AI for Bookkeeping Firms: Routing Receipts and Client Requests
Learn how a bookkeeping firm can route receipts and document requests through AI agents now, with a staff accountant approving every categorization.

An AI agent can help a bookkeeping firm by collecting receipts and client document requests, reading them, proposing a ledger account, and placing each item in a review queue. A staff accountant approves every categorization before anything posts to a client's books. The firm keeps the decisions, and the agent removes the sorting and chasing that fills the week.
Key takeaways
- Start with one intake route, such as receipts from one group of clients, before adding more channels.
- Let agents draft categorizations and document requests, but never let them post to the general ledger without approval.
- Give each agent only the memory, tools and permissions its single job requires.
- Measure accuracy by how often accountants change proposals, not by how fast items move through the queue.
- Treat receipts, invoices and client emails as untrusted input, because any document can contain text meant to steer the agent.
Where the hours go in a bookkeeping week
Most of the time a bookkeeping team spends on receipts is not spent on the ledger itself. It goes to collecting the files, separating photos from emails, identifying which client a file belongs to, asking for the missing page of a statement, and remembering which card merchant corresponds to which supplier. The entry is often the shortest step. An accountant who opens a folder of undated images has to do intake work before any accounting starts.
Document requests follow the same pattern. A request goes out, the client replies with one of three files, the accountant notes what is still missing, and the chase repeats. The work is repetitive, easy to delay, and easy to lose track of across a large client list. The same intake problem appears in import and export document tracking, where status depends on which document arrived and which one is still open.
That is the work an agent should absorb. The goal is not a system that closes books on its own. The goal is a queue of clearly labeled items, each with a proposed treatment and a visible reason, so the accountant spends attention on judgment instead of on sorting.
Which tasks an agent should route, and which it should never touch
Draw the line before building anything. A practical decision guide for a bookkeeping firm looks like this:
- Receipt intake: the agent may receive, read, identify the client and date, and store the file with its original source. It may not discard a file.
- Extraction: the agent may propose vendor, amount, date and tax fields, and it must mark any field it could not read clearly.
- Categorization: the agent may propose a ledger account with a short reason. It may not post the entry.
- Document requests: the agent may draft a request from an approved template and queue it. A person releases it to the client until the template and tone have been reviewed.
- Ledger posting, reconciliations marked complete, tax positions and any client advice: staff only.
- Chart of accounts, bank rules and recurring entries: staff only, with a written change record.
A step-by-step plan for the first pilot
Start narrow. Choose one client group or one receipt channel, usually a set of clients who already send photos through a single route. Then follow these steps in order:
- Step 1: Write the agent's single job in one sentence, for example: route receipts from these clients into the review queue with a proposed account and a source reference.
- Step 2: Define the success measure before the first run. Use the rate at which accountants accept proposals unchanged and the count of items that reached the wrong client folder, not throughput.
- Step 3: List the memory the agent needs: each client's chart of accounts, known vendor names, prior approved categorizations and the fiscal calendar. Keep each client's memory separate.
- Step 4: Grant tools narrowly. Read access to the intake folder and write access to the review queue only. No posting rights and no access to bank credentials.
- Step 5: Run in shadow mode for a period agreed with the accountant lead. Proposals are visible but not used, and they are compared with what staff actually booked.
- Step 6: Review every disagreement together, adjust the instructions or memory, and only then move the queue into daily use.
- Step 7: Schedule a recurring evaluation and a permission review, so access that is no longer needed is removed.
How the approval step works in daily use
Each proposal should arrive as a card the accountant can act on in seconds. It shows the source image, the extracted fields, the proposed account, the reason in plain language, and any field the agent flagged as unreadable. The accountant accepts, edits or rejects. Rejections are the most useful signal, so each one should carry a short reason code such as wrong client, personal expense, split receipt or new vendor. Those patterns become client memory, but they do not give the agent permission to change its own behavior.
