly why nobody catches them.

Which Uku plan do I need?

The Public API v3 is available on the Elite plan, and only Company Owners and Admins can create API keys. Reading your own data needs a Read-scoped key and nothing more.

Is it safe to let an AI touch my accounting data?

Every check in this article is read-only — the prompts ask questions, they do not change anything. Keys are scoped to Read, Edit or All, so a Read key literally cannot write. If you later automate writes, the CLI asks before each batch and an interrupted run resumes without creating duplicates.