AI Agents for Auditors: Chasing Evidence Requests From Clients

Yes, for the chasing. An agent can track the request list, draft reminders and pre-screen uploads, provided a named person approves what leaves the firm in the early rounds. It cannot accept evidence, run external confirmations or judge reliability. Those steps stay with audit staff, and each needs a dated sign-off.

The standards draw that line, not the software. Under the international auditing standards the auditor keeps control of confirmation requests, follow-ups included, and evaluates the reliability of information the client produces. A chasing agent works on the client's side of the table, asking for things; everything that judges evidence stays on the auditor's side.

Follow me on Instagram@sagnikteaches

Requests the agent may touch, and requests it may not

Sort the request list before any automation is switched on. The table maps each kind of request to the agent's role, with the paragraph of the standard that explains why. Paragraph numbers are those of the international text; national versions can add paragraphs, so ask your quality lead to confirm yours.

Connect on LinkedInSagnik Bhattacharya

"None" for confirmations does not mean AI has no place in that work. It means the process lives outside the chasing agent, run by the audit team, with addresses checked and replies returning to the firm. If a client starts routing its bank's replies through the finance team, that is a matter for the partner, not for a reminder.

Subscribe on YouTube@sagnikteaches

The tracker the agent reads and writes

Everything the agent does is a read or a write on one tracker, a shared spreadsheet or a table in your practice software. Build it first, with one row per request item:

• Identity: client, engagement, item number, description copied word for word from the request list, audit area.
• Dates: requested on, due on, last chased, chase count.
• People: client contact (the only address the agent may write to), staff owner, approver.
• Evidence: status, link to the stored file, and the acceptance initials and date.

Statuses are where control lives, so decide who may set each one.

The agent copies item descriptions and never rewords them, because a drifted description means the client sends the wrong thing, and the file then shows a request that no one approved. Seeing the tracker as the single source also means a person can reconstruct, at any time, what the agent knew when it wrote each message. The sibling tutorial on human approval steps in AI automations covers the mechanics of holding a draft until someone approves it.

A chase cadence with named approvers

Chase per client, not per item: one digest email listing everything outstanding for that engagement, so the client sees a single message. The cadence below assumes a request list with a due date; replace the day counts with your own.

Approval is a tick in a column, with initials, by a person on an approved list, and the sending step reads only rows with that tick. Three rules sit on top. The agent writes only to the address in the tracker's client contact column. It never adds recipients, such as the client's directors or those charged with governance, because widening the audience is a partner's decision. And it never mentions fees, penalties, the report date or anything about findings.

Add a column for the date the client promised. When a contact says an item will arrive on a named date, a person records it, and the agent leaves that item out of digests until the day after, then quotes the promise: "You told us on [date] that item 18 would arrive by [promised date]." Quoting the client's own words does more than a generic nudge, and it stops the agent chasing something the client has already scheduled.

Drafting the reminder: prompt, output and the edits

The prompt keeps the agent to a narrow job. Notice what it forbids as much as what it asks for.

You draft client reminder emails for a small audit firm. You do not send them.

Input: client contact first name, engagement name, a list of outstanding request items
(number, description, original due date) and the chase round (1, 2 or 3).

Write one email that:
- opens by naming the engagement and the date the items were due
- lists each outstanding item with its number and exactly the description given
- says what helps: original or system-generated files covering the whole period
- asks for a date by which the client expects to send each group of items
- is polite and plain; round 2 is a little firmer; never threatens
- never mentions fees, penalties, our report date, audit findings or conclusions
- never adds, removes or rewords items, and never advises on the client's accounts

Output: subject line, body, and a one-line note for the approver listing anything you were unsure of.

An illustrative round 2 draft, with the placeholders left in:

Subject: [Engagement name]: 9 items still outstanding

Hello [first name],

Thank you for the files sent since my last note. Nine items from our request list were due on
[due date] and have not reached us:

12. Aged debtors listing at the year-end date
14. Bank reconciliations for all accounts at the year-end date
18. Fixed asset register with additions and disposals for the year
(six more items listed in the same way)

Original or system-generated files covering the whole year are the most useful. If it is easier,
send each group as it is ready. Could you tell me by [date] when you expect each group?