Approval has to be real, not a formality. If a queue is cleared in bulk each afternoon without anyone opening the source images, the control exists only on paper. Items with low legibility, unusual amounts or unfamiliar vendors should require an explicit decision, while routine items from known vendors can move faster. The thresholds belong to the accountant lead, they should be written down, and they should be reviewed whenever the client mix changes. Queues of this kind resemble timesheet approvals at staffing agencies, where a person signs off before anything is billed.
Illustrative scenario: a small firm receives a photographed fuel receipt from a client who is the sole owner of a business, and a personal card is visible in the corner of the image. The agent proposes a vehicle expense account, flags that the card may belong to the owner, and marks the item as needing a personal-use decision. The accountant sends a short question through the approved template and holds the item until the answer arrives. Nothing posts in the meantime.
Protecting the queue from manipulated documents
Receipts, invoices and client emails are content an agent reads, and that content can include text written to influence it. A scanned invoice might contain a line asking the system to ignore its rules or to classify a payment as a refund. The OWASP guidance on prompt injection treats this as a core risk for any language model that processes external content. The practical answer is structural: separate document content from instructions, never let text inside a document grant permissions, and log every tool call.
In practice, the agent cannot widen its own access, cannot post entries and cannot send messages outside approved templates. Its outputs are checked against the chart of accounts before they reach the queue, so an account code that does not exist is rejected automatically. Anything unusual, such as a changed bank detail inside an invoice, goes to a person as a flag, never to an automatic correction. The OWASP Authorization Cheat Sheet is a useful reference for deciding who may do what in these workflows.
What the agent should remember, and what it should not
Memory is what makes the second month more accurate than the first, but it needs boundaries. Store approved decisions, client-specific vendor names, the version of each chart of accounts and the rules the accountant has written down. Do not store raw bank credentials, full statements beyond what a task requires, or notes from one client that could leak into another client's work. Each client's memory should be scoped, expirable and auditable, so the firm can answer what the agent knew when a given proposal was made.
The NIST AI Risk Management Framework is a useful reference for these choices because it asks organizations to map, measure and manage AI risk with named people responsible for each part. Firms handling personal data also need to check where records are stored, who processes them and which privacy rules apply in each jurisdiction where clients are located. Those rules sit alongside the agent design and should be reviewed with counsel.
How Be Human runs this as a managed service
Be Human designs the first engagement around one useful role, here the receipt and request intake agent. The work defines the memory and tools that role needs, the approval boundaries, and a measurable success criterion agreed with the accountant lead. The agent keeps its identity, memory, channels and permissions under the firm's control, and the firm can see what the agent knows and what it has done. After launch, managed operations cover queue monitoring, recurring evaluations against accountant decisions, maintenance when a document source or accounting integration changes, permission reviews, incident response and controlled improvements to instructions and memory. The service is delivered remotely, in English or Portuguese. If you want to find which of your intake and follow-up tasks is the best first candidate for an agent, take the free Be Human AI diagnostic.
Frequently asked questions
Can an AI agent post entries directly to a client's books? In the approach described here, no. The agent proposes categorizations and places them in a review queue, and a staff accountant decides what posts. Posting rights are withheld from the agent so that a wrong proposal cannot reach the ledger without a person seeing it first.
What should a bookkeeping firm measure during a pilot? Track how often accountants accept proposals unchanged, how many items reached the wrong client, and how much staff time intake takes compared with the period before the pilot. Do not judge the pilot on throughput alone, because a fast queue that is approved without review creates risk instead of value.
Does this replace bookkeepers? It takes over sorting, chasing and first-pass proposals, which are the parts of the week that delay judgment. Accountants still review, decide and sign off on every posted entry. The role shifts toward review, exceptions and client communication.
How does the firm keep client data separate? Each client has scoped memory, restricted tool access and its own audit trail, so the agent working on one client cannot read notes from another. Permissions are reviewed on a schedule, and access that is no longer needed is removed. Data handling rules for each jurisdiction should be confirmed with the firm's advisers.
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


