How to make an accounting firm AI-native: a playbook

In October 2025 Deloitte agreed to refund part of the A$440,000 (about US$290,000) it was paid for a report to an Australian government department. The report cited papers that do not exist and put words in a federal judge's mouth. A revised version disclosed that Azure OpenAI had been used[62]. A large firm's review process let it through.

In June 2026 the IRS Office of Professional Responsibility cited that report in its first guidelines on AI in tax practice. They say that "practitioners cannot rely solely on AI; human scrutiny and editing are essential", and that a firm should "fairly credit to the client's account any cost reductions" AI brings[13].

Staffing makes it harder. Accounting degrees fell 6.6% to 55,152 in 2023 to 2024[32]. Starting pay for master's hires rose 17% in two years[31]. Tax professionals e-filed 72.8 million individual returns by April 17, 2026, 53% of all e-filed returns[21], and the deadline did not move.

Agents can now do most of the preparing: sort the uploads, enter the figures, list what changed since last year, draft the workpaper. But each return still has one signing preparer, each audit one engagement partner, and the reviewers between them have no more hours than before. The question for a firm is whether its reviewers, its client consents and its records can keep up with what agents will draft. This playbook is about getting those three ready first.

Four things set the AI-native firm apart:

1. The firm plans review capacity. It sets thresholds, sampling and a daily limit for each reviewer before agent volume arrives.
2. Every agent action runs under a named professional's authority. The agent can do no more than that preparer may do, for that client and year.
3. Client data stays in its lane. Return information serves only the return, SSNs stay onshore, and one client's file never informs another's draft.
4. Pricing and staffing change on purpose, and clients are told how AI is used.

No firm will replace its tax software, its clients' ledgers or its practice management system to become AI-native. Agents will work across them, and each product's own AI acts under its own login. None of them can say, across all of them, which agent touched which client's return, for which preparer, and who cleared it. The IRS expects firms to document their AI steps[13], and the Safeguards Rule expects them to log what authorized users do[22].

OrchKernel is a control layer that sits between agents and those systems. Agents ask it before they act. It checks what the agent may do and for whom, holds what needs a reviewer, hides fields the agent should not see, and keeps a tamper-evident record. It is not a tax engine, a ledger, a practice management system or an e-file provider, and it never signs anything. The blueprint near the end maps it to the firm's 18 control points.

Few accounting firms publish their numbers, but two public ones show the shape. At CBIZ, operating expenses were 86.4% of its $2.758 billion revenue in 2025, down from 88.9% in 2024, and "the majority of our operating expenses relate to personnel costs"[27]. Andersen reported cost of services at 71.0% of revenue in 2025, up from 63.1%, in a year that included its listing[28]. No public dataset breaks down costs for small and mid-sized firms; the AICPA MAP Survey does, for members only.

CBIZ tells investors its tax and accounting work has "a heavier volume of activity during the first four months" of the year[27], and extended returns make a second peak in the fall. Hiring cannot follow that curve, so the same reviewers absorb it. That is why every change in this playbook is tied to the calendar.

Enrollment in accounting programs rose 12.4% to 266,506 in spring 2025, but degrees awarded fell, and new CPA Exam candidates dropped from 42,626 in 2023 to 28,082 in 2024 after a rush to sit the old exam[32]. Starting salaries reached $67,750 for master's hires and $60,834 for bachelor's hires[31].

Government data is mixed: the occupation is projected to grow 5% over ten years[25], while employment at accounting and tax firms fell 0.6% in the year to September 2026[26]. The series measure different things, so read the gap as a hint.

Andersen says "the substantial majority" of its revenue is billed on time and materials[28]. The AICPA says many firms have moved to value pricing while others still bill hours[40]. Where fees are hours times rate, an agent that cuts preparation time cuts revenue unless pricing changes, and the IRS has now said the savings should reach the client[13]. Growth is slowing at the same time: median net client fee growth was 6.7% in the 2025 MAP Survey, against 9.1% in the previous one[31].

