AI for accounting firms: the work it can take on, and the client data rules that come first

How do accounting firms use AI today?

A lot of a firm’s hours go to moving information from one place to another: from a client’s upload into a workpaper, from a bank statement into the ledger, from an IRS notice into a reply. That is the part of the work AI changes.

The Bureau of Labor Statistics covers this in its outlook for accountants and auditors, last updated in October 2026. It notes that they may use AI and robotic process automation to work more productively, and that automating routine tasks such as data entry will make their advisory and analytical work more prominent.

AI in accounting shows up in four places today:

• Inside the ledger and tax software: suggested categories for bank transactions, matches for reconciliations, and reading of uploaded documents.
• In the client portal: sorting and naming what clients upload, and spotting what is missing.
• In a general assistant: drafting client emails, engagement letters and memos, and summarizing long documents.
• In custom automation: moving data between systems that don’t connect, with AI reading the unstructured parts on the way.

What accounting tasks can AI automate?

AI handles the reading, the matching and the first draft. The reviewer handles the judgment. Here is how common tasks split:

AI for accountants is a drafting and sorting aid, and the person who signs is still the preparer. A model can write a fluent summary of a tax rule that cites a section that doesn’t say what the summary claims. Every figure and citation it produces stays a draft until someone checks it against the source.

Which AI is best for accounting?

There’s no single best AI for accounting, because firms use it for different jobs. Work through the options in this order:

1. What your software already includes: the AI features in your ledger, tax and practice software work on data that is already there, under a contract you’ve already reviewed.
2. A general assistant on a business account: for drafting and summarizing, on terms that keep your inputs out of model training and let you control who has access.
3. A document tool: for reading receipts, statements and tax forms in volume, when your portal doesn’t do it well.
4. Custom automation: for work that crosses systems your products don’t connect, such as client uploads that need to reach the right workpaper.

Whichever you pick, put the same five questions to each tool:

• Training: are client inputs used to train the provider’s models?
• Retention and location: how long is data kept, and where?
• Access: is there multi-factor sign-in, user roles and a log of who opened what?
• Contract: does the agreement set security expectations you can check? The FTC’s Safeguards Rule expects your contracts with service providers to spell them out.
• Exit: can you export your data and have it deleted when you leave?

What rules apply to client data in an AI tool?

Three federal requirements shape what a firm may do with client data. Each applies to an AI tool the same way it applies to any other vendor:

• Section 7216 of the tax code: anyone in the business of preparing returns, or of providing services connected with preparing them, who knowingly or recklessly discloses a client’s return information, or uses it for anything other than preparing the return, commits a misdemeanor. The penalty is a fine of up to $1,000 in most cases, up to a year in prison, or both. The Treasury regulations list uses that need no consent. For anything else, they require the taxpayer’s written consent, signed and dated, before the information is used or disclosed, and a consent given afterward doesn’t count.
• The FTC Safeguards Rule: the FTC lists tax preparation firms among the financial institutions it covers. A firm has to oversee its service providers, with contracts that spell out security expectations and periodic reassessments, encrypt customer information on its systems and in transit, and use multi-factor authentication. Since May 2024, a firm must also notify the FTC no later than 30 days after discovering that the unencrypted information of at least 500 consumers was acquired without authorization.
• A written information security plan: in August 2026, the IRS reminded tax and accounting professionals that federal law requires them to create and maintain one. The requirement comes from the Gramm-Leach-Bliley Act, which treats tax and accounting professionals as financial institutions, and the IRS points to Publication 5708 as a template.

So before client data goes into a new tool, add the tool to your written plan, check its contract against the Safeguards Rule, and decide whether each use of return information is one the regulations already allow or one that needs the client’s signed consent first.

Is it safe to put client financial data into an AI tool?

It can be. Safety depends mostly on the account and the setup. A free personal account with a client’s return pasted in fails most of the tests above: nobody at the firm controls retention or training, there’s no contract with security terms, and the use isn’t in your written plan. A business account set up for the firm can pass them.

A setup that holds up has five parts:

• A firm account: the firm controls who has access, how long data is kept and what happens when someone leaves.
• Terms you have read: inputs kept out of training, a stated retention period, and a contract with security expectations.
• The least data the task needs: mask Social Security numbers and account numbers unless the task requires them.
• Multi-factor sign-in and a log: so you can tell who used it, when and on which client.
• Consent where it’s needed: for return information, the client’s written consent before any use the regulations don’t already allow.

What does a reviewer still check?

Everything that goes out under the firm’s name. AI output is a draft, and the review is what turns it into work product:

• Figures against the source: totals tied to the bank statement, the invoice or the form itself.
• Judgment calls: classification, materiality and positions on a return. Software can suggest, and the accountant decides.
• Citations: every code section, ruling or form it names, opened and read.
• Anything sent to a client: the dates, amounts and promises in an email or letter.
• The trail: what the tool did, what the reviewer changed and who signed off.

Keep reviewing at full depth until the error rate on a task has been measured over a full cycle, such as a month-end close or a filing season. Only then decide whether sampling is enough.

Will AI replace accountants?

Not according to the Bureau of Labor Statistics. Its October 2026 outlook projects employment of accountants and auditors to grow 5% from 2025 to 2035, faster than average, adding 79,400 jobs, with 115,300 openings a year on average. There were 1,595,200 accountant and auditor jobs in 2025.

The change BLS describes is in the mix of the work: routine tasks such as data entry get automated, and advisory and analytical work takes a larger share of the job. For a firm, the useful question is what its accountants’ hours go to.

When is plain automation enough?

Not every slow task needs AI. When the input is structured and the rule never changes, plain automation costs less to run and is easier to check:

• Bank feed rules: the same payee always goes to the same account.
• Recurring entries: monthly accruals and depreciation posted on a schedule.
• Requests and deadline notices: document requests and filing-date notices sent on a calendar.
• E-signature flows: engagement letters sent, signed and filed without retyping.

AI earns its running cost where the input is unstructured: a scanned receipt, a client’s email, a notice in its own format. For the plain parts, the Zapier, Make, n8n or custom code guide and our guide to whether to build or buy software cover the trade-offs.

Where should an accounting firm start?

1. Update your written security plan: list the AI tools in use, who may use them and on what data. If you don’t have a plan yet, IRS Publication 5708 gives you a template.
2. Pick one task: high volume, a clear right answer and low risk, such as sorting client uploads or matching bank lines.
3. Check the tool against the rules: training, retention, contract terms, multi-factor sign-in, and whether the task touches return information that needs consent.
4. Test it on last season’s files: run the tool on finished work and compare its output with what your team produced.
5. Set the review rule: who checks the output, at what depth, and what gets logged.
6. Measure a full cycle: hours saved, errors caught and running cost over a month-end close or a filing season, before you add a second task.

Our AI readiness checklist has more on getting a firm’s data and systems ready.

Our page for accounting firms says what we would build for a practice like yours, and a 15-minute call is enough to talk through a first task.