Across fifteen real accounting firms worldwide, from solo bookkeepers to a 250-person practice, AI is doing the same job: it takes the repetitive, high-volume work, the invoice coding, the receipt pile, the first draft of a memo, the routine tax-office letter, and a credentialed accountant keeps the judgment and signs off. The firms reporting the biggest gains did not chase a grand AI strategy. They picked one painful task, pointed a reviewed tool at it, and kept a human in charge of the result.
If you run a firm, lead a department, or sign the returns, you have heard two years of noise about AI in accounting. Some of it is real. A lot of it is software companies describing a future that has not arrived. This piece does something narrower and more useful. It looks at fifteen named, small and mid-sized firms around the world, what they actually use, and what actually happened, with a source for each.
None of them are the Big Four. They are practices the size of yours: a solo in Tennessee, a 27-person CPA firm in New York, a Belgian close team of nine, a 250-person London firm. The tools differ, the countries differ, the numbers differ. The pattern does not. Read them as a working menu, not a sales pitch, and pay as much attention to the limits as to the wins.
Key takeaways
- Fifteen real firms, from a solo bookkeeper to a 250-person practice, across the United States, Canada, the United Kingdom, Belgium, and Australia.
- The pattern is identical everywhere: AI takes the repetitive, high-volume work, and a credentialed accountant keeps the judgment and the sign-off.
- Largest reported gains: Countsy saves about 300 hours a month with Vic.ai; Yeti Books about 50 hours a week on receipts with Keeper; HHL cut categorization time about 75 percent with Truewind; Eventus turned a four-hour memo into 30 minutes with OpenAI and retrieval.
- Tools range from general assistants like Claude and ChatGPT to specialized platforms like Vic.ai, Dext, Silverfin, Truewind, FYI, and Digits. No single tool wins; each firm matched a tool to one task.
- Confidentiality is the non-negotiable first step: decide what client data may enter a tool, in writing, before anyone uses one.
- The honest limit: AI clears the routine work and leaves the exceptions to people, and the reported metrics come from the firms and their vendors, not an independent audit.
| Firm | Location | AI tool(s) | Reported result |
|---|---|---|---|
| Countsy | United States | Vic.ai | ~78% of invoices on autopilot, ~300 hours/month saved |
| Gravita | London, UK | Silverfin Assistant | ~100 AI checks per file surfaced 4 real errors a reviewer would have spent 30+ min each to find |
| HHL Advisors | New York, US | Truewind | ~75% less categorization time, 5-6 hours/week saved per person |
| Count | North Sydney, AU | FYI + ATOMate | ~95% of tax-office correspondence automated, one fewer admin role at same output |
| Your Bottom Line | Oshawa, Canada | Custom GPTs + own tax app | ~30% faster closes, tax-planning prep cut ~90% |
| Eventus Advisory | Arizona, US | OpenAI Assistants + retrieval | Memo drafting cut from 4 hours to 30 minutes |
| Good Measure Financial | Knoxville, US | Claude + Google Apps Script | A 20-hour monthly report now compiles on its own |
| Invenio Advisors | Cleveland, US | ChatGPT / Claude / Gemini | Builds its own tools, no outsourced coding since 2023 |
| Yeti Books | Flagstaff, US | Keeper / Double | ~50 hours/week saved on receipt processing |
| GW CPA | Maryland, US | Custom GPT + Ask Blue J | Faster tax research plus a GPT that drove ~50 new subscribers |
| The Free Minded Accounting Group | Tempe, US | Digits (autonomous GL) | Cleanups that took weeks now finish in days |
| PKF Newcastle & Sydney | Australia | FYI | Year-end setup 20-30 min to two clicks, signing 30 min to ~5 |
| Avail CPA | Lethbridge, Canada | Dext | "Twice the work in the same time," bookkeeping became a service line |
| VP Accountants | Keerbergen, Belgium | Silverfin Assistant | Flags deferred costs and post-year invoices at line-item level |
| Best Suited | London, UK | Karbon AI + Dext + ChatGPT | Honest "marginal wins" on time-saving and organizing information |
Countsy: invoice processing that mostly runs itself
Countsy runs finance and accounting as a back office for venture-backed startups, and at peak it handles up to 3,500 invoices a month. Volume like that used to mean a room of people coding invoices by hand. The firm now routes accounts payable through Vic.ai, an AI tool that reads each invoice, predicts the coding, and pushes the clean ones straight through.