Thomson Reuters found 21% of tax firms using generative AI at the organization level in early 2025 and 34% by late 2025, with 14% of tax firms using agentic AI[49,50]. It sells AI tax products and screened respondents for familiarity with AI, so treat those as upper bounds. Among 1,808 CPA tax preparers surveyed in June 2026, 65% use AI for tax research and 16% have no plans to use it, down from 24.3% a year earlier[44].

Firm-level change is rarer. In the 2025 MAP Survey, 13% of firms said they had implemented AI and automation successfully[31]. Karbon, which sells practice management software, reports that 98% of accounting professionals use AI (vendor survey)[51]. The 98% counts people who have tried a chatbot; the 13% counts firms that changed how a return or a close gets done.

Clients mostly do not know. In the 2025 Thomson Reuters survey, about 59% of tax-firm clients did not know whether their firm used generative AI[49].

Named products are examples from the sources, not recommendations, and their performance figures are their own.

Language models are poor tax calculators on their own. In TaxCalcBench, frontier models calculated fewer than a third of federal returns correctly[46]; its authors work for a tax software company. A 2026 preprint found the best plain model right on 66% of 51 held-out returns, while model-written code checked against symbolic tax rules got all 51 right[47]. The practical reading for a firm: let agents extract, organize, draft and check, and let the tax software calculate.

Tools for firms got the venture money. Basis raised $100 million at a $1.15 billion valuation in February 2026[53]; Rillet raised $100 million at $1 billion in August 2026 and claims more than 600 customers[56]; Black Ore raised $60 million in 2023[54].

AI-led buyers bought firms. Thrive Holdings owns and runs more than 70 businesses, including accounting and IT services, raised over $2 billion at a $12 billion valuation in August 2026[71], and made OpenAI an owner in December 2025[72].

Private equity bought the mid-tier. Grant Thornton (2024)[75], Baker Tilly, which then combined with Moss Adams[76], Schellman (2026)[74] and Crowe (2026), where Crowe LLP keeps "all attest services" and a new Crowe Advisory LLC does tax and advisory[73]. If the two halves share an AI platform, that is a new independence question.

The best-known AI-adjacent firm failed. Bench, a venture-backed bookkeeping and tax firm, shut on 27 December 2024 after burning $135 million[66]. Its founder raised $10 million in 2026 for Synthetic, pitched as "a fully autonomous AI bookkeeper"[68]. MAVI, an offshore talent marketplace, argues AI will remove entry-level roles and deepen the mid-level shortage[70], a founder's view rather than data.

An individual return from upload to the log, in a firm at Stage 3. A CAS close and an audit workpaper follow the same shape.

Six stages, ordered by how far an agent's work travels: inside the firm, to clients on templates, into a return or ledger under review, and only then, for non-attest clients, on its own. Each stage is also tied to the calendar. Launch in the change window from May to early October, prepare and run in shadow from November, and freeze from mid-January to April 15.

Written for a firm starting in October 2026. Ninety days ends just before busy season, so the plan ends with a freeze, not a launch. Most of October is partner and admin work, so it can run around the October 15 deadline for extended individual returns.

An agent can prepare each of these. A named person makes the decision and answers for it, however good the agent gets.

Danny Werfel, a former IRS Commissioner, gives the main risk a name. Control dilution is when "the speed and volume AI introduces causes existing review and sign-off processes to become less rigorous in practice, even if they remain unchanged on paper. More returns get processed with less scrutiny per return." He describes the matching failure at the junior level, when "a junior associate accepts the AI suggestion without independent verification"[43].

Nothing in the firm's procedures changes when this happens, which is why it is easy to miss. The bottleneck and the main risk sit at the same desk. The figure below uses round numbers to show the mechanism.

What the gate does. It does not make reviewers faster. Each reviewer gets a daily release limit, and every return arrives with its change list, sources and a sample of unchanged fields to check. When prepared work outruns review, the queue grows where a partner can see it, and the firm decides: add a reviewer, extend, or reprioritize. Without the gate, reviewers make that choice silently by spending less time on each return.

