92% of accounting professionals are now using AI in some form. That number is from Karbon's State of AI in Accounting 2026 Report, and it caught us sideways because the same report had it at 9% in 2024. Two years. From single digits to near-universal adoption.
But here's the part nobody puts in the headline: 37% of the time AI saves accountants gets eaten back up by reviewing and correcting its outputs. Accounting Today calls it the rework tax, and it's the reason most "AI for accountants" articles feel a little hollow when you actually try to use them.
We build AI systems for service businesses, including a handful of small accounting firms. So this isn't a vendor pitch or a Stanford research summary. It's what's working in 2026 for solo CPAs, two-person bookkeeping shops, and 5-10 person firms. What to automate first. What to leave alone. The tools we actually pay for. And the workflows we've watched fall flat.
If you're running a small firm and trying to figure out where to start, this is the playbook.
AI for accountants in 2026: the honest picture
When people say "AI for accountants" they're usually mashing together three different things.
The first is generative AI — ChatGPT, Claude, Gemini. Good at writing, summarizing, explaining tax code in plain English to clients. Bad at math without help, bad at consistency, bad at citing sources unless you build the right scaffolding around it.
The second is agentic AI — systems that take actions, not just produce text. Pulling data from a bank feed, categorizing it, posting to the GL, flagging exceptions. This is where the real time savings live, and it's what most accountants haven't actually deployed yet.
The third is ML-powered platforms — Vic.ai, Digits, Trullion, Botkeeper. Purpose-built tools that learned on accounting data. You don't write prompts. You upload documents and the platform does its thing.
All three are useful. They cost different amounts. They fail in different ways. And the firms getting real ROI in 2026 are mixing all three on purpose, not picking one and calling it a strategy.
The numbers worth knowing:
- Stanford Graduate School of Business research from 2025 found accountants using AI close their monthly books 7.5 days faster.
- Gartner's 2024 Productivity Impact Survey put gross time savings at 5.4 hours per week.
- The same Karbon report shows firms with a documented AI policy save 17% more time than firms winging it.
Net out the rework tax and you're looking at maybe 3.4 real hours saved per week per person. Not a revolution. Not nothing either. At a 5-person firm billing $200/hour, that's around $176,000 a year of capacity unlocked if you bill it out or close earlier.
That's the honest picture. Now let's get into what to actually do with it.
What AI actually does for accountants right now
Five categories, and they don't all matter equally for small firms.
Bookkeeping and AP/AR
This is where the dollars are. Receipt capture, bill coding, bank reconciliation, AP processing. The platforms that work well here learned on millions of transactions, so they're not guessing — they know what a Verizon bill looks like for a marketing agency vs. a contractor.
Tools we see working in small firms: Dext for receipt capture, Vic.ai for AP, Ramp for cards plus AP, and the native AI categorization inside QuickBooks Online and Xero. Botkeeper if you want a managed layer on top.
If you serve small business clients on the bookkeeping side, the SMB-owner perspective on this same problem is AI bookkeeping for small business — worth sending to clients who keep asking "should I just use the AI?"
Tax preparation
According to Thomson Reuters' 2025 Future of Professionals Report, AI cut tax prep time by 55% on average, and some firms hit 80% automation on individual returns. The wins are mostly in document extraction (K-1s, 1099s, W-2s, brokerage statements) and first-pass return prep.
Tools: TaxGPT for research, Aiwyn for engagement-to-return workflow, the new AI features inside CCH Axcess and Lacerte, and Claude or ChatGPT for client memos.
Audit and compliance
Trullion for lease accounting and revenue recognition. MindBridge for risk-flagging on full ledgers. For small firms doing reviews and compilations, this category is overkill unless you have a niche. Skip it for now if you're sub-$2M in revenue.
Client advisory services (CAS)
This is the one most firms talk about and the one we see most struggle with. The pitch is: AI does the compliance work, you spend more time on advisory. Real talk — advisory takes a different skill set than tax prep. AI doesn't give you that skill set. It just frees up the hours. What you do with those hours is on you.
Useful tools here: Karbon for client comms, ChatGPT or Claude for advisory memo drafting, Spotlight or Fathom for management reporting visualizations.