The numbers Countsy reports are specific. About 78 percent of invoices clear on autopilot without a person touching them. The time spent on each remaining invoice fell from four or five minutes to under two. Add it up and the firm saves roughly 300 hours a month, which Colman Edwards on the Countsy team describes as the difference between hiring for growth and absorbing it with the staff already on payroll.
The honest part: the 22 percent that does not clear is the hard 22 percent. Odd vendors, split coding, anything that needs judgment still lands on a human desk. Countsy did not remove the accountant. It removed the part of the job that never needed one. Source: vic.ai case studies.
Gravita: a second set of eyes on every file
Gravita is a London firm of around 250 people, large enough that a slip on one client file can be expensive and quiet. They started running client accounts through Silverfin Assistant, an AI layer that runs about a hundred automated checks over a set of accounts before a human signs off.
On one client, those checks surfaced three or four planning opportunities the team had not flagged, plus four genuine errors. Russell Frayne, the firm's director of transformation, made the point that matters: each of those errors would have taken an accountant more than thirty minutes to find by hand, if they found it at all. He calls the tool "like having a friend looking over your shoulder."
This is the use case skeptical partners tend to respect, because it is about risk, not speed. The AI does not file the accounts. It reads them first and raises a hand. The reviewer still decides what to do with each flag, which is exactly where a senior accountant should be spending the hour. Source: Silverfin customer stories.
HHL Advisors: bookkeeping categorization, cut by three quarters
HHL is a New York CPA firm with about 27 people. Categorizing transactions is the kind of work that fills a junior's week and never gets cleaner. The firm brought in Truewind, an AI bookkeeping tool that learns how a given client's transactions should be coded and does the first pass.
Partner Corbin Hanus reports that Truewind cut his categorization time by about 75 percent, and that across the team it gives back five to six hours a week per person. That is not a moonshot figure. It is a steady weekly recovery of time that used to go to repetitive coding, freed up for review and client work.
What makes HHL a useful example is the size. This is not a tech startup or a hundred-partner firm with an innovation budget. It is a mid-sized practice that picked one painful task and pointed a tool at it. Source: truewind.ai customer spotlight.
Count, North Sydney: tax-office correspondence on near-total automation
Count is a North Sydney firm that decided its biggest time sink was correspondence with the tax office. Routine letters, confirmations, and standard replies ate admin hours every week. They combined FYI, a document and process automation platform, with a tool called ATOMate that handles tax-authority documents.
The firm now automates roughly 95 percent of that correspondence. The result they state plainly is that they carry one fewer full-time admin role and still produce the same output, in their words "if not better quality." Barry Tang framed it as capacity, not cuts: the same team covers more clients without the back office growing.
This is the example that answers the headcount question directly, and it deserves an honest reading. A role did go away. It was an administrative one built around moving paper, and the firm chose to let attrition and automation cover it rather than replace it. The accountants stayed. Source: FYI customer stories.
Your Bottom Line: a mid-sized firm building its own tools
Your Bottom Line, or YBL, is a roughly 30-person firm in Oshawa, Ontario. Rather than wait for a vendor to ship the feature they wanted, they built their own. The firm uses custom GPTs, AI meeting analysis, and a tax-planning application they wrote in-house, work that won them a 2025 Karbon AI innovation award.
The outcomes COO Mike Libbey reports are concrete: onboarding about 25 percent faster, month-end closes around 30 percent faster, and tax-planning prep time cut by roughly 90 percent. The tax-planning number is the eye-catching one, and it comes from turning a manual research-and-assembly task into a guided application the team feeds and reviews.
The lesson here is not that every firm should write software. It is that a mid-sized practice with no formal engineering team can now assemble useful tools out of off-the-shelf AI, because the building no longer requires a developer for every step. Source: The Globe and Mail.
Eventus Advisory: a four-hour memo in thirty minutes
Eventus Advisory is a small advisory firm where Glenn Hopper, its finance lead, did something most accountants have not tried yet. He built a custom assistant on the OpenAI Assistants framework with retrieval over the firm's own reference material, so the AI answers from approved sources rather than guessing.
The task he points to is drafting technical accounting memos. What used to be a four-hour job became a 30-minute one. The AI assembles a first draft grounded in the firm's documents, and Hopper edits and signs. The retrieval piece is the quiet hero, because it is what keeps the draft anchored to real guidance instead of plausible invention.
