Accounting workflow automation: what AI can actually run in your firm.
Of 36 common accounting workflows, roughly 30 percent can run end to end today, about 47 percent need a human approving before anything posts, and the remaining 23 percent stay human. Which is which depends on your ledger. Set it below and the matrix adjusts.
2 of 36 workflows are limited on QuickBooks Online, because its API does not expose what they need. They stay listed, downgraded, since that gap is the thing worth knowing before you scope a project.
| Workflow | Function | Ceiling | Size |
|---|---|---|---|
| Close and reconciliation | End to end | M | |
| Firm operations | End to end | S | |
| Close and reconciliation | End to end | S | |
| Accounts receivable | End to end | M | |
| Accounts receivable | End to end | M | |
| Accounts payable | End to end | S | |
| Tax and compliance | End to end | S | |
| Accounts payable | End to end | S | |
| Reporting | End to end | M | |
| Tax and compliance | End to end | S | |
| Firm operations | End to end | S | |
| Accounts payable | Approval | S | |
| Tax and compliance | Approval | M | |
| Close and reconciliation | Approval | M | |
| Reporting | Approval | L | |
| Accounts receivable | Approval | M | |
| Firm operations | Approval | S | |
| Close and reconciliation | Approval | S | |
| Accounts receivable | Approval | S | |
| Close and reconciliation | Approval | M | |
| Accounts payable | Approval | S | |
| Firm operations | Approval | M | |
| Close and reconciliation | Approval | M | |
| Reporting | Approval | M | |
| Close and reconciliation | Approval | M | |
| Reporting | Approval | L | |
| Tax and compliance | Approval | M | |
| Accounts payable | Approval | M | |
| Firm operations | Human | M | |
| Reporting | Human | S | |
| Accounts receivable | Human | M | |
| Close and reconciliation | Humanledger limit | L | |
| Accounts payable | Human | L | |
| Tax and compliance | Human | M | |
| Accounts payable | Humanledger limit | M | |
| Accounts payable | Human | M |
Select a row for the human gate, the failure mode and the build size.
The short answer
An AI agent can own the high volume, rule-heavy, reversible parts of an accounting workflow: reading documents, matching, classifying, chasing, flagging. It cannot own the parts where the output is a professional opinion or where the action cannot be undone. Almost every failed accounting automation project we have seen failed by putting an agent on the wrong side of that line, not by picking the wrong model.
How to read the three levels
The classification is a ceiling, not an ambition. A workflow marked as needing approval does not become straight-through because the model improved; it stays supervised because the consequence of being wrong is borne by someone with a licence.
- Runs end to end. The agent completes the task. Humans audit a sample afterwards rather than checking each item. Appropriate when errors are visible and cheap to reverse.
- Needs approval. The agent does the work and stages the result. A person approves before it posts, sends or settles. Most accounting work sits here.
- Human decides. The agent drafts, assembles and prepares. A person decides every item. This is where regulated judgment and irreversible actions live.
Your ledger decides more than your budget
The same workflow runs at different levels on different systems, because what the API exposes sets the ceiling. NetSuite and Sage Intacct reach multi-entity and intercompany work that QuickBooks Online and Xero cannot. Spreadsheets expose nothing at all, so every row drops to draft-only and the honest first project is moving the data somewhere an agent can reach. The matrix downgrades rather than hides those rows, because the gap is the useful information.
What to build first, and what to leave alone
Good first builds share three properties: they are high volume so the saving is real, they are reversible so a mistake is recoverable, and they fail visibly rather than silently. Invoice capture, tax document classification, client document collection and card coding all qualify.
The ones to leave until the programme has a track record are payment execution and vendor bank detail changes. Both are fraud surfaces, and an incident on either one tends to end the whole initiative regardless of what else was working.
The audit trail is the product
For any workflow that touches the ledger, provenance matters more than accuracy. You need to replay how an entry came to exist: the inputs, the rule or prompt version, the model version, the named approver, the timestamp. An auditor asking about a journal eighteen months later is a normal event, and a system that cannot answer creates work rather than saving it. Provenance designed in at the first build is cheap. Retrofitted, it is not.
Questions firms ask before scoping
What can I actually automate in my accounting firm this quarter?
Start with the workflows that are high volume, rule-heavy and reversible. Invoice capture, tax document classification, GL coding suggestions, client document collection and credit card coding all fit that description, ship in days to weeks, and fail visibly rather than silently. The matrix marks these as good first builds. Avoid starting with anything that moves money or reaches a client unsupervised.
Can AI handle accounts payable and invoice coding without a human reviewing it?
Capture and extraction can run end to end with a sampled audit behind it. Coding should not. An agent proposing a GL code from vendor history is reliable for vendors it has seen and confidently wrong for ones it has not, so the review gate belongs on any code the agent has not used for that vendor before. Payment execution stays human entirely: that is the fraud surface.
Can AI do month-end close?
Parts of it. Checklist orchestration, anomaly detection and reconciliation matching are agent work. Judgment is not. Accruals and journals need a qualified approver before posting, and flux commentary needs the controller to edit before it reaches anyone. The realistic outcome is a close that takes fewer days with the same sign-offs, not a close that runs without accountants.