If the audit may be delayed, please let me know.

Kind regards,
[your name]

Approver note: item 31 has no description in the tracker, so I left it out.

Two edits belong on that draft. First, "If the audit may be delayed, please let me know" suggests a consequence the firm has not agreed to say, and timing messages are the partner's call, so the approver deletes it and the prompt gains one more forbidden topic. Second, the approver note is valuable: item 31 has no description, so a person fixes the request list rather than letting the client guess. Keep an eye on the ratio of approver notes to drafts, since a rising number points to a messy request list, not a bad agent.

Pre-screening an upload without accepting it

When a file lands in the engagement folder, the agent can run mechanical checks and write the result beside the tracker row. It cannot, and must not, conclude that the evidence is good. The checks it can run:

• The file opens and is legible, not a password-locked or blank scan.
• The entity name matches the engagement.
• The date or period covers the year-end the request names.
• The document type matches the item description.
• The total agrees to the figure a staff member keyed in from the trial balance.
• It is not a copy of something already uploaded for another item.

You pre-screen one uploaded file against one request item. You do not accept evidence.
Input: the item number and description, the engagement's entity name, the year-end date,
the control total entered by staff (if any) and the file text.
Report only these: file legible (yes/no); entity name matches (yes/no/different, quote it);
period or date covers the year-end (yes/no, quote it); document type matches the description
(yes/no); total matches the control total (yes/no, give both figures); duplicate of an earlier
upload (yes/no). Then write PASS TO STAFF if every answer is yes, otherwise FLAG with the reasons.
Never say the evidence is sufficient, reliable or acceptable.

The last line of that prompt matters most. Words such as sufficient, reliable and acceptable belong to the auditor, so the agent is forbidden to use them. An illustrative run on six uploads for one client:

A pass means "worth a person's time", not "accepted". The aged debtors total matching the trial balance shows only that the client's schedule agrees with the client's own ledger. Whether the ledger is complete, and whether the schedule is reliable enough to rest a conclusion on, is the evaluation ISA 500 puts on the auditor. Likewise the lease row may be a false alarm, because the tenant could be a trading name; the staff owner decides.

Returning an item shows the same split. The agent's first suggestion for the current-account reconciliation was "Please resend the correct document". A staff member rewrote it to name the problem: "The reconciliation for the current account ends on 30 November. Please send the version at the year-end date, with the bank statement it was prepared from." The agent drafts; the person decides whether the message is specific enough to get the right file back first time.

Two realistic mistakes show why the status rules exist. A client's reply says "attached" and carries no attachment; an automation that trusts the words marks the item received, which is why the status changes only when a file is stored. A client uploads a bank statement printed from an edited spreadsheet; every automatic check passes, and only a person reading the source would doubt it, which is why originals matter in the standard.

Sign-off steps and what the file must show

ISA 230 requires the file to record who performed the audit work and when it was completed, and who reviewed it and when and to what extent. An agent's log is not a preparer's sign-off, so keep two records apart: the log shows what was sent and received, and people's initials show what was concluded.

The point of this table is that "the agent chased it" never appears in a signature column. If a regulator or peer reviewer asks who decided an item was good enough, the answer is a name and a date, written by a person.

A client's four weeks in numbers

The main example is a distribution client with 62 items on its request list, due on day 14. The numbers are illustrative; the shape is typical of a chase that works.

Of the 59 files received, the agent flagged 7: three wrong period, two unreadable, one name mismatch and one total that differed. Staff cleared the name mismatch as a trading name and returned the other six. So first-time acceptance was 53 of 59, about 90%. Time for this client: writing three digests by hand at 14 minutes each is 42 minutes; reviewing three drafts at 3 minutes each is 9, a saving of 33 minutes. Across 38 clients that is 1,254 minutes, roughly 21 hours, against perhaps 8 hours of set-up in the first season. Pre-screen flags add review time, but each flag is a problem that would otherwise have surfaced later, in the middle of testing.

Spreadsheet agent, practice platform or audit request tool

Three routes exist, and the right one depends on how many engagements you chase at once.

With fewer than about 30 audit clients and one or two seniors, the first two routes usually cover the need: start with rule-based reminders and add an AI drafting step only where wording takes time. A firm running hundreds of engagements will benefit from a purpose-built platform, so long as its prescreen agent follows the same status rules as above. For the practice-platform decision itself, see the comparison of TaxDome and Karbon.