This is AI moving up the value chain, from data entry into knowledge work, which is the part senior professionals usually assume is safe. The judgment did not move. The blank page did. Source: Journal of Accountancy.
Good Measure Financial: a solo firm and a spreadsheet that builds itself
Good Measure Financial is a small Tennessee firm run by Ben Curtis. There was no innovation team and no software budget. Curtis used Claude together with Google Apps Script to pull data out of the tools he already pays for, Karbon and Gusto, and assemble a monthly reporting and operations dashboard on its own.
His line about it is the one to remember: "What used to take 20 hours a month now runs on its own." The AI did not replace his accounting. It wrote the connective code that a small firm normally cannot justify paying a developer to build, and now the report compiles itself each month while he reviews the output.
This is the most repeatable example in the set, because almost nothing about it is special to Good Measure. A general AI tool, a scripting layer, and the data you already own can remove a recurring chore for a one-person shop. Source: Karbon resources.
Invenio Advisors: three people who stopped outsourcing code
Invenio Advisors is a tiny Ohio practice, a director and two staff, led by Don Tomoff. Small firms usually hit a wall when they want a custom report or a small internal tool, because hiring a developer for a one-off rarely makes sense. Tomoff went the other way and used a mix of ChatGPT, Claude, and Gemini to build those tools himself.
His marker for the shift is blunt: he has not had to outsource coding since January 2023. The work is internal reporting and data tools, the sort of thing that used to sit on a wish list because the cost of building never cleared the value of having it. Now the model writes the first version and he refines it.
The honest caveat is that this takes a certain kind of person, someone willing to tinker and check the output line by line. It is not push-button. But it shows that the floor for building useful things inside a firm has dropped, even for a three-person team. Source: Journal of Accountancy.
Yeti Books: fifty hours a week back from the receipt pile
Yeti Books keeps the books for home-service businesses, plumbers, electricians, and the like, and serves more than 180 of them. That niche comes with a flood of receipts and small transactions. CEO Mike Ventrella moved receipt processing and a chunk of client communication onto Keeper and its Double tooling.
The firm reports saving about 50 hours a week on receipt processing and another 10 to 12 hours a month on client messages, while collapsing three separate tools into one. For a bookkeeping shop, where margins live and die on time per client, recovering 50 hours a week is the difference between turning away work and taking it.
The unglamorous truth in this one is the point. The win was not strategy or advisory. It was receipts, the lowest-status task in the building, and that is exactly where high-volume AI earns its keep first. Source: Double customer stories.
GW CPA: AI that also brought in clients
GW CPA is a roughly 13-person Maryland firm where tax partner Barrett Young used AI on both sides of the business. On delivery, the firm uses Ask Blue J for tax research, a tool trained on tax law that answers questions with citations. On growth, Young built a custom GPT he named Generations Advisor and put it in front of prospects.
That second move is the unusual one. The custom GPT drove around 50 new subscribers in three months. It worked as a always-on intake and education tool, answering questions and warming people up before they ever spoke to a person. Most AI accounting stories are about cutting cost. This one added revenue.
It is worth being clear that a research tool like Ask Blue J still needs a professional to weigh the answer, and a marketing GPT needs a real service behind it. GW had both. The AI widened the top of the funnel and sped up the research; the firm closed the work. Source: Journal of Accountancy.
The Free Minded Accounting Group: autonomous ledgers, with an asterisk
The Free Minded Accounting Group, TFMA, serves more than 125 clients out of Tempe, Arizona, and it is the one firm here using an autonomous general ledger tool, Digits, that aims to keep the books current on its own rather than just assist a bookkeeper.
Their reported result is that cleanup engagements which previously took weeks now finish in a matter of days. That is a real jump, and it points at where the category is heading: software that maintains the ledger continuously instead of a person reconciling it after the fact.
Here is the asterisk, and the firm would not disagree. "Weeks to days" is a softer metric than the hours-saved figures elsewhere on this page, and autonomous bookkeeping is the newest and least proven slice of this market. It is the frontier example, included because the frontier matters, and read with the caution the frontier deserves. Source: Insightful Accountant.
PKF, Newcastle and Sydney: year-end setup in two clicks
The PKF offices in Newcastle and Sydney went after the small recurring frictions that no one remembers to fix. Using FYI for document and workflow automation, they rebuilt how year-end jobs get set up and how documents get signed.
