AI to Chase Client Documents at an Accounting Firm: A Practical Workflow

AI to Chase Client Documents at an Accounting Firm: A Practical Workflow

To use AI to chase client documents at an accounting firm, start with a short workflow: a reminder calendar for each client, an automatic check of what arrived against the month's expected list, and a clear escalation to the accountant when the deadline is at risk. The AI drafts, organizes and checks what is objective. The accountant still decides what to do about a client who does not respond.

Key takeaways

• AI works best as a pending-items assistant: it reminds, receives, checks and records, without interpreting the client's tax situation.
• Each client needs a list of expected documents for each month. Without it, there is nothing to check against.
• Reminders follow a ladder: early notice, reminder, direct request and human escalation, all tied to the closing date.
• The accountant steps in by rule, not by accident: close deadline, unreadable document, mismatch or prolonged silence.
• Start with a small group of clients, measure what changes and only then expand.

Why document chasing stalls the monthly close and what AI can take over

In many firms the monthly routine is familiar: the team messages clients asking for bank statements, invoices, timesheets, expense receipts and sales reports. Half respond, some send incomplete files and others forget. The person doing the chasing is often the same analyst who should be closing the books. The problem is not a lack of method. Chasing is repetitive, depends on memory and uses up skilled attention.

This task has traits that suit automation: it recurs, it has clear rules (what is expected, by when) and it produces a verifiable result (the document arrived or it did not). It also has parts that do not suit automation: negotiating a deadline with a client in difficulty, deciding whether a missing item changes how the period is assessed, or interpreting an unusual document. That is why the split of responsibilities has to be explicit.

What the AI can take over:

• Sending reminders on the channel the client already uses, in the tone the firm defines.
• Receiving files and photos, identifying the document type and the period it covers.
• Comparing what arrived with the client's expected list for that month.
• Recording the status of each item and building the pending-items summary for the team.
• Answering simple questions, such as which format is accepted or where to find a statement.

What stays with the accountant:

• Any tax, fiscal or labor guidance.
• Accepting deadline exceptions or substitute documents.
• Firm-toned follow-ups, changes to contract terms and conversations about repeated missing documents.
• Validating documents that the check flagged as doubtful.

Write this boundary down before choosing any tool. The article on what not to delegate to AI in your business can help you formalize it with your team.

How to build the reminder workflow, step by step

The workflow below is a starting point. The days are relative to each firm's closing schedule, which you define.

• Define the list of expected documents per client and per month. Separate what is fixed (bank statement, payroll) from what is occasional (asset purchase invoice, new contract).
• Define each client's delivery deadline, calculated backward from the filing obligations the firm must meet.
• Set up the early notice: a message a few days before the delivery window opens, stating what will be needed and by when.
• Set up the first reminder for when the delivery window opens, with the list of items and how to send them.
• Set up the second reminder for a midpoint, naming only what is still missing, never the whole list again.
• Set up the direct request for close to the deadline, with the final date and a notice that the accountant will be brought in.
• Program the escalation: if the deadline passes without a response or with a critical item missing, the AI stops chasing and opens a case for the responsible accountant, with the full history.
• Record every message sent, every file received and every check decision in a log the team can consult.
• Review the workflow at the end of the month: which reminders worked, which clients needed escalation, which items on the list were wrong.

For the channel, WhatsApp is usually the one clients use most, but messages initiated by the business have their own template and consent rules. Read the guide on WhatsApp outbound messages, templates and opt-in before writing the reminder text, so you do not build a workflow that depends on practices the platform does not accept.

Checklist: what the agent verifies when a document arrives

Receiving the file is only half the work. A statement for the wrong month or a cropped photo creates as much rework as a missing document. The check can combine simple rules with AI reading of the content, and it should always produce one of three results: accepted, accepted with a note, or returned to the client. When the document is a PDF or an image, field extraction needs rules and human review for doubtful cases, as explained in the article on extracting data from PDFs with AI.

Checklist for each document received:

• The file opens and is readable, with no cropping, blur or missing pages.
• The document type matches an item on the client's expected list.
• The document period is the period being collected.
• The account holder or company on the document matches the client record.
• The document is complete: all accounts, all bank accounts or all units expected.
• It is not a duplicate of something already received that month.
• If there is a relevant amount or total, it was read and recorded for the accountant's later review, with the AI drawing no tax conclusions.
• The file was saved in the folder and under the naming standard the firm uses.

If any item fails, the AI replies to the client specifically (for example, asking for the statement of the other account) and updates the pending status. If the failure involves a question of interpretation, it forwards the case to the accountant instead of replying to the client.

When to escalate to the accountant: a decision guide

Escalation is what separates a reliable workflow from a message blaster. It must be defined by rule, so the accountant neither receives everything nor receives it too late.

Escalate immediately when:

• A filing obligation is close and a critical item is still unanswered.
• The client says they do not have the document, that they lost it, that the company changed banks or that something new happened (a new branch, a sale of an asset, a termination).
• The client asks a tax, labor or corporate question.
• The document received shows signs of inconsistency with the client record or with previous months.
• The client shows dissatisfaction, threatens to leave or complains about the follow-ups.