Measure reviewer minutes per return from Stage 1. If the number falls in Stage 3, be able to say why, and show that post-filing errors did not rise.

The loudest AI failures in this sector so far are reports, not returns. Each real case below is tied to the control point that addresses it.

Scenarios, not reported events, each mapped to the control that stops it.

No rule bans AI in this work, and none moves a duty to the software. The 2026 IRS guidance applies existing Circular 230 duties to AI[13], and the AICPA and Werfel have formed a council to refine his risk framework[41].

The hardest open question is the first. Section 7216 allows disclosure to software contractors but says nothing about a hosted model that drafts work. Until counsel or the IRS says otherwise, treat it as a disclosure that may need the taxpayer's consent. The IRS guidance adds that client data must be handled "using only secure, enterprise-approved AI"[13].

Across these rules, four requirements keep recurring. The exact obligation differs by rule and by service, but a firm that builds these four for every client has the base the others extend.

The junior job moves from keying to verifying each figure against its source and explaining every change from last year. That is the job that prevents Werfel's failure case[43], and it has to be taught to people who have never prepared a return by hand.

Reviewer hours are the scarce input, so they get planned like a budget: release limits, change lists, sampling, and a weekly look at minutes per return. Partners remain the owners of judgment and signature, unchanged in law; the IRS says "final decisions must always rest with qualified professionals"[13].

Circular 230 already puts procedures on whoever has "principal authority and responsibility for overseeing a firm's practice"[12], and SQMS No. 1 makes the managing partner ultimately responsible for the quality system[35]. Someone has to own the approved tools, the agent playbooks and the evidence, and put widely used AI tools through change management[42]. In a 20-person firm that is part of a partner's job.

The AICPA and NASBA have added a licensure path of a bachelor's degree, two years of experience and the exam, enacted in 14 states by May 2025[33]. If agents absorb the first drafts that used to be that experience, the firm has to design work that still teaches and still qualifies. No source we found defines what qualifying experience looks like when agents prepare the first draft. Only 5% of firms in the MAP Survey had formal training for staff moving into new roles[31]. Cutting junior hiring without a plan means fewer reviewers in five years.

Offshore staff can cost "as little as 25%" of a US hire, according to one conference speaker[34], an anecdote rather than data. Agents now take on much of the data entry those teams did, and the SSN rule limits what offshore staff can see[4], so field-level masking is what keeps them useful for review support.

The IRS guidance says billing for "time that was not actually spent or double billing for AI-assisted tasks may violate § 10.27", and that cost reductions should be credited to the client[13]. The AICPA answers that this ignores value pricing and the cost of AI itself; it has asked the IRS for clarification, and its CEO called the language not authoritative[40]. In the 2025 Thomson Reuters survey, 6% of tax firms expected rates to rise significantly and 37% slightly, and 25% planned to pass AI costs through to all clients (vendor survey)[49].

Whatever your pricing, be able to show a client what AI did on their work and how the fee reflects it. That record is cheap to keep if every agent action is already logged by client.

OrchKernel is the layer between AI agents and the systems a firm runs on, as described in the missing layer. Agents ask it before they act; it checks the rules, holds what needs a person, and records what happened.

What it is not. OrchKernel does not replace the tax software, the general ledger, practice management, the portal or the audit platform; they stay the systems of record. It does not calculate tax, transmit returns or sign anything.

The controls a firm needs whatever tools it uses, who owns each, and what enforces it.

OrchKernel is source-available under the Business Source License 1.1 and runs on your own servers, so you can read the code that enforces these controls. Running it in production for your own firm is covered; the license page says what needs a commercial license.

Record the baseline before Stage 1, then track the same numbers at each stage. Public benchmarks barely exist for this sector: the MAP Survey detail is for AICPA members, and other benchmark studies are paid. We have not quoted any number we could not source.

Sources were read in October 2026; dates are publication or data dates. Last reviewed October 2026.