Confidentiality and independence limits for the agent

Audit files hold payroll, bank and contract detail, so choose where the agent runs with care. Keep documents in the firm's own storage and give the AI step access to text and metadata only where it needs it. Business plans of ChatGPT and Claude are set not to train on business content by default, but confirm the setting and read the vendor's data terms before any client file passes through. The tutorial on using ChatGPT with client data safely lists the questions, and AI wording for engagement letters covers telling clients.

One independence point needs your ethics partner. An agent that chases is fine; one that helps the client prepare or correct its schedules is a different matter, because it may drift into doing management's work or reviewing its own output. Keep the agent's instructions limited to asking for items, and have it refuse questions about how a schedule should be built.

Starting small: three clients, one cycle

Pilot on three clients for a single fieldwork cycle, with every message approved by the senior, before touching the rest. Measure four things: days late per item before and after, first-time acceptance of uploads, edits the approver makes per draft, and wrong-recipient or wrong-client emails, which must be zero. One wrong-client email means pause the agent, find the cause in the tracker or prompt, and restart only after a person has tested the fix. Treat any of these as a stop signal and tell the partner:

• An email goes to an address that is not in the client contact column.
• A draft mentions fees, penalties, the report date or audit findings.
• The agent sets a status that only staff may set.
• A client complains about tone or frequency of reminders.
• The tracker and the engagement folder disagree about which files have arrived.

After one cycle, relax approvals for round 1 digests if edits per draft are near zero, and keep rounds 2 and 3 with managers and the partner. The wider tutorial on chasing missing client records has email wording you can adapt, and the working paper drafting tutorial shows where the same assistant stops.

Questions audit firms ask before switching on a chasing agent

Can the agent contact the client's bank for confirmations?

No. Paragraph 7 of ISA 505 says the auditor keeps control of external confirmation requests, including follow-up requests, and requests must be set up so that responses come back directly to the auditor. Run confirmations in a separate process the audit team controls, never through the client or a client-facing chasing agent.

Should we tell clients that an AI drafts our reminders?

Yes, in plain words in your engagement terms or privacy notice, together with how their documents are stored and which tools see them. The email still comes from a named person who approved it and answers for it. Ask your ethics or quality lead to agree the wording, and keep it consistent with your firm's AI policy.

Is the chase log part of the audit file?

Routine reminders are rarely significant matters, but ISA 230 does require documentation of significant matters discussed with management, so file any exchange where a client withheld, changed or disputed evidence. Keep the full chase log under your own retention policy either way, because it shows who sent what and when.

What happens when a client replies to the agent's email?

A person handles it. The agent only drafts outbound messages. Replies land in the senior's mailbox, any attachment is saved to the engagement folder for pre-screening, and a text reply is never answered by the agent. Questions about the audit itself, timing or fees go to the senior or partner.

Further reads

• Accountants AI Risk Checklist: Error Points, Each With a Named Check — A named check for each error point in an accounting workflow.
• Can AI Select an Audit Sample? Document and Defend It — Where AI helps with sample selection and how to defend it.
• AI for Accounting Teams in a Small Firm: Roles, Prompts and Sign-Off — Who writes prompts, who reviews, and how sign-off is shared.
• Accounting Firm Standard Operating Procedures With AI: Record to Publish — Turning a working method into a written procedure with AI.
• AI Agents for CPA Firms: Starting Jobs and a Review Gate for Each — Choosing the first agent jobs and a review gate for each.
• 7 AI Mistakes That Put an Accounting Firm's Client Trust at Risk — Seven AI slips that cost an accounting firm client trust.
• Spotting AI-Generated Evidence: A Checklist for Litigators — A working checklist for images, audio, video and documents: red flags, metadata checks, questions for the other side and when to instruct an expert.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: ISA 230, ISA 500, ISA 505 and ISA 580 (international auditing standards, paragraph references as published; national versions can add paragraphs); Suralink and Fieldguide websites; Karbon and TaxDome pricing pages; Zapier's Human in the Loop help article; n8n documentation on human review and the Wait node. Read in November 2026.