fficary
PricingCompareMigrateResourcesHow it works
Sign inStart free

Resources / AI

AI

AI for accounting firms: what is safe to automate, and what is not

Firms are already using AI. The question is not whether — it is which tasks can touch client data, who trains on it, and what still needs a human.

8 min read · 2026-08-23

Someone in your office is already using AI for something the partner has not approved. That is the real adoption curve. Tools spread faster than policy. The firm still signs the return.

This is not an argument against AI. It is a sorting guide. Automate the boring, inspectable work. Keep a human on anything that changes a client’s legal or tax position. Write down where the data goes.

Three buckets of work

1. Draft and fetch — usually fine

  • Summarize a long client email into a task list.
  • Draft a reminder in the firm’s tone, then send it after a person reads it.
  • Pull a vendor name and amount off a receipt image.
  • Suggest a category the firm has used for that merchant before.

2. Propose a ledger change — fine with a confirm step

A suggestion that a human can accept or reject belongs in an inbox, not a silent background job. ficary posts high-confidence, rule-backed activity and holds the rest. That is the pattern to copy even if you never use ficary: AI proposes, the ledger only moves when a rule or a person says so.

3. Legal, tax, and “just send it” — not fine

  • Filing or e-signing anything.
  • Deciding whether a payment is deductible, capitalized, or personal.
  • Pasting a full return, SSN, or payroll file into a consumer chatbot.
  • Letting a model invent journal entries to “make it balance.”

Questions that belong in the vendor meeting

Accounting Today has been telling firms to stop accepting the slide and start reading the terms. Use these out loud. Ask the rep to scroll to the paragraph, not paraphrase it.

  1. Is client content used to train any model — yours, a subprocessor’s, or a future one?
  2. Where is inference run, and who can see prompts in logs?
  3. Is there a signed DPA, and does it cover AI subprocessors by name?
  4. Can the firm bring its own key or model later without a migration?
  5. What does a deletion request actually delete — records, embeddings, backups?

What ficary will and will not claim

Today, ficary can use configured AI providers for extraction and category help. Launch checks require real provider setup; we do not treat those features as ready on placeholder keys. Isolated databases per firm and private inference are the enterprise direction, not a live production claim. If a salesperson cannot say that sentence, do not buy from them either.

Put it in the WISP, or it did not happen

The FTC Safeguards Rule expects a written information security program. If staff can paste client data into a personal ChatGPT account, your WISP is fiction. A usable policy is short: approved tools, banned tools, what data may enter a prompt, and who reviews model output before it hits a client or the ledger.

IRC 7216 is the tax-specific tripwire: disclosing return information to a third party generally needs consent in the form the regulations describe. “The AI vendor is a processor” is a legal conclusion, not a slogan. Get counsel involved before you pipe returns through a new model.

A practice you can start this week

  1. Inventory every AI tool already in use, including browser extensions.
  2. Kill consumer accounts for client work in writing.
  3. Move category and receipt help into a system that shows a confidence and a confirm button.
  4. Keep tax research in tools with a DPA, or do it without client identifiers.
  5. Review one week of model suggestions the way you review a junior’s workpapers.

For the buying side of the same problem, see How to choose accounting practice management software.

Questions firms actually ask

Can a CPA firm put client data into ChatGPT?

Not on a consumer plan, and not without a written process. Return information and client files can trigger IRC 7216, ethical confidentiality rules, and the FTC Safeguards Rule. Use a contracted vendor with a DPA, or keep the data out.

Does “enterprise AI” mean the vendor will not train on my data?

No. Marketing pages and the actual terms often disagree. Ask the vendor to point at the clause. If they cannot, treat training as allowed.

What accounting work is actually a good fit for AI today?

Drafting emails, summarizing a thread, suggesting a category, and extracting a receipt. Posting to the ledger, signing a return, and deciding tax treatment still need a named human.

Keep reading

AI

An accounting firm AI policy you can put in the WISP this week

A one-page AI policy for CPA and bookkeeping firms: approved tools, banned pastes, confirm-before-post, and the questions vendors must answer in writing.

Buying

How to choose accounting practice management software

A buying guide for CPA and bookkeeping firms: workflow, email, portal, AI terms, TCO, and the ledger question most demos skip.

CAS

What is Client Accounting Services (CAS)? A firm guide

CAS is outsourced accounting a firm runs for clients — bookkeeping through advisory. Here is how to price it, staff it, and keep the books honest.

Books first

See the leftover inbox on your own activity.

Start free. Invite your accountant when the books are worth handing over.

Start freeTalk to us
fficary

The money OS for owners and firms. ficary posts the obvious, collects what you invoice, and holds the leftovers for you.

Product

PricingCompareMigrateResourcesHow it worksPractice softwarePrivate AISecurity

Company

ContactTermsPrivacySecurity

Account

Sign inStart free