What can AI still not do in an accounting firm?
Anything where the output is a professional opinion or an irreversible action. Signing a return, accepting an engagement, releasing a payment, deciding a write-off, resolving a client dispute. Agents also cannot be relied on to notice that a question was the wrong one, which is most of what experienced accountants are actually paid for.
How do I automate bank reconciliation in Xero or NetSuite?
The mechanics are the same on both: pull the bank feed, match against ledger entries, propose the exceptions and route them. What changes is what the API exposes. NetSuite and Sage Intacct support multi-entity and intercompany work that QuickBooks Online and Xero cannot reach, so the same workflow runs at a different level depending on the ledger. Set your ledger in the matrix and the rows adjust.
What if our data lives in spreadsheets?
Then none of this applies yet, and the matrix will tell you so. An agent needs a system to act against. If the ledger is a spreadsheet there is no API, no permission model and no audit trail, so every workflow drops to draft-only. The first project is getting the data into a ledger that exposes an API. Anything else is building on sand.
How do I keep an audit trail for AI-generated journal entries?
Store the inputs, the prompt or rule version, the model version, the proposed output, the named human approver and the timestamp, against the entry itself. The test is whether you can replay how an entry came to exist eighteen months later when an auditor asks. If provenance is not designed in from the first build, it does not get retrofitted cheaply.
How do I stop client data leaking between clients?
Isolation has to be architectural, not a prompt instruction. Separate retrieval indexes per client, permission scopes enforced at the data layer rather than the model layer, and no shared fine-tuning across client data without explicit consent. Cross-client benchmarking is the workflow most likely to leak, which is why it belongs in the architecture conversation early rather than as a later feature.
What should an accounting firm never automate first?
Payment execution, vendor bank detail changes, engagement acceptance and anything client-facing that could read as advice. These are either fraud surfaces or regulated judgment. They are not permanently off limits, but making one of them your first project is how a programme gets shut down after a single incident.
Can this work without an IT team?
Yes, though not without someone owning it. The ledger integrations are the technical part and an implementation partner handles those. What cannot be outsourced is deciding where the human gates sit and who approves what, because those are firm policy decisions. Budget for a named internal owner even if you buy the engineering.
Where the pressure to automate is coming from
The staffing picture is the reason this question is urgent rather than interesting. The AICPA has documented a sustained decline in accounting graduates entering the profession while a large share of licensed CPAs approach retirement, and firms have been competing for a shrinking pool of qualified staff for several years. That is the actual driver: not a desire to cut headcount, but an inability to hire the headcount the work requires.
It changes what a sensible first project looks like. If the goal were cost reduction you would automate the largest cost centre. If the goal is capacity, you automate the work that is stopping qualified people from doing qualified work: document chasing, coding suggestions, reconciliation matching, classification. That is why the good-first-build markers in the matrix sit on volume workflows rather than on the technically impressive ones.
- AICPA Trends Report, on accounting graduate and CPA pipeline numbers.
- Journal of Accountancy, ongoing coverage of staffing and technology adoption in practice.
- QuickBooks Online API and Xero Accounting API, the reference for what each ledger actually exposes.
A realistic first ninety days
Firms that get value quickly tend to follow the same shape, and it is unglamorous. Weeks one and two go on picking a single workflow and writing down what good looks like, including the error rate you would accept and who reviews what. Weeks three to six build it against real data rather than a sample, because the exceptions are the work. Weeks seven to ten run it in parallel with the existing process, which feels wasteful and is the only way to know whether it is right. Weeks eleven and twelve either cut over or stop.
The step firms skip is parallel running, and skipping it is what produces the projects that get quietly abandoned in month four. An agent that is 95 percent accurate sounds excellent until you learn that the 5 percent clusters on exactly the cases that matter. You only find that out by running both and comparing.
What this costs, honestly
The build sizes in the matrix are bands rather than prices because the price is set by your data, not by the workflow. A firm with a clean ledger, consistent coding and one entity will pay a fraction of what a firm with four acquisitions and inconsistent charts of accounts pays for the identical workflow. Anyone quoting a fixed price without looking at your data is quoting the demo, not the build.
The band that surprises firms is the small one. An S workflow is genuinely days to a couple of weeks, which means the first project should not be a procurement exercise. If a partner proposes a six-month evaluation before automating invoice capture, the evaluation costs more than the build.
Method and limits
Each workflow is classified by the level a competent implementation can reach today with the named human gate in place, not by what a vendor demo shows. Ledger support reflects whether the platform exposes what the workflow needs through its API. Build sizes are relative bands, S through L, deliberately not dollar figures, because cost depends on your data quality and integration surface far more than on the workflow itself. Where a workflow could be argued into a higher level, it is classified conservatively. Nothing here is transmitted anywhere; your selections stay in your browser.
Want agents like these in your stack?
Book a free assessment, we'll map where an AI agent creates real leverage in your workflows and scope the first one to ship.