Invest1 publisher3 min readPublished
Karbon finds nine in ten accounting clients want to be told when AI touches their work
The survey covers 350 small-business owners who handle their own accountant relationships, and 89% want some transparency on AI against 21% of firms with a documented policy, two figures counted on different populations.
The Investor · Invest desk
What happened
- Adam Shay, a CPA, asked on LinkedIn why AI should require a different level of disclosure than the technology clients have accepted in accounting work for decades.
- Of those respondents, 89% want some transparency about how and when AI is used in their work, and 57% want full transparency.
- Only 21% of accounting firms have a documented AI policy or strategy, according to the research CPA Practice Advisor cites.
Compiled by The InvestorSomething wrong?How this is made
Why it matters
- contradiction The two percentages behind the 68-point gap are counted on different populations, and the article describes the 350-person client sample but not a firm-side one, so the spread measures client demand against firm practice.
- constraint Clients will hand AI the anomaly checks and the processing, so the advisory and crisis hours a firm bills at its top rates are the ones an AI rollout is least licensed to touch.
- decision A firm now has to choose between cataloguing every AI feature buried in its software and writing down how it reviews AI output, and the article argues the catalogue gets less useful as the features multiply.
- exposure With no policy on file, the professional who signs the work is the only control standing between a tool whose output varies and a client's return.
The 89% and the 21% count different people. The 89 is client-side, drawn from 350 small-business owners and leaders who personally manage the relationship with an outside firm, which puts roughly 312 respondents in the transparency column [1][2][17]. The 21 counts firms. CPA Practice Advisor does not describe a firm-side sample [18]. Subtract one from the other and you get 68 points [13].
The more useful split is inside the client sample. Some transparency draws 89% and full transparency draws 57%, so about 32 points of respondents want to be told something without asking to see everything [2][15].
The task-level answers matter more for revenue than the disclosure question does. In Adam Shay's summary of the findings, 61% want the trusted adviser role handled entirely by a human and 58% say the same about crisis management [5]. Clients will accept AI identifying anomalies, processing information and supporting routine work, and turn uncomfortable when it interprets a complicated situation or delivers difficult news [6]. If a firm's top rates attach to advisory and crisis work, the permission clients are granting covers the other end of the book. The 39% and 42% who did not choose an entirely human answer are the population a firm would have to win over to move AI up the rate card [16].
The reason the tax-software comparison breaks down is in the failure mode. Deterministic systems are expected to return the same result from the same data and instructions [7]; generative AI output can vary on similar input, and can read as authoritative while being incomplete, misleading or wrong [8]. So the control is the reviewer, not the software. A qualified professional still has to evaluate the output, weigh the facts and stand behind the conclusion, because AI cannot take on professional responsibility for the quality of the service [10]. CPA Practice Advisor wrote that "The AI generated it" will never be an adequate explanation when a client receives incorrect advice [11].
Which is why I think the 21% is a drafting backlog. About 79% of firms have a document to write [14]. A written policy is cheap to produce, and whether anyone opens the file and checks the output is a separate question. I would expect that share to climb quickly while client-reported comfort barely moves.
The case against that: if the 57% who want full transparency start putting the question into engagement terms, disclosure becomes a condition of the work, and firms that skipped the software inventory the article calls impractical and unhelpful [9] get asked for one anyway. There is a third path, in which nothing much happens, because the 350 respondents were answering a survey, and stated preferences on governance are famously cheap. The recommended posture, per the same article, is "Yes, we use AI, and we have a system for managing its risks" [12], and a firm can say that sentence with a two-page policy or with a real review process behind it. A second wave of the same survey, matching the documented-policy count against client comfort, would separate the two.
What to watch
- A second wave of the Karbon survey, showing whether the documented-policy share rises while client-reported comfort with AI stays flat.
- Small-business clients writing an AI disclosure clause into engagement letters or RFPs, which would move the question out of marketing and into contract terms.
- Any professional standard or insurance requirement that asks firms to produce the documented review process.