Escalate at the end of the reminder ladder when:

• The delivery deadline passes and any mandatory item is still missing.
• The same client is late month after month, so the accountant can consider a conversation about routine and deadlines.

Do not escalate, resolve it automatically:

• A request to resend an unreadable file.
• A question about the accepted format or the sending channel.
• Confirming receipt and saying thank you.

An escalated case should arrive with everything the accountant needs: who the client is, what was requested, when, what arrived, what is missing and the last part of the conversation. If the accountant has to ask what already happened, the escalation failed.

Illustrative scenario: the client who replies but never completes

Imagine, as an illustrative scenario, a firm with a services company that always sends the statement for one account and forgets the second. The early notice goes out at the start of the month. The client sends one PDF after the first reminder. The check compares it with the expected list and sees that two accounts are registered but only one arrived.

Instead of repeating the whole request, the agent replies: we received the statement for the main account, the one for the secondary account is still missing for this month. The client sends a crooked photo. The check marks it as unreadable and asks for a new photo, explaining how to frame it. At the direct request, the second statement is still missing, and the agent warns that the accountant will be notified if it does not arrive by the deadline.

The deadline passes. The agent opens a case for the responsible accountant with the history: two accounts expected, one received, the second requested twice, one unreadable attempt. The accountant calls or writes personally, finds out the secondary account was closed, updates the client record and adjusts the expected list for the coming months. The workflow stopped chasing something that no longer existed, and the lesson was stored in the record, not in someone's memory.

Data protection, WhatsApp and document security

Client documents contain personal and financial data. Before turning on any automation, the firm should know what data the agent will see, for what purpose, where it is stored and who has access. For official guidance on data protection in Brazil, the starting point is the website of the ANPD, the national data protection authority. If the project processes data at volume or in a sensitive way, check with your data protection officer or legal counsel whether to produce an impact assessment, as described in the guide to a data protection impact assessment for AI projects.

Good operating practices:

• Limit the agent's access to what is necessary: it needs to read the pending list and save files to the client's folder, not browse the firm's entire history.
• Avoid sending full documents to services the firm has not evaluated. If in doubt, address it in the internal AI use policy.
• Treat content received from clients as untrusted input. A file or message may contain instructions written to manipulate the agent, and OWASP's material on prompt injection prevention describes practical defenses, such as separating instructions from data and restricting the allowed actions.
• Give the agent only small, reversible actions: send a standard reminder, record a status, open a case. It should not change records, issue payment slips or answer questions about taxes.
• Keep the log of messages, files and decisions, so you can audit errors and prove what was requested and when.
• If the check classifies a client in a way that affects something important to them, review the design with attention to the right to review automated decisions.

On the channel, handle the number and consent consistently: the client should know they are talking to an assistant of the firm and always have the option to speak to a person.

Applying the method: map, limit, pilot, hand over

The method we use in AI agent projects fits this case well, without requiring a large rollout.

Map: document how chasing happens today. Who chases, through which channel, on which day, how many clients are late and which documents are missing most often. That picture is the baseline for the pilot.

Limit: decide what the agent can read, send and record, and what stays out of its reach. Write the escalation rules before configuring any message.

Pilot: choose a small group of clients with similar routines, run one monthly cycle with a full log and define the success metric in advance, such as time until the list is complete or the volume of manual follow-ups. The article on how to measure an AI automation pilot details decision criteria for expanding, adjusting or stopping.

Hand over: at the end of the pilot, give the team the workflow documentation, the rules, the reminder texts and the procedure for adjusting lists. The firm must be able to operate and change the workflow without depending on whoever built it. For building agents with few moving parts and clear control, Anthropic's guidance on building effective agents recommends starting simple and adding complexity only when there is a proven need.

Frequently asked questions

Can the AI chase the client on its own, with nobody at the firm watching?

It can send standard reminders and record replies, as long as the scope is limited and escalation follows rules. Firm follow-ups, deadline exceptions and any tax guidance should go through a person. We recommend that someone on the team reviews the pending-items summary at a set frequency, especially at the start.

Do I need to change my accounting software to use this workflow?

Usually not. The workflow starts with the list of expected documents, a messaging channel and a place to save files and statuses, which can be a spreadsheet or an organized folder. Integrations with the firm's system can come later, once the pilot shows they are worth the effort.

What should I do about a client who keeps sending the wrong document?

Use the check to return the specific reason with each submission and record the repetition. If the pattern persists for months, the case goes to the accountant, who can talk about the client's routine, simplify the list or agree on an easier way to send files.

How do I know if the pilot is working?

Compare it with the starting baseline: how many manual follow-ups the team still does, how long it takes for each client's list to be complete and how many cases reach the accountant. Also watch for check errors and client complaints. If the numbers improve without adding rework, it is worth expanding to another group.

Sources

• ANPD: Autoridade Nacional de Proteção de Dados
• OWASP: prevenção de prompt injection
• Anthropic: Building effective agents