Setting up a year-end job fell from 20 to 30 minutes of clicking to, in partner Stacie Shaw's description, "two clicks and a three-second automation." Signing went from about 30 minutes to roughly five. Her best line is about memory, not time: "No one has to remember it anymore." The process now carries itself, so it does not depend on the one person who knew the steps.
These are not dramatic numbers on any single job. They are small wins repeated across hundreds of jobs a year, which is how real capacity gets created in a firm. Source: FYI customer stories.
Avail CPA: twice the work in the same hours
Avail CPA is a larger North American practice, more than 150 people, with about 20 in bookkeeping. They standardized on Dext, an AI capture tool that reads source documents and extracts the data, feeding into Xero and QuickBooks.
Chelsey Deitz puts the result simply: "We can do twice the amount of work in the same amount of time with Dext." That throughput let the firm turn bookkeeping from a cost center into a standalone service line they actively sell, which is a strategic outcome, not just an efficiency one.
The example matters because of scale. Doubling throughput in a 20-person bookkeeping team is a different management problem than a solo automating a spreadsheet. It needs training, consistent process, and buy-in. Avail did that work, and the tool paid off because the process around it was solid. Source: Dext blog.
VP Accountants: catching what the eye skips
VP Accountants is a small Belgian firm of roughly seven to thirteen people. They use Silverfin Assistant during close and review, and their example is the most technical on this page, which is why a CPA audience tends to trust it.
The AI flags specific things at the line-item level: deferred costs that should be spread, invoices dated after year-end that landed in the wrong period. As the team describes it, the system "identifies items we wouldn't catch ourselves." These are not headline errors. They are the quiet ones that pass a tired human eye on the third file of the day.
That is the value for a small firm with no spare reviewer. The AI is a consistent second pass that never gets tired or rushed, on exactly the technical details where a missed item turns into a restatement. The accountant still decides; the AI just makes sure the item gets seen. Source: Silverfin customer stories.
Best Suited: the realist on the panel
Best Suited is a small London firm, and director Dominic Ahern is included here precisely because he does not oversell it. The firm uses Karbon AI, Dext, and ChatGPT across organizing information and routine tasks, and Ahern's summary is measured: "We're seeing some marginal wins in terms of time-saving and organising information."
Marginal wins. Not transformation, not ten times the output. After fourteen firms reporting large numbers, that honesty is worth sitting with. For many practices, especially early on, AI is a collection of small improvements that add up rather than a single dramatic before-and-after.
This is the healthiest expectation to start with. Aim for marginal wins on tasks you understand, stack them, and let the compounding show up over months. Firms that expect a miracle in week one are the ones that quit in week three. Source: AAT Comment.
What still does not work, and what every one of these firms kept human
Fifteen success stories can read like a clean line, so here is the other half, drawn from the same firms.
The last slice is always the hard slice. Countsy automates 78 percent of invoices, which means a quarter still needs a person, and it is the messy quarter. Every honest firm reports the same shape. AI clears the routine and leaves the exceptions, and the exceptions are where judgment was always the point.
Confidentiality is not optional and not automatic. Accounting runs on client data that is regulated and privileged. The firms doing this well decide in advance what may go into a tool and what may not, and they favor tools with retrieval over a firm-approved source, like the Eventus setup, rather than pasting client detail into a public chatbot. If you take one rule from this page, take that one.
The numbers come from the firms and their vendors. The outcomes here are reported, not independently audited. They are credible and attributed to named people, but they are also the best cases the vendors chose to publish. Treat them as what is possible, not what is guaranteed on your files.
Autonomous is still early. The Free Minded Accounting Group shows where the category is going, but fully autonomous bookkeeping is the youngest and least proven part of this market. The safe posture today is AI that drafts and flags, with a human who reviews and signs.
It rewards a tinkerer, or training. Good Measure and Invenio got their wins because someone was willing to experiment and check output carefully. Firms without that person inside usually need a method and some training to get the same result, which is the gap a structured program is meant to close.
How to think about your first step
If these firms point at one move, it is this: do not start with a strategy, start with a task. Pick the single most repetitive, highest-volume job in your week. For most practices that is transaction categorization, document capture, or a first draft of something you write often. Point one tool at that one task, keep yourself on the review, and measure the hours for a month.
Then decide what stays in the building and what never leaves it, in writing, before anyone pastes a client file anywhere. Get that confidentiality rule down first, train the team on the one workflow, and only add a second task once the first is boring. That is the unglamorous order every firm on this page followed, whether they said so or not.