How accounting firms can automate document chasing with AI

Ask anyone who works in a small accounting or bookkeeping firm what eats their week, and document chasing comes up fast. Bank statements, receipts, payroll reports, signed letters: the work can’t start until they arrive, and getting them means writing the same polite email again and again, then keeping track of who replied.

It’s usually the first thing we’d automate. It happens often, everyone dislikes it, and the risk is low because the automation asks for things rather than deciding anything.

Why is document chasing worth automating first?

Three reasons.

• Volume. A firm with a few hundred active clients sends dozens of requests and follow-ups a week, more around deadlines.
• It’s pattern work. Work out what’s missing, ask for it, wait, ask again more firmly. That’s a loop software handles well.
• Low risk. A slightly clumsy reminder is an easy mistake to recover from. A wrong tax position isn’t. Starting with chasing lets your team build trust in automation where mistakes are cheap.

How does automated document chasing work?

Every automation in our Capacity Playbook follows the same shape: trigger → AI step → human check → output. For chasing, it looks like this.

1. Agree what “complete” means. For each job type (year-end accounts, a VAT or sales tax return, a monthly bookkeeping close), write down the documents you need. This is the step most firms skip, and it’s the one that makes the rest work.
2. Put the checklist where your jobs live. Your practice management tool (Karbon, TaxDome or similar) or, to start, a shared spreadsheet: one row per client per job, one column per document.
3. Let a workflow tool find the gaps. n8n, Zapier or Make checks each open job against its checklist on a schedule, or when a deadline gets close.
4. Let the AI write the request. Claude or ChatGPT drafts a short, specific message for each client: what’s missing, for which period, and by when. Specific requests get answered faster than “please send your documents”.
5. Keep a person on the send button at first. For the first few weeks, someone approves every message. Once the drafts are consistently right, routine reminders go automatically and only unusual cases get flagged.
6. Escalate on a timetable. A friendly first ask, a firmer second, and on the third a task for a person to pick up the phone.

The output is clients getting clear requests, and a job record that shows what was asked for, when, and what came back.

Where does a person stay in charge?

Automation handles the chasing. People handle the relationship. In practice:

• A person approves the message templates and the tone before anything goes live.
• A person approves each message for the first few weeks.
• Anything unusual goes to a person: a client who’s gone quiet after three reminders, a reply that sounds unhappy, a gap that suggests something bigger is wrong.
• A person decides when it’s time to call.

For more on where to draw these lines, see Is it safe to let AI read your clients’ emails?

What tools do you need?

Usually less than you’d think:

• A source of truth for jobs and checklists: your practice management tool or a spreadsheet.
• A workflow tool: n8n, Zapier or Make for the schedule and the logic.
• A language model: Claude or ChatGPT, on a business plan or through the API, for the drafting.
• Your email: Microsoft 365 or Google Workspace, so messages come from your firm, not a no-reply address.

Before building anything, check what your practice management tool already does. Many have client requests and automatic reminders built in. If yours does, use them for sending and add AI where it helps: working out exactly what’s missing, and writing the request in plain language.

How many hours does it save?

Don’t trust anyone’s headline number, including ours. Work out your own:

Hours saved per week = times per week × minutes each × share automated ÷ 60

Say your team sends about 60 chasing emails and follow-ups a week. Each takes about 6 minutes once you count looking up what’s missing, writing the email and logging it. If automation handles 70% end to end:

60 × 6 × 0.7 ÷ 60 = about 4.2 hours a week

That’s our estimate for an example firm, not a measured result. For a firm with a few hundred active clients, we’d expect something in the range of 3–8 hours a week, and more in the weeks before a deadline. To see what those hours are worth in revenue, read the capacity math.

What goes wrong?

The common failure modes, so you can avoid them:

• No agreed checklist. If the team doesn’t agree what “complete” means, the automation asks for the wrong things. Fix the checklist first.
• Out-of-date job data. If documents arrive but nobody marks them received, clients get chased for things they’ve already sent. Make “received” easy to record, or detect it automatically from the inbox or portal.
• Too many reminders. Cap them. After the third, a person takes over.
• Nobody owns it. Automations break quietly when a tool changes. Someone should check weekly that it’s still running.

Where to start

Pick one job type, ideally the one with the most volume this quarter, and write its checklist. Run the automation for that job type only, with a person approving every message. After two or three weeks, look at what the team changed in the drafts, tune it, and then expand.

If you’d like help, that’s what our AI Ops Audit is for: in two weeks we map how your firm actually runs, score every automation opportunity, and get one working quick win live. Chasing is often that quick win. Book a 30-minute call, or start with the free Capacity Playbook for nine more automations like this one.