What 437 accounting professionals, most of them small-firm owners, told us about how they actually use AI today, what's blocking them, and where the profession is heading next.
I started the AI Lab because of what I keep seeing in this profession: the burnout, the workload, the people who love this work being worn down by it. If accountants actually learn to use AI, not collect ideas about it, but use it, they get real time back in their lives. This report is what 437 of them told us about where that journey actually stands. The weekly call where members show each other real work is free, and always will be.
Rebecca Driscoll, CPA · The AI Lab for Accountants
AI is already in the building. Nearly everyone here uses something. The real story is depth: how far past the chatbot firms have gone, and what stops the rest.
Among these applicants, 45% haven't gone past dabbling with their main assistant, while 32% use it daily, including 18% building custom workflows, projects, and MCPs. The distance between the two halves widens every month. (These are self-selected, AI-curious firms; read it as the engaged segment splitting, not the whole profession.)
Today's top uses are tax research (57%) and client email (42%): asking and drafting. What they most want to learn flips to automation and workflows (53%) and building their own tools (30%). The profession has learned to talk to AI; it now wants AI to work.
The share of power users more than triples from solo firms (10%) to firms of 51–200 (35%). Yet firm-built, shareable work is still rare everywhere: fewer than 1 in 16 report having built something they could hand to a peer, distinct from the 18% "building" for themselves.
Client-data safety, PII, and firm policy show up again and again as what firms need before going further, often as an external constraint (IT departments and policies that block AI outright). Tellingly, only 12% say they want to learn security. They don't want a course. They want a ruling: "is my exact setup allowed?" As the calls below show, the field's dominant safety proxy is simply wrong.
Roughly one in nine respondents, unprompted and in their own words, asked to see what other firms are actually doing with AI. Nobody asked for theory.
% of respondents who answered · n = 416
multiple answers allowed · n = 419
This is the small-firm profession: over half run or work in firms of 2–10 people, and tax dominates the work mix, but most firms wear several hats.
Every respondent placed themselves on a usage ladder. The result isn't a hump in the middle; it's an early half and an advanced half, with the middle thinning out. The ladder measures depth of hands-on use of a primary assistant; AI use in general is near-universal (see the multi-model reality below).
% of all respondents · n = 437
% within each size building custom work · 51–200 is n = 20, directional
Ask what accountants do with AI now and you get research and writing. Ask what they want to learn and the answer flips to automation, integration, and building. Read the two charts together (graphite is now, purple is next), and the trajectory is the whole story.
% who described current uses · n = 370
% who answered · n = 344
"Getting over the hump of using AI for one-off stuff and into recurring applications."
Firm owner, 2–10 person firm
Respondents run an average of 1.7 other AI tools alongside their main one. ChatGPT is nearly ubiquitous; Gemini and Copilot follow. This is why "45% still early" is about depth, not exposure. Most of that 45% already use AI daily somewhere; they just haven't gone deep with a primary assistant.
multiple answers allowed · n = 411
The survey says security is the blocker. Our discovery calls with firm owners say why: not that AI is unsafe, but that no one can confidently answer "am I allowed, and how do I set it up?" The field's safety shortcuts are wrong, and even the most qualified people can't resolve the rules. Four patterns from the calls:
The field's dominant proxy for security, and a misleading one. A widely-shared community spreadsheet uses SOC 2 as its only signal. SOC 2 says nothing about whether a vendor trains on your client data; that's a contract term, in a document nobody reads.
Firms were routinely unaware their tools claim training rights: one proposal tool asserts a "perpetual, irrevocable sublicense" over client inputs; a common ledger appears to train with no opt-out. Owners went quiet when shown the clause in their own stack.
A CISA- and CISSP-certified IT auditor with ~700 clients still could not resolve whether his setup triggers §7216 consent, and called a major vendor's own terms self-contradictory. If he can't settle it, no solo firm can. That gap is the whole problem.
A 25–30-person firm, stuck talking to IT about a policy that was never written, afraid of the model "seeing things… without it escaping." Meanwhile a partner presented a Claude summary that mislabeled a client's line item as "vinegar." The client replied, "we don't sell vinegar."
"I want to know what I don't know. I learn so much from seeing what others are doing that I might not have even thought about."
Firm owner, on the one thing they want most
The survey is a snapshot. The calls are the tell. Four shifts we're betting the next two years on, and none of them are five years out. They're happening in the firms in this sample, this month.
The doing of the work is compressing toward zero: preparation, drafting, cleanup, first-pass review. What's left, and what gains value, is judgment: deciding what's worth doing and standing behind the answer. The firms that win won't be the ones that produce faster; they'll be the ones that advise better on top of production that's nearly free.
Firms won't stop at delegating tasks. The capable ones in this survey are already building their own CRMs, document-sweep pipelines, a "firm OS." That's more advanced than the profession's reputation suggests, and it has a consequence: a firm that builds software becomes the data controller of software its WISP has never heard of. The compliance surface grows with the ambition.
Capability is no longer the constraint; permission and articulation are. The scarce skills become telling the model precisely what you want, and being able to say "yes, safely." When an $8M firm's IT shut down Claude, Cursor, and a finished payroll agent over §7216 and SOC 2, the build wasn't the problem. The ruling was.
As the tools converge, the difference between two firms is the humans, and whether they've done the paperwork to move fast without getting shut down. A firm with a current WISP, a documented §7216 posture, and a real AI policy gets to sprint. The firm without one gets benched by its own IT department. Compliance stops being the thing that slows you down and becomes the thing that lets you floor it.
The single most common request in this survey, unprompted, was a window into other firms. That request is the AI Lab. And it maps cleanly onto the three things holding firms back.
Every safety call traces to primary source (§7216, the FTC Safeguards Rule / WISP, Circular 230, the AICPA Code), verified by a CPA and Certified Fraud Examiner, not an AI's guess. So a member gets a defensible answer to "is my exact stack allowed, and where do I write it down?"
Live sessions, a growing playbook, and real workflows and skills, because the members told us plainly that "by the time you watch a course it's dated." The point of every session is that you leave having built or fixed a real thing in your own firm.
One in nine of you asked for exactly this. Every week you see what another firm shipped, what worked, and what broke: proof, not theory. That's the thing no course, vendor, or influencer can hand you, and it's what the AI Lab is for.
The rest of the market sells speed and hopes the compliance works out. We do both: the wins that save you time, anchored by the guardrail that keeps you out of trouble, because we made the safe path the easy one.
Free. Built by and for practicing accountants.
Survey of 437 accounting professionals who applied to The AI Lab for Accountants between May 5 and July 22, 2026. Respondents are self-selected and AI-curious by definition; this sample almost certainly overstates AI adoption relative to the profession at large, and every finding should be read as a portrait of the profession's AI-engaged segment, not the whole. Percentages for each question are calculated on the respondents who answered that question (n noted per chart). Free-text responses were theme-coded; a response can count toward multiple themes. The adoption ladder measures depth of hands-on use of a primary assistant and is not a measure of whether a firm uses any AI (most do). Qualitative sidebars are drawn from 15 recorded discovery calls with firm owners (2026); quotes carry role and firm-size attribution only, with individual permissions confirmed before publication.