Internal firm operations
Hiring, content, proposals, scheduling. AI is great here because the stakes are low. A bad LinkedIn draft costs you nothing. A bad GL entry costs you a client. Start here if you're nervous about AI in client work.
The 5 highest-ROI AI workflows to automate first
If you're going to do one thing this quarter, do one of these. In order.
1. Receipt and bill processing
What it does: client emails a photo of a receipt to a shared inbox, AI extracts the vendor, date, amount, GL code, and posts it to QBO or Xero with the receipt attached. Or upload a stack of bills and the system runs them through the same pipeline.
What to use: Dext (now part of Sage) is the safest pick. Vic.ai if you want AP-specific learning. Ramp if your clients use Ramp cards already and you want the whole loop in one place.
Time saved: We've measured this with one of our clients — a 6-person bookkeeping shop. They went from 11 hours/week on receipts and bills to 2.5 hours of review. About 8.5 hours back per week across the firm.
Gotchas: Sales tax codes are still a coin flip for AI in multi-state firms. Build a review step. Don't auto-post without human eyes for the first 60 days.
2. Bank reconciliation and categorization
What it does: AI watches transactions hit the bank feed, matches them against open bills/invoices, suggests categories based on past behavior, and flags anything weird. Native QBO and Xero AI have gotten genuinely good at this in the last 18 months.
What to use: Whatever GL you're already on, turn the AI features on. Digits if you want a much smarter layer that learns across your whole client base. Botkeeper if you want it managed for you.
Time saved: Around 60-70% reduction in time-per-rec once it's trained on 3-4 months of your client's transactions.
Gotchas: First two months will feel slower because you're correcting it. That's the rework tax in action. It pays off month 3.
3. Client onboarding
What it does: New client fills out an intake form, AI generates a draft engagement letter, a request list customized to their entity type, and a Karbon or Financial Cents project plan. You review and send.
What to use: A form (Typeform, Tally, Jotform) feeding into Gumloop, which calls Claude to generate the docs, which then drop into Karbon or Google Drive.
This is the kind of workflow we build for clients all the time — it's stupid simple and the time savings are immediate. Going from "I'll get to that onboarding next week" to "they're set up by tomorrow morning" changes how prospects feel about you on day one.
Time saved: Roughly 2 hours per new client. Not huge per client, but if you onboard 50 clients a year, that's 100 hours.
4. Tax document extraction
What it does: Client uploads a folder of 1099s, K-1s, W-2s, brokerage statements. AI pulls the numbers, maps them to the right lines on the return, flags anything missing.
What to use: TaxGPT or Aiwyn if you want purpose-built. DIY with Claude or ChatGPT if you have unusual docs — Claude is currently better than ChatGPT at long, structured tax documents in our testing.
Time saved: Thomson Reuters' 55% figure tracks with what we've seen. Some firms hit 80% on simple individual returns. Don't expect that on returns with rental properties, K-1s from multi-tier partnerships, or anything with foreign income.
Gotchas: Always verify the numbers against source docs. AI hallucinates digits sometimes — usually swaps two numbers or drops a zero. A two-pass review (AI extracts, human verifies, AI generates summary) catches almost all of it.
5. Advisory narrative writing
What it does: You give the AI a P&L, a budget, and some context on the client's quarter. It drafts the narrative section of your management report or the talking points for your QBR.
What to use: Claude is the best at long-form business writing right now. ChatGPT for shorter punchier pieces. Don't paste client data into the free consumer versions — more on that below.
Time saved: 30-45 minutes per client per month, more if you're doing detailed quarterly reviews.
For orchestrating these five workflows, the workflow automation platform comparison breaks down which platform fits which kind of shop. Don't pick one before reading it.
Build vs. buy: when to use a specialist tool, ChatGPT, or a custom agent
Three buckets. Each has a place.
Specialist accounting AI ($200-2,000+/month per seat): Vic.ai, Trullion, Botkeeper, Digits, Karbon's AI features. These tools learned on accounting data. You don't write prompts. They cost more because they save you from having to think about prompts and integrations. Worth it if you're doing a specific workflow at volume.
Generic LLMs with good prompts ($25-30/month per seat): ChatGPT Team, Claude for Work, Gemini Workspace, Microsoft Copilot. Cheap, flexible, terrible at sticking to a process unless you build the scaffolding. Best for one-off tasks: client memos, research questions, drafting emails, explaining things to staff.
Custom AI agents on no-code platforms ($50-200/month plus setup time): This is where most firms haven't gone yet and where the biggest leverage hides. You take a process unique to your firm — say, a custom client check-in workflow — and build a no-code AI agent that runs it.
We build with Gumloop as our primary workflow platform. The interface is genuinely good, the LLM integrations are clean, and you don't have to fight it to get a working flow. Make and N8N are fine too — many firms use them — but Gumloop has been the fastest for us when building accounting workflows specifically. We've tried all four (yes, including Zapier for the simple stuff) and Gumloop wins on AI-heavy work.
For the AI development itself — writing prompts, testing them, building the logic — we use Claude Code. It's our daily driver. Not because it's the only option, but because it's the best one we've found for the kind of iterative agent-building work small firms need.
A typical small-firm AI stack looks like:
- One specialist tool (usually Dext or Vic.ai for AP)
- ChatGPT Team or Claude for Work for staff
- Gumloop for one or two custom firm-specific workflows
- Native AI inside QBO/Xero/Karbon
That's it. You don't need 17 tools. You need the four that fit your actual work.
Honestly, most "best AI tools for accountants" listicles are written by affiliate marketers, not practitioners. They list 30 tools because more tools means more affiliate links. Real firms run 4-6. If a vendor isn't earning its monthly fee with measurable time savings, kill it at the next renewal. The AI tools we actually use post has the short list.
Client data security: what every accountant needs to lock down before using AI
This is the one section we'd ask you not to skim.
Public ChatGPT and Claude consumer accounts train their models on your inputs by default. That means if you paste a client's trial balance into the free version, it's in the training data. Some accountants treat this as a technicality. The state CPA boards do not.
The fix is straightforward:
1. Use the enterprise versions. ChatGPT Team or Enterprise, Claude for Work, Microsoft Copilot for Business. All of them contractually do not train on your inputs. The team plan is $25-30 per seat per month. Cheaper than one billable hour of cleanup.
2. Build an anonymization habit. Even on enterprise plans, don't paste SSNs, EINs, or full client names if you can avoid it. Replace with "Client A" or "Vendor 1." Especially for testing prompts.
3. Read the AICPA and your state board guidance. The AICPA put out updated guidance in late 2025 on AI use in attestation work. Your state board may have its own rules — some are stricter than AICPA. CPA.com has a good clearing-house page.
4. Document a firm AI policy. Five bullets is enough to start:
- Approved AI tools (list them)
- Approved use cases (list them)
- Prohibited inputs (PII, full client docs, anything in attestation engagements unless tool is approved)
- Review requirement (human signs off before anything goes to client)
- Training requirement (annual, including new staff)
Firms with a documented AI policy save 17% more time according to the Karbon 2026 report. That stat surprised us at first — until you realize the firms with a policy actually rolled out AI on purpose. The ones without are just dabbling.
5. Vendor due diligence. Ask any AI vendor: Where is data stored? Who has access? Is it used for training? Is there a BAA available? If they hedge, walk.
What AI still can't do (and probably won't in 2026)
We're going to disappoint some of the AI maximalists here.
Knowing your client's business. AI doesn't know that this client's "consulting revenue" is actually three different revenue streams that need to be split for tax planning. It doesn't know that their landlord's a flake and rent payments are always late. Context is the part you keep.
GAAP and tax law interpretation in edge cases. AI is great at well-documented common situations. It's confidently wrong in edge cases. K-1s from multi-tier partnerships. ASC 842 lease accounting on weird modifications. Section 199A with multiple businesses. Treat AI output in these areas as a starting point, never the answer.
Final review and sign-off. This is on you. State boards have been clear: AI is a tool, not a licensed professional. The signature is yours.
The 37% rework tax. Accounting Today's number isn't going to zero this year. AI gets better, but new use cases keep finding the failure modes. Budget for the review time when you're calculating ROI.
Client relationships. Nobody hires you because you have the best AI stack. They hire you because they trust you with their numbers. AI can free up time to be more present in those relationships, or it can let you onboard 30% more clients and dilute the time you give each one. That's a choice you make, not one the tool makes for you.
Honestly, this is the section most "AI for accountants" articles skip, and it's the section that determines whether AI actually makes your firm better. Frame AI as a junior staff member you supervise, not a partner you trust. That mindset survives contact with reality.
If you're a sole-prop CPA doing complex high-net-worth work, big chunks of this article don't apply to you. Your edge is judgment, not throughput. Use AI for admin and content. Skip the agentic workflows until your volume justifies them.
How to roll out AI in a small accounting firm in 90 days
Here's the plan we use with firms we onboard.
Days 1-30: One workflow, one tool, baseline the savings
Pick receipt and bill processing. It's the easiest win. Sign up for Dext or turn on Vic.ai or activate Ramp's AI features depending on your stack. Run it for 30 days on every client.
Document the time you spent on receipt and bill processing the month before. Document the time the month after. Multiply across your client base. That's your baseline ROI and the number you'll show partners (or yourself) when defending the spend.
Don't add a second tool yet. Don't write a policy yet. Just prove one thing works.
If you also handle AP for clients, our invoice automation guide goes deeper on the AP side of this same workflow.
Days 31-60: Write the policy, add the second workflow
Write the 5-bullet firm AI policy from the section above. Send to staff. Make them sign it (digitally is fine). This is the step most firms skip and it's why their time savings get eaten by chaos.
Then add bank rec and categorization as workflow #2. Native QBO/Xero AI is fine to start. Add Digits or Botkeeper if you want a smarter managed layer.
By end of day 60, you should have:
- Two workflows running
- A policy on the wall
- A measured baseline of time saved
- Staff trained on what tools to use, when, and how
Days 61-90: Build one custom workflow with Gumloop
This is where most firms stop and shouldn't. You've got the standard stuff humming. Now pick one thing unique to your firm — a custom client onboarding flow, a monthly close checklist generator, a tax-season status update email — and build it as a custom Gumloop workflow.
This is the workflow your competitors don't have. It's the one that compounds. And it's the one that turns AI from "a thing we use" into a real operational advantage.
That's the kind of system we build for clients when they bring us in. Not because they couldn't do it themselves, but because the difference between "we use AI" and "AI runs while we sleep" is mostly setup time most firms don't have.
When we set this up for a 4-person bookkeeping firm in Q1, the custom Gumloop workflow alone saved them 6 hours a week. The off-the-shelf tools were saving 12. The custom one was the difference between "AI is nice" and "AI is the reason we took on 14 new clients without hiring."
That's the goal. Not bigger firm. Better firm.
Frequently asked questions about AI for accountants
Will AI replace accountants?
Short answer: no. Longer answer: AI will replace accountants who don't use AI. The 92% adoption number from Karbon's 2026 report tells you the field has moved. Accountants who supervise AI well will do more work for more clients with the same hours. Accountants who pretend it doesn't exist will lose work to firms that use it. The job changes. It doesn't disappear.
Can I use ChatGPT for accounting?
You can, but not the free consumer version with client data. Use ChatGPT Team or Enterprise, which contractually does not train on your inputs. For non-client work — research, drafting, learning — the free version is fine. For anything touching client data, you need an enterprise plan plus an anonymization habit plus a firm policy.
What's the best free AI tool for accountants?
If you want truly free, Google Gemini in the free tier is solid for research and drafting. Claude has a free tier that's good for longer documents. But for client work, free tier tools are not appropriate. Budget $25-30/month per seat for a proper Team plan. That's table stakes in 2026.
How much time does AI actually save accountants?
Gartner says 5.4 hours per week in gross savings. Accounting Today says 37% of that gets eaten by reviewing AI outputs. Net it out and you're looking at roughly 3.4 hours of real saved time per week per person. Firms with documented AI policies save 17% more than that. Books close 7.5 days faster on average per Stanford GSB. Tax prep is 55% faster per Thomson Reuters. Real numbers, not hype, and they compound month over month once your workflows are dialed in.
Originally published at brothersautomate.com. James and Brendan Pinder are co-founders of Brothers Automate, where they build AI automation systems for service businesses doing $1-5M.
Top